Skip to main content

Development of a novel control technology for ensuring high quality Powdered Infant Formula

Final Report Summary - PURE-FORMULA (Development of a novel control technology for ensuring high quality Powdered Infant Formula)

Executive Summary:
Powdered infant formula (PIF) is a granulated product whose quality and safety depend on tight control of its physical and chemical properties. Particle size and shape determine how well the powder will dissolve and reconstitute into a liquid to be consumed, and also affect the bulk density of the powder, while its nutritional value depends on its having the correct material composition. The PIF industry is very valuable economically, with global sales estimated at US$24 billion in 2015.

Bulk density is a critical property for the proper and safe conversion of the powder into a liquid drink, as the measurement of the correct amount of powder to add to water is done volumetrically using a scoop. The potential exists to overfeed or underfeed the infant if the powder being used is too dense or not dense enough, and this is related to particle size.

PIF is a blended food, made by combining basic ingredients to make a wholesome and complete diet for a baby, so it is essential that the constituent parts are present in the appropriate amounts, including protein, fat, moisture and carbohydrates (lactose).

Moisture content must be below the level that enables it to be stored for its planned shelf-life. If moisture is too high, the powder can potentially decay, which can be very harmful to infants if consumed.

Currently the most common practice in the PIF manufacturing industry is to measure physical density and chemical composition separately once per day, based on a combination of samples collected during the day. This can potentially lead to the loss of a day’s production, if a major deviation from specifications is detected.

Process analytical technology (PAT) systems provide the opportunity to control processes in real time with refresh rates of less than a minute, so that any potential process deviation can be detected and corrected at an early stage. The PureFormula technology is a hybrid system that measure the physical properties of particle size using direct image technology and the chemical composition using NIR technology.

Project Context and Objectives:

The technological problem or need
Good nutrition is essential for the growth and development that occurs during an infant’s first year of life. When developing, infants are fed the appropriate types and amounts of foods, their health is promoted.
In tune with our changing pace of life, the rising incidences of working mothers, increasing disposable incomes and other drivers, the numbers of babies that are bottle-fed and indeed the numbers of infants that are weaned off breast milk to infant formula, means that at some stage in their young lives, a very high percentage (>70%) of European and indeed global babies and infants consume powdered infant formula (PIF).
PIF is a milk-like food which is intended to supplement or replace the milk of the baby's mother. PIF is the most widespread and established alternative to the breastfeeding of the newborn and is characterised by a rapidly growing market.
PIF is unique among foods as it is the sole source of nutrition for babies and infants. Its quality, nutritional profile and safety are thus intrinsic for the health and wellbeing of the babies and infants that depend on it for their survival and growth. EC directive (2006/141/EC) states that “the essential composition of infant formulae and follow on formulae must satisfy the nutritional requirements of infants in good health as established by generally accepted scientific data”.
Infant formula is produced by either dry or wet-blending of ingredients including milk proteins, lactose, vegetable oils, minerals and vitamins. Wet blending, the more common approach, is followed by a drying step to produce agglomerates. The powder is reconstituted directly before use by adding water.
The spray drying of powdered infant formulas is considered a relatively gentle drying procedure. However, degradation reactions as the loss of available lysine due to the Maillard reaction with lactose commonly occur during spray drying of infant formula. Milk proteins contain relatively high amounts of lysine which is an essential amino acid that can only be metabolized in the human body if it has a free ε-amino group. Consequently, the protein quality, i.e. the nutritional value and the digestibility, depends on the loss of available lysine. The use of high temperatures during PIF manufacture can cause a great degree of protein denaturation. This is a major limiting factor, considering the already naturally lower lysozyme and lactoferrin content of cow’s milk, which is the base of PIF. These valuable ingredients are severely denatured by excessive heat treatment. As infant formulae and follow-on formulae are sophisticated products that are specially formulated for their intended purpose, essential requirements on protein, including minimum and maximum levels of protein and minimum levels of certain amino acids, should be monitored (directive 2006/141/EC). The protein requirements specified in this Directive should relate to the final products as such, prepared ready for consumption.
Since most infant formulae are required to be instant, the preferred drying approach is typically designed to allow for product agglomeration during drying.
Agglomeration improves the reconstitution properties. It is therefore important that the degree of agglomeration and the compactness of the agglomeration can be controlled. The identified key physical parameters governing the agglomeration process include granule size/shape and the key chemical parameters include moisture content. Properties of individual agglomerates determine bulk powder properties such as powder flowability, bulk density and solubility. These are important product quality characteristics since, for example, the product is measured out on a volume basis (in scoops) whereas nutritional requirements are provided on a mass basis. Thus any changes in particle properties post drying, during conveying and packaging will have implications. For dry blending process the ingredient distribution within granules and between granules has a critical effect on content uniformity, nutrient absorption rates and ultimately infant safety.
Moreover, as the initial raw material is cow's milk, one of the principal ingredients is fat globules. When the powdered milk is dissolved in hot water in preparation of the milk formula, these fat globules float freely in the hot water, forming a so-called oil/water emulsion. Since it is believed that the particle distribution of these fat globules can to some degree affect their digestion and absorption in the body, this is one of the physical properties that must be evaluated in the quality control process.
A modern spray drying, agglomeration & fluidised bed set-up for PIF manufacture has little process monitoring except for psychometric air conditions and visual particle inspection.
The current method of determining the end-point of PIF processes for spray drying and agglomeration is typically a combination of (1) off-line analysis of size and shape characteristics, (2) off-line analysis of moisture content and (3) end product analysis of content uniformity. Particle size and shape characteristics are important as they govern bulk, flow and solubility properties. The moisture content is important as it governs thermal degradation during the drying process and product shelf-life. Content uniformity is critical as set out by directive (2006/141/EC).
PURE-FORMULA’s development of a hybrid technology that is capable of assessing the physical and nutritional characteristics of PIF in real time would significantly advance process control leading to greater product quality, process validation, and ultimately infant safety. Moreover, a technology capable of providing information on the physical and chemical nature of agglomerates as they form during drying would represent a global breakthrough. The ability to identify the chemical signature of particles spatially will allow profiling of ingredient distribution and ultimately content uniformity for dry blending. Spatial-spectral responses can be used to map agglomerate moisture profiles and detect protein degradation due to thermal damage during drying for wet PIF manufacture. Overall the combined ability of NIR spectroscopy to detect PIF control parameters such as protein denaturisation coupled with agglomerate morphology will provide information on product quality and on structure-functionality relations. The output of this project will be an advanced control strategy for end point determination of the highly variable but critical process of PIF manufacture for both dry and wet methodologies.
To this end, the proposed PURE-FORMULA project aims to develop a technology that is a hybrid of established imaging platforms. On-line physical characterization of size and shape will be determined with the EYECON technology (that provides 3D images of particles) owned by INNOPHARMA. To this technology, multipoint NIR measurement technologies, based upon novel miniaturized MEMs spectrograph advances, will be integrated, to provide nutritional and moisture information of PIF. Data fusion techniques will be employed to combine physical and spectroscopic data. This combined approach will be carried out in real-time to provide on-line agglomerate characterization.
In short, the development of a technology capable of monitoring critical physical and chemical parameters of both dry and wet PIF production processes in-line would significantly increase process understanding and control thereby assuring product quality, infant safety, and lean operational costs. The impacts of the PURE-FORMULA technology for improved PIF formulation and nutritional value, baby health and wellbeing, increased consumer confidence and peace of mind, as well as for the competitiveness of the European infant formula are high. Moreover, this solution has the potential to extend to other food instant powders, such as soups, juices, coffee.

Scientific and Technological objectives
The overall objective of this project will centre on the development of a hybrid device based on collimated 3D imaging of particles’ physical parameters (EYECON technology) and novel miniaturized NIR multipoint or quasi chemical imaging technologies that provide spatial measurements of moisture and ingredient identity from the forming PIF agglomerates. The R&D effort will centre on engineering design, software/algorithm design, user interface design; and integration and validation within PIF manufacturing equipment.
In this project a device for measuring physical and chemical properties of agglomerates will be designed, realized and tested. Technology challenges relate to building a cost efficient, robust, user friendly device which produces multi point or quasi imaging of chemical information (such as moisture distribution, or ingredient concentration) coupled with 3D physical information (particle size and shape) by one single instrument. Spectral information will be generated with a multichannel detector using Fiber-optic coupled multi-point Fabry-Perot technology. With this technology quasi-images or measure in several spatial points by fiber-optics can be obtained. This technology is available to the consortium and has successfully been developed into other devices such as unmanned remote sensing aircraft. VTT patented calibration methods specially designed for in-line analysers will also be employed. Physical information will be produced with a camera and LED illumination based device (EYECON). Novel 3D algorithms are utilized in the calculation of physical information. The technology platform will be selected in tune with industry requirements, as this is a marker-driven approach. Figure 3 shows the proposed integrated design of the proposed technology

In order to realise this, the scientific and technological objectives, and corresponding Performance Indicators, that will be fulfilled during the PURE-FORMULA project are provided below. Every effort has been made to ensure that these objectives are S.M.A.R.T. (Specific, Measurable, Achievable, Realistic and Timely) and clearly linked to the Work Plan in Part A of the DoW:
1. To implement a ‘bottom-up’ approach whereby the needs and specifications of companies from the PIF industry will be consulted, as will regulatory aspects, the findings from which will be used to guide the definition of the specifications for the PURE-FORMULA system.
2. To build a laboratory scale test rig integrating physical imaging and multipoint NIR chemical measurement technology for the physical and chemical characterization of PIF agglomerates. The device will enable physical and chemical information to be obtained with one single setup. The scale-up parameters for developing a pre-competitive PURE-FORMULA prototype for industrial-scale PIF agglomeration monitoring purposes will have been defined. To evaluate the effectiveness of the combination of direct imaging and multipoint NIR spectroscopy measurements for the characterization of the physical and chemical properties of agglomerates against reference methods and to define the limits of detection of the approach. The prototype will be evaluated in laboratory scale fluidised bed dryers. The sensitivity, SNR, selectivity and other key parameters will be tested. The performance of the test rig will identify parameters used to design the pre-competitive prototype.
3. To draw up the designs of the pre-competitive PIF agglomerate characterization hybrid physical imaging and NIR- spectral measurement system in keeping with the industry specifications defined as well as the laboratory parameters defined and to assemble the system hardware. Consideration of integration challenges: retrofitting the technology inside commercial PIF agglomeration facilities.
4. To develop general software for the operation of the PURE-FORMULA system, including synchronized control of the different subsystems, user application and the analytical software that will process the data obtained by the system. This analytical software will use the algorithms and spectral databases developed, in order to characterize the physical and chemical properties of the agglomerates under analysis and readily display the results via an ergonomic User Interface.
5. To integrate the system hardware, software and User Interface in order to provide a pre-competitive prototype that can be validated in industry and to carry out pre-validation tests with the system to ensure its proper functioning before shipping to industry test-sites.
6. To test and evaluate the PURE-FORMULA system in commercial PIF production sites in order to assess the efficacy of the developed system to effectively characterize the physical and chemical properties of agglomerates during processing in a production environment. Design issues arising from the industrial tests will be fed back.
7. To carry out optimisation work on the prototype based on feedback from the field trails, and to carefully outline scaling-up rules and development work for full production.
8. To facilitate the uptake of the PURE-FORMULA results, by the participating SMEs, as well as a wider audience of actors and stakeholders from the PIF value chain, by carrying out a comprehensive series of knowledge transfer and training activities to on the one hand show the validity for the system for reliable, in-line agglomerate characterisation in commercial PIF production facilities, and on the other hand to capacitate end-users about its usability and to outline its benefits for facilitating quality control of the agglomeration process, increased product quality and reliability, reducing plant down-time, reducing costs and trained operator intervention, etc.

Project Results:
The development of the PureFormula technology encompassed the initial development of a bench top test rig and then using the learnings to develop a pre-competitive prototype. The development included both hardware and software. The prototype was then trialed in industrial environments. The physical measurement test rig was built to be able to determine particle size measurement on a wide particle size specification. This is due to the fact that raw materials for the process were down to 10μm in size while the spray drying system can generate agglomerates greater than 1000μm in size and in some cases up 4000μm in size. The system therefore had to be built in order to analyse particles over this wide size range. The test rig was built with two cameras integrated to cover the size range in WP2 and subsequently refined during the execution of WP4 and WP5. The system consists of a micro and macro camera to cover small and larger size particles. Depth of Field (DOF) measurements for the micro camera were made with Multifunction Target 4X-20X (Target 1).
For the macro camera measurements, 1951 USAF Hi-Resolution Target, 3" x 3" Positive (Target 2) No. 64-862 and Depth of Field Target 5-15 (Target 3) No. 54-440 were used. In the setup, the macro camera is facing towards the measuring surface and the micro camera is in a 12º tilt position against it. In the experiments, test targets were placed on the measuring surface and moved with the x,y,z -stage to complete the tests. The microscope lens has a fixed aperture, and the tests with the macro camera (50mm lens) were made with aperture 8, which was found the most suitable for this type of imaging in earlier tests.
The 180 lp/mm pattern of Target 1 was used for DOF definition.
The depth of field for the macro camera has been determined through target 3. Target 3 has 5.0 lp/mm and 15.0 lp/mm patterns when imaged in 45º position (3.54 lp/m m and 10.61 lp/mm in 0º position). The DOF imaging was made using the 45º position. The reported DOF was based on the contrast dropping by 50 % from the maximum value.
The chemical test rig was built to facilitate multipoint assessment, with consideration given to the materials that are used in the constitution of PIF and the area of their peaks of interest. To this end the spectral wavelength range was determined as 1520nm to 2200nm. This range covers all the materials used in the manufacture of PIF and follow on milk powders.

The testing program for the image analysis component of the technology consists of verifying that the system can process the information generated by the imaging system and can provide accurate results for the physical attributes of the sample. Tests were made in laboratory with test targets placed on a linear x-, y-, z-stage.
Depth of Field (DOF) measurements for the micro camera were made with Multifunction Target 4X-20X (Target 1). For the macro camera measurements, 1951 USAF Hi-Resolution Target, 3" x 3" Positive (Target 2) No. 64-862 and Depth of Field Target 5-15 (Target 3) No. 54-440 were used.
In the setup, the macro camera is facing towards the measuring surface and the micro camera is in a 12º tilt position against it. Test targets were placed on the measuring surface and moved with the x,y,z -stage to complete the tests. The microscope lens has a fixed aperture, and the tests with the macro camera (50mm lens) were made with aperture 8, which was found the most suitable for this type of imaging in earlier tests.

In current tilt position, the micro camera gets less light than macro camera. In the performance tests, the difference was fixed by changing the LED pulse width. The pulse width control is common for both cameras, so in the final configuration, a neutral density filter in the macro camera should be used to get even illumination for both cameras. Suitable filters are: ND8 and ND4, available form Edmund optics. ND8 (0.9) Item No. 59-173 and ND4 (0.6) Item No. 59-168.
The aim was to have a quite small process window size. In parallel orientation, the total imaging area is circa 3.9 cm × 2 cm. By flipping the mounting plate between the camera and linear stage and by tilting the micro camera, the image area is circa 2.3 cm × 2cm, being closer to the Particle sizer’s area, which is circa 1cmx1cm. Parallel orientation requires extra LEDs (2 pcs) for micro camera illumination, which is not implemented to current setup. Tilted orientated micro camera gets less light than the macro camera due to the fact that it is little off the main illuminated area. By making new mounting plate, the micro camera can be tilted more to get into the micro image area, inside the macro area without micro lens blocking the macro area. A small portion of the micro lens is seen on the left side of macro camera image. The relationship and positioning of the two cameras were further enhanced in the activities of work packages 4 and 5.

Laboratory performance tests of the chemical measurement test rig were also carried out. The thorough performance tests were performed on four channels.
The measurement set up utilised a combination of a single point probe with integrated halogen illumination and an ocean optics fibre optics probe with separate light source. These were connected to the Multipoint NIR system and the software resided on the system laptop. The following samples were used in the laboratory performance tests.
Optical Teflon OP.DI.MA foil. Used as a standard white sample.
Wavelength standard MRC-910-1920x wavelength standard with many narrow absorption bands in the NIR range.
2, 25, 50, 75 & 99% reflectance standard FGS-02-01c Fluorilon Gray Standard. Used for linearity testing.
LWP1700 Optical long-pass filter, LWP-1700-25. Used for testing the stray light.
SWP1670 Optical short wave pass filter, SWP-1670-25. Used for testing the stray light.

The measurement parameters used in the tests are as follows. (the parameters are typical to normal measurement conditions)
Wavelength range 1520 - 2140 nm.
Wavelength sampling step 10 nm.
AD sampling frequency 25 kHz Default
Sample averaging 2000

Baseline correction (Depends on the measurement) Baseline correction was not used in the stationary tests, since it is not needed when the sample does not move. The signal repeatability was tested by measuring the optical Teflon sample 100 times in succession. Signal repeatability and signal-to-noise ratio results are calculated based on these measurements.
Signal repeatability results: Channels 2 and 3 have a lower level of signal than channels 6 and 7. These differences are caused by different sensitivity of the channels inside the MultiEye. The repeatability of the signal of all probes is very good.
The signal-to-noise ratio (SNR) can be calculated based on the repeatability results reported in the repeatability tests. The signal was calculated as the mean of the 100 measurements, and the noise was calculated as the standard deviation of the measurements. This was done for each spectral point, and the signal was then divided by the noise wavelength-by-wavelength. The SNR is above 200 for all channels at almost the entire wavelength range. Channel 6 has a higher SNR than the other channels. The difference to channels 2 and 3 is explained by the higher level of signal as shown in the repeatability study. This reasoning would also predict the SNR of channel 7 to be similar. However, it appears that there is more noise in channel 7 than in other channels. The reason for this is probably the electrical characteristics of the pixel corresponding to channel 7. The SNR levels of all channels are lower at longer wavelengths. This is because the signal levels are also lower there.

The illumination setup and the probe used have major effects on the results. To demonstrate this, the same signal-to-noise analysis was also done for measurements done with an Ocean Optics fibre probe. The different probe gives a several times higher SNR. It can also be seen that increasing the sample averaging from 500 samples to 2500 samples gives no benefit in the SNR. The noise levels are similar as with the single-point probe.
The signal stability was analysed by taking measurements from the optical Teflon sample overnight at one minute intervals. The results contain information on signal intensity stability. The changes between different measurements are seen to be offsets in the whole wavelength range and don’t change the shape of the spectrum. The signal levels shown as a function of time for selected wavelengths. Some smoothing was applied to reveal the underlying trend from the measurement-to-measurement differences. These results confirm that the spectral shape does not change as all the wavelengths move in unison. The changes in signal levels stay within a ±1 % range.
The wavelength axis accuracy was examined by measuring the wavelength standard sample with all channels. The wavelength step in this measurement was 1 nm to ensure the highest resolution spectrum. The number of averaged samples was also increased to 10 000. The results are compared with a measurement made with a reference instrument (Varian Cary 5000) from the same sample. The absorption band maximum positions measured by the MultiEye are very close to the ones measured with the reference instrument, and the shape of the absorption bands is also quite identical.
The most important features of the dark current of any NIR instrument include the dark current level, dark current noise and dark current stability. The dark current repeatability was studied by disconnecting all probes from the MultiEye, and then measuring 100 spectra. It should be noted that the dark current level of different channels is dependent on the properties of the detector preamplifier, and is thus somewhat arbitrary. Therefore, the dark current levels cannot be compared to each other.
The dark current noise of each channel was computed as the standard deviation of the dark current repeatability measurement. There are significant differences in the noise levels between channels. It is difficult to say whether these differences originate from the detector elements, from the preamplifier or from something else. Because in this section the dark current noise was plotted for all 8 channels it is most likely the detector configuration which accounts for these differences in noise levels. Detector configuration was addressed for future commercial multipoint systems, as previously stated for the purposes of this project we will utilise only 4 channels.
The dark current stability was studied by letting the device cool to room temperature, disconnecting all probes, and then beginning to measure spectra immediately after powering it on. Measurements were made at an interval of one minute overnight.
There is a clear warm-up period of approximately two hours. During this time the dark current level raises significantly in all channels. After that the stability of the dark current is good. This warm up period can be stabilised by a regular automated dark reference function. This will need to be considered for future NIR devices.
Crosstalk between channels was measured by connecting one channel at a time to the single-point probe (this channel is called the “connected channel” hereinafter), while the other channels were disconnected (these channels are called “the other channels” hereinafter). The dark current measurement was taken first, and then the spectra of all channels were measured and the dark current correction was done. Ideally, the result should be that only the connected channel shows signal, and the other channels should have a zero signal. Anything different from zero for the other channels is channel-to-channel crosstalk.
To analyse the crosstalk levels, the relative crosstalk was calculated by dividing the spectra of the other channels by the spectrum of the connected channel, and then multiplying by 100 to convert the result to percentages. At almost all positions, the crosstalk levels stay below 2 %. As the desired crosstalk level would be below 1 %, it can be concluded that the crosstalk performance of the MultiEye may need to be improved. Spectral resolution of the instrument depends on the optical properties of the Fabry-Perot used in the instrument. In order to optimize the resolution, as high orders as possible should be used in wavelength calibration.
The spectral stray light was measured by connecting channel 6 to the multi-point probe. The dark current was measured first, and a normal measurement was taken from the optical Teflon sample. Then, a 1700 nm long-pass filter (LWP1700 in short) was placed into a filter holder in the fibres connecting the probe to the MultiEye. The LP1700 filter blocks very effectively shorter than 1700 nm wavelengths (5 % transmission at 1700 nm, average transmission 0.1 % below 1700 nm). A similar measurement was made using a 1670 nm short-pass filter (SWP1670 in short).
The relative stray light is simply the “with filter” spectrum divided with the “without filter” spectrum. The result was then converted into percentages. The signal level should be about 0.1 % in the range 1520 - 1600nm (according to a reference measurement with Varian Cary 5000), but it is about 6 %. Thus, the spectral stray light level is about 6 %, again this is an area that should be considered for improvement in a commercial NIR system.
Linearity measurements were by measuring Avian Technologies Fluorilon grey standards of levels 2 %, 25 %, 50 %, 75 % and 99 %. Measurements were made with the single-point probe. This probe has a long depth of focus and thus the signal level is not sensitive to small changes in the distance of the sample. This is important for the linearity measurement since the only thing affecting the level of the signal should be the different reflectivity between different reflectance standards. Unfortunately the single-point probe also had a downside of collecting low amounts of light from outside the reflectance standard. This stray light was eliminated as much as possible by having a black background around the reflectance standard, but some of it may have remained nevertheless. The stray light collection problem could were avoided by using the Ocean Optics fibre probe, but that probe has a very short depth of focus which would were a bigger problem in the context of the linearity measurements. Looking at a few wavelengths and plotting the dark-corrected signal level as a function of calibrated reflectance give a series of intensity plots for each measured standard. Each of the lines in this figure corresponds to a different wavelength. Nonlinearity at some wavelengths would present itself as non-straight lines but all of lines are fairly straight. The different lines have different slopes but this only corresponds to different signal levels at different wavelengths and is to be expected.
Ideally all the points should lie on the 45-degree slope, normalized intensity is gotten by renormalizing the signal level so that each of the measurements from the 99 % standard corresponds exactly to the calibrated reflectance they should have. To have a better idea about the deviations from the ideal the difference between the normalized intensity and the calibrated reflectance was plotted. It was seen that the 25 %, 50 % and 75 % levels are well within ±1 %. The lowest intensity is about 1.5 % higher than it should be. A possible cause of this (and perhaps the average positivity of 25 % deviations) is the stray light from the background that could not be eliminated.

In summary the physical test rig, performed well in the experiments. The tests conducted with the physical measurement test rig included depth-of-focus, resolution and imaging area measurements. Test targets were used in the tests. According to the measurements, the resolution of the macro camera is ~30 μm and the resolution of the micro camera ~3 μm. The measurement area for the macro camera is 21mm × 16mm and for the micro camera 1.28mm × 0.93mm.
The chemical measurement test rig, performed well in the tests. The signal-to-noise ratio was up to 6000, and the stability of the signal was good. The differences between the channels were small enough to not cause any performance problems. The dark current characteristics were mostly good, but it should be noted it takes some time for the dark current to stabilize when the instrument is cold-started. This means the dark current measurement has to be taken frequently enough, to avoid artefacts in the absorbance spectra.
The channel-to-channel crosstalk was 1 – 2 %, which is a little bit more than expected, but still quite acceptable. The spectral resolution is wavelength-dependent, ranging from about 20 nm to 60 nm. This was expected based on the design. The signal linearity was so good that the deviation from linearity could not be measured. However, the spectral stray light was too high, up to 6 – 8 %, whereas a stray light level of 2 – 3 % would were expected based on the spectral characteristics of the Fabry-Perot interferometer used. This is not a big problem, but it should be noted that the present chemical measurement test rig is not so good for transmission measurements because of the high spectral stray light.
All in all, it can be concluded that the performance of the chemical and physical test rigs is well in accordance with the specifications set up in Deliverable 1.1 and with the objectives of WP2.

When carrying out experiments in the laboratory, powder sample was circulated in measuring chamber with pressured air and the powder was imaged with two cameras through a glass window. Two different brands of PIF were measured: Nutrilon and NAN. The PIF samples were circulated behind a glass window with pressurized air.

The reproducibility of the measurement results is quite good as the results from different measurements agree well. However, comparison to the Malvern results shows that the D10 and D50 are overestimated. D90 agrees reasonably well with the Malvern results.
The overestimation of the sizes in the samples has two possible reasons: overestimation of individual particle sizes and/or an unrepresentative sample of particles, which are analysed. The used algorithm fits an ellipse to each found particle. The fit ellipse is an ellipse that has the same second moments as the particle projection in the image plane. Due to these facts when the particle shape is not very much like an ellipse, the calculated ellipse (and the volume calculations based on it) may overestimate the size of the particle. This would lead to larger than actual D10, D50 and D90 values. On the other hand, the used version of the software collects all the observed ellipses from both the micro and macro cameras during the configured time window and then calculates the distribution parameters based on this set of individual ellipses. However, this leads to a sampling error, since all found ellipses are treated equally in calculating the distribution, but only micro camera can see the smallest ellipses. The volume from which small particles can be found is much smaller due to the fact that the micro camera image area is smaller than the macro camera area by a factor of about 280, and in addition, the micro camera depth of field is smaller than that of the macro camera. This sampling bias should be taken into account in calculating the distribution, but correcting the bias is not straightforward because the particle sizes that each camera can observe overlap.
The MultiEye has eight channels altogether. However, in the PIF tests reported here, only four of the channels (2, 3, 6 and 7) were used. These are the centre most channels of the detectors.
The MultiEye spectrometer is connected with the multipoint probe. The probe, in turn, is attached to a rotation rig. The rotation rig has a large rotating wheel, on top of which the samples are placed. The speed of rotation can be changed from zero to about 20 – 30 RPM, slightly depending on the load of the wheel. The rotation system is used to simulate a moving sample.
The multipoint probe has a halogen bulb, a multi-faceted ellipsoid back-reflector for illumination, an elliptical mirror for signal collection to a fibre bundle. The fibre bundle has eight fibres arranged in a row. The multipoint probe can be used to measure spectrum from a set of eight points in a row.
The powdered infant formula (PIF) samples used in the tests were Cow & Gate and SMA. The sample was piled near the edges of the bowl, since the measurements are taken from there. The surface of the sample was deliberately made uneven so as to simulate measurement distance variations. Rapid measurement distance variations should cause interference in the spectra of a scanning instrument like the MultiEye. The full spectral range of the MultiEye was used, with the typical measurement parameters. The baseline correction function was in use in some tests, and not in use in others. The rotation speed was varied from zero (stationary) to 0.13 m/s and 0.3 m/s.
From the results of measuring a stationary sample without the baseline correction and with the baseline correction from the Cow & Gate PIF sample with channel 2, several observations can be readily made. Similar spectra with the other measurement channels showed no discernible difference from channel 2.
• There are no essential differences between the spectra measured with the different channels.
• The most remarkable differences between the “without baseline” and “with baseline” spectra are seen in the wavelength range 1920 – 1980 nm. There are differences in the spectra in this range even within the two sets.
• In addition, there are small baseline changes in the “without baseline” spectra.
• The 1920 – 1980 nm range is related to one of the NIR water absorption bands. Therefore, the differences in this region are caused by the drying effect of the heat from the probe to the sample.
• The baseline differences probably also originate from the sample moisture changes. The baseline variations are, of course, not seen in the “with baseline” spectra.
When the sample moves, the spectra get distorted in the “without baseline correction” case. This is expected, since the small sample-to-probe distance variations cause small intensity variations. Because different wavelengths are measured at different times by the MultiEye, these intensity variations show up in the final spectrum. However, these intensity variations almost completely disappear when the baseline correction is on. The small intensity variations are still there, but they are continuously monitored by the reference arm. After computing the absorbance spectrum, these small variations are compensated for by using the data of the reference arm. This compensation works extremely well in this case.
The baseline correction works also well with the SMA PIF powder. The baseline correction transforms the almost useless spectra to high-quality ones.

Single ingredient lactose was also tested under these controlled conditions to obtain a spectral signal and compared to full spectrum NIR reference characterisation.
Variable moisture content of powders and agglomerates was evaluated to obtain a drying profile.
Protein denaturing was monitored from the spectral response as a function of heating conditions.
MultiEye NIR spectra of various commercially available Powdered Infant formula (PIF) powder were recorded under controlled static bench-top conditions. MultiEye NIR spectra of agglomerates of Powdered Infant Formula (PIF) prepared by dry granulation technique were recorded separately and analysed by multivariate statistical analysis.
Various PIF brands and PIF agglomerates were imaged using RGB particle imaging system. These images were processed to get physical properties of size and shape and compared to reference measurements from stereo microscopic image analysis.
NIR spectra of PIF ingredients i.e. lactose and casein, were recorded. NIR spectra of binary blends of lactose and casein at various ratios were also recorded and multivariate statistical techniques were utilised for quantification of casein in these blended ingredients.
PIF SMA2 and PIF SMA2 agglomerates were stored under various humidity conditions (approximately 33%, 53% and 84% at room temperature) for 5 days to modify their moisture content. NIR spectra of powders and agglomerates were recorded before and after humidity treatment and correlated to moisture content in order to obtain the drying profile. Correlation was carried out by multivariate statistical analysis which allows building a calibration model to predict the moisture content. Thermo Gravimetrical Analysis (TGA) was used as a reference method for moisture content determination and is based in the difference in sample weight before and after heat treatment (105°C for 24h).
Protein denaturation in PIF proteins and PIF SMA2 as a function of heating conditions, 80oC and 105oC, was monitored by recording spectral response after 10minutes and 24h heat exposure.

Measurements were performed with the 4 probes arranged in square shape, set perpendicular with the samples and at 1 mm distance. Known amount of sample (3.5g for moisture analysis, 5g or 12g for protein denaturation studies) was placed in a one-use aluminium dish and its surface was flattened to reduce the impact of distance difference between the sample and the probes. Samples were scanned at the full wavelength range and using a 5nm scanning interval. For each sample, measurements were taken at five different regions of the sample surface, acquiring 20 measurements (4 probes by 5 replicates). For samples containing a ratio of casein and lactose at various levels both ingredients were weighed at the proportional weight to obtain 30 g of final weight. Samples were manually blended in a plastic resalable bag for 1 minute.
Raw data was pre-processed by Standard Normal Variate (SNV) and for moisture analysis Savitzky- Golay smoothing was also applied.
For moisture analysis PIF SMA2 and SMA2 agglomerates (1.0 mm) samples were accurately weighed in aluminium pan and placed at various relative humidity (RH) conditions for 5 days. Samples were weighed again after humidity treatment and NIR spectra of the samples were collected by MultiEye before and after humidity treatment. Samples were placed into an oven at 105oC for 24h in order to determine the moisture content of the samples before and after the humidity treatment. NIR spectra of all samples were correlated with the moisture content obtain by Thermo Gravimetrical Analysis (TGA).
Protein denaturation assessment utilised the following: PIF SMA2 and PIF ingredients: casein, whey protein isolate (WPI), lactose and their mixture casein: WIP: lactose (1:1:1) were subjected to 80oC (Memmert oven, Fisher Scientific, Dublin, Ireland) and 105oC (Gallenkamp BS Oven 250 Size 2, AGB Scientific, Dublin, Ireland) for 10 minutes and 24h). NIR spectra were collected by MultiEye before (room temperature) and after heat treatment

Physical characterising.
The Particle sizer has a field of view of approximately 9mmX6mm. The distance of the imaging system from the particles was constant for all different particle category sizes. This distance was approximately 10mm. For each sample, images were taken at various regions of the sample surface, while image acquisition was running in a continuous mode at the acquisition rate of 1 second per image. The number of images recorded and analysed varied depending on the size of particles, due to the constant field of view. Minimum one image was taken for particles below 100 μm and maximum of 30 images was taken for particles above 1000 μm. Images for each sample were processed in a batch mode.
Samples were also imaged by a reference method; stereomicroscope (SZH10) equipped with Moticam 580 camera 5 MP. Each sample powder was placed on the slides and viewed under the imaging system. A 4-dot calibration slide (0.07 0.15 0.6 1.5mm) and micrometer cross-hairs (10mm in 0.01mm increments) were also pictured to ensure accurate calibration for accurate measurements. Three images were taken for each product. Images were processed individually and then parameters were calculated and averaged.
ImageJ Fiji was used for image processing: images were first converted to 8-bit greyscale (Type > 8-bit), thresholding was adjusted to select specific pixel intensities and regions (Image > Adjust > Threshold). The image was then converted to binary mode with black and white colours only (Process > Binary > Make Binary). Noise was removed using the smoothing mask (Process > Smooth) or the noise filters (Process > Noise > (1) Despeckle or (2) Remove outliers). Touching particles were individually separated using a segmentation algorithm (Process > Binary > Watershed). The image analysis does not take into consideration incomplete particles at the peripheries of the image. Scales for particle sizing were defined using known pixels to distance values, based on dot and cross-hair calibration slides of 10, 70, 150, 600, 1000 and 1500 μm references. In the program, these pixel to size values were entered (Analyse > Set scale). Particles were then analysed with required criteria selected.
Key conclusions:
MultiEye NIR can identify various ingredients in PIF brands, such as proteins and carbohydrates over the NIR range 1515 to 2295nm, when analysed under controlled static bench-top conditions. Results show that MultiEye NIR followed by multivariate statistical analysis such as PCA, can possibly differentiate between various PIF brands, probably based on the moisture content. By applying multivariate statistical analysis, such as PLS, MultiEye NIR is capable of quantifying PIF protein, casein in binary powdered mixtures of PIF ingredients, casein:lactose. These results show potential of MultiEye of determining the levels of carbohydrates and proteins in PIF. The particle sizer is capable of recording images, which followed by image analysis can estimate size and shape of particles. These results were comparable with that obtained from the reference method, Stereomicroscope. In future, efforts were made to use image analysis algorithm developed by DIT team to estimate the particle size and shape.
MultiEye NIR followed by multivariate statistical analysis such as PLS can measure the moisture content in PIF and consequently a drying profile can be obtained. Efforts were made to validate this model to allow quantification of moisture from test samples. However, efficacy to measure moisture level decreases as particle size increases, as it was found with PIF agglomerates of size 1mm. This could be due to larger particle size and spaces between these particles, allowing relatively lesser number of particles to be detected by probes at constant field of view. As the granule size increases, the measured sample appears more heterogeneous. MultiEye show promise at monitoring denaturation of proteins when PIF is subjected to heating conditions.
The various ingredients of PIF were blended at controlled ratios and analysed using the bench-top approach. Ingredients of various particle size and shape distributions and were studied and correlated to off-line reference techniques.
Chemical identification of controlled blended ingredients were made at varying ratios. Approaches to assess ingredient distribution from the multipoint spectral signatures were examined for dry bending of PIF.
For wet manufacture, blends of manufactured agglomerates with controlled moisture content was assessed and an overall moisture content and distribution profile was produced with a view to spray drying end point determination for wet blend PIF.
Multivariate statistics techniques such as PLS (Partial Least Squares) was utilised to construct models to predict the nutritional and moisture profiles of granules.
Executed activity:
PIF and granules of PIF was imaged using the new Particle sizer prototype. These were correlated to off-line image analysis reference technique, Stereo microscope. Algorithm scripts written on R statistical computing package were used to calculate and study particle size and shape distributions. Standard alumina spheres with size ranging from 100um to 2500um, and cellets from size 100um to 1000um were first studied to validate the new Particle sizer prototype and also validate the algorithm scripts written for particle size and shape analysis.
NIR spectra of PIF ingredients i.e. lactose and casein, were recorded. NIR spectra of dry binary blends of lactose and casein at various ratios were also recorded and multivariate statistical techniques were utilised for quantification of casein in these blended ingredients.
Wet granules of lactose as well as PIF were prepared, NIR spectra of these wet granules were recorded during drying over 3.5 hours to assess overall moisture content and distribution profile. Multivariate statistical techniques were utilised to construct models to predict the moisture profiles of these wet granules. A gravimetric method bydrying samples in oven at 105°C to a constant weight was used as a reference method during construction of models to predict moisture profile.
Multivariate statistical techniques such Partial Least Squares (PLS) were utilised to construct models to predict the nutritional profiles (protein profile) of PIF from the spectral response after heat exposure of PIF at 80oC and 105oC for 10 minutes and 24h. FT-IR Spectroscopy - Attenuated Total Reflectance (ATR) was used as a reference method during construction of models to predict nutritional profile.
Moisture profile analysis:
A gravimetric method was used as a reference method to determine moisture content in wet granules during drying. Wet granules were sampled every 30 minutes and their moisture content during various stages of drying was determined. NIR spectra of all samples were correlated with the moisture content by chemometrics method to obtain a prediction model for moisture profile.
The new version of the MultiEye instrument with 4 channels and reference arm was used to measure spectral data during drying of wet PIF granules. Measurements were performed with 1 probe set perpendicular with the samples and at ~20 mm distance. The use of a collimator lens at the end of the probe allowed this large distance between the sample and the probe. Samples were scanned at the wavelength range of 1560 to 2100nm and using a 5nm scanning interval.
The SNV NIR spectra of wet lactose granules were analysed during various stage of drying, i.e. 0 – 3.5h sampled every 30 minutes. The spectra showed absorption band at 1940nm, related to water. This absorption band decreased as time progressed, indicating drying of wet lactose granules. Decrease in levels of moisture over 3.5h was confirmed with offline gravimetric analysis.
Partial Least Square (PLS) was applied on the full spectral data with gravimetric analysis as a reference method, to build a prediction model for moisture profile. Results generated by cross validation with “leaving-one-out” option indicated that five components in the calibration model were needed for high determination coefficient (R2=0.9603) and low Root Mean Squared Error in Prediction (RMSEP=2.005%). The Cross-validated prediction model for moisture content in PIF for the five-factor model showed that the points for each time point correspond to each one of the 4 analysed spectra recorded by each probe.
Casein quantification:
The spectra of pure 100% casein showed absorption bands around 2055 nm, characteristic of proteins (Khodabux et al., 2007). These absorption bands decreases as the casein content decreases in the mixture of Casein:lactose. The main absorption bands of lactose, from 1480 to 1550 nm and around ~2100nm were found to increase with an increase in the lactose content.
Partial Least Square (PLS) was applied on the full spectral data to build a PLS model to predict the casein content in the mixture. The distribution of the scores of thetwo main principal components; casein and lactose were displayed in two clearly separated clusters and the corresponding binary mixtures were scattered according to the concentration range.
After PLS treatment, cross validation with “leaving-one-out” option was applied. The cross-validated root mean squared of error prediction (RMSEP) and the determination coefficient (R2) was obtained as a function of number of factors or components for the calibration model. In order to avoid overfitting, a minimum number factor
model was chosen, 3 factors. Variations in the RMSEP were small for number of factors n=3 (RMSEP=2.982%, R2=0.987). Similarly, total variance of 98.92% was observed for samples spectra at various weight proportion mixtures.
Cross-validated prediction by the model with 3 components is showed that the points for each concentration correspond to each one of the 4 analysed spectra recorded by each probe.
Nutritional evaluation:
An evaluation was conducted of the SNV NIR spectra of PIF protein, casein before and after treatment to high temperatures 80°C and 105°C for 10 minutes and 24h, a commonly encountered phenomenon in spray drying process. The spectra showed absorption band at 1940nm, related to water (Corredor et al., 2011) and at 2055nm related to protein (Khodabux et al., 2007). These absorption bands decreased as with an increase in temperature (80°C and 105°C) and time of incubation (10 minutes and 24h). Decrease in absorption band at 1940nm can be related to drying of wet PIF granules, while a decrease at 2055nm could probably be related to protein denaturation.
Protein denaturation at higher temperatures was confirmed with offline FTIR-ATR analysis.
PCA was applied on NIR spectra in the range 1900-2100nm to cover absorption bands for water (1940nm) and protein (2055nm). This distribution can be related to combined effect of decrease in moisture content and protein denaturation. This was confirmed by the loadings plot where PC1 showed the score distribution was related to decrease in moisture content at 1940nm while PC2 showed the score distribution was related to the denaturation of protein at 2035nm, which is close to 2055nm for proteins as reported in the literature by Khodabux et al., 2007.
Partial Least Square (PLS) was applied on the full spectral data with FTIR-ATR as a reference method, to build a prediction model for the level of casein denaturation.
Results generated by cross validation with “leaving-one-out” option indicated that two components in the calibration model were needed for high determination coefficient (R2=0.9865) and low Root Mean Squared Error in Prediction (RMSEP=0.2478%).

By applying multivariate statistical analysis, such as PLS, The MultiEye NIR is capable of quantifying PIF protein, casein in binary powdered mixtures of PIF ingredients, casein lactose.
The MultiEye NIR and its new version, with an option to be equipped with collimator, allowed spectral measurements of wet lactose granules and wet PIF granules, in-process, during drying on a vibrator. By applying multivariate statistical analysis such as PLS, the MultiEye NIR is capable of measuring the moisture content during drying of wet granules, consequently a drying profile can be obtained, and potentially end-point of granulation could be determined. Gravimetric method to measure moisture content was used as a reference method while generating the prediction model for moisture profile. This model requires further validation.
By applying multivariate statistical analysis such as PLS, the MultiEye NIR is capable of measuring level of PIF protein denaturation, as a function of heating conditions, conventionally encountered during processing stages in food and pharmaceutical industry. FT-IR - Attenuated Total Reflectance (ATR) was used as a reference method while generating the prediction model for the level of protein denaturation. This model requires further validation.

Tests were performed to assess the ability of the test rig to quantify particle properties during motion. The bench-top unit was employed to image particles of defined characteristics as they move past the field of focus utilising a moving stage. The stage velocity was increased and any limits of detection for particle characterisation identified.
A second DIA arrangement was assessed where a narrow stream of particles was passed through the field of focus under controlled free-fall conditions. The ability of the system to cope with particle velocity and the percentage of particles analysed during free-fall was assessed.
The ability of the new Particle sizer prototype to quantify particle properties during motion was assessed. PIF granules of various sizes were imaged as they move past the field of focus utilising a moving stage. The stage velocity was increased at three levels, minimum, medium and maximum. R algorithm scripts developed by DIT team were
used to calculate particle size and shape.
A narrow stream of particles was passed through the field of focus under controlled free-fall conditions and the ability of the system to cope with particle velocity and the percentage of particles was assessed during free-fall conditions. Free-fall speeds of 3/4 levels were investigated.
PIF and granules of PIF was imaged using the new Particle sizer prototype. Standard alumina spheres with size ranging from 100um to 2500um, and cellets from size 100um to 1000um were first studied to validate the new Particle sizer prototype and also validate the algorithm scripts written for particle size and shape analysis.
Test were executed under varying conditions, vibrating speeds, free fall, and vertical movement and compared against static measurements. Results were also compared to a stereo microscope method.
The Particle sizer prototype is capable of recording images. Image analysis was carried out on R statistical computing package using algorithm scripts developed by DIT team. Particle size and shape estimated by using these scripts were comparable with the standard sizes or with that obtained from the reference method, stereo microscope. The Particle sizer prototype is also capable of recording images in dynamic conditions, conventionally encountered in food and pharmaceutical industries. Particle size and shape measured under dynamic conditions were comparable with that measured under static conditions.

The test rig was also integrated with a laboratory scale spray dryer, and challenged with the measurement of material during the drying of skimmed milk powder. Spray drying was conducted under different operational conditions such as solids content of the concentrate (10-40%), feed rate (2-30% of pump’s capacity, inlet air temperature (150-180°C) and aspiration rate 60-80% of compressor’s capacity.

Commercial Powdered Infant formula (PIF) and skimmed milk powder were acquired from the Irish market. Standard sugar spheres of various sizes (100, 200, 500, 750,
1000, 1500 and 2500 μm) were supplied by Hanns G. Werner GmbH + Co. KG (Germany).
Spray dryer: A single-state spray dryer (Büchi Mini Spray dryer, Büchi Labortechnik AG, Switzerland)
PureFormula test rig

The effect of the operational factors during the spray drying process was investigated. Four spray drying batches of skimmed milk concentrate were performed under different operational conditions.
The performance of these test-batches allowed for a better understanding of the process and identification of the predominant factors. System cleaning was found to be essential for keeping the equipment in good working order as spray dried powder was not collected under non-cleaned conditions. However, as it happened in test 2, an excess of humidity in the system caused a similar effect since the presence of liquid water in the drying chamber stopped the concentrate from drying. Then, milk concentrate (probably with water) was collected instead of powder. A notable amount of milk concentrated was also collected in the condensate vessel.
Besides the cleaning, the solids content of the concentrate and the feed rate also played an important role in preventing the nozzle from clogging. During test 1 spray drying, milk concentrate overflowed from the nozzle probably due to an excessive feed rate and solids content, eventually causing the absolute obstruction of the nozzle.
Feed rate needed to be increased over time. This fact is probably due to particle accumulation inside the nozzle, requiring a pressure increase as the process advances. The outlet air temperature is subject to many factors. Test performance indicated the outlet air temperature is highly influenced by the inlet air temperature and especially the feed rate. The reason is probably that such factors regulate the humidity of the system, which will determine the evaporation rate and consequently the outlet temperature. The aspiration rate showed to have a lower-grade effect on the outlet temperature.
The fourth spray drying batch was conducted satisfactorily. Approximately 300 ml of milk concentrate (32 g of milk powder) were processed and 16 g of spray dried powder were collected. In other words, the yield of the process was around 50%.

The effect of measuring through glass was assessed. Spectra of a standard reference (FW-WCVisNIR-O2, Avian technologies LLC, Sunapee) were recorded in both modes: directly and through the glass of the spray dryer collecting vessel.
The results indicated that the combination of MultiEye with fibre probes is capable of measuring through glass since the recorded spectra presented minor differences between measuring modes. However, the use of the collimator was found to perform poorly and further work will be required to optimize the collimator with respect to obtaining signals through curved glass.

The performance of the MultiEye as an in-line measuring system for the spray-drying process of skimmed milk was assessed. The fibre probes were attached to the spray dryer. Spectra of the spray dried powder were recorded during the spray drying process every 2-3 minutes. The process consisted of three batches where 100ml of milk concentrate was spray dried.
The spectra recorded in-line were pre-processed by SNV and Savitzky Golay smoothing. In order to compare the spectra, spray-dried powder was also measured off-line, without glass and using the same sensor configuration.).
Despite sharpness on the in-line spectra, spectral similarities to the off-line spectra were noticed, especially at the regions where lactose and water have absorption. However, water peak in the in-line spectra is more noticeable indicating that the moisture content of the powder was probably higher during the process, above all, batch 2. The moisture content of the spray dried powder was 5.8%, which was analysed by gravimetric method (Memmert oven, Fisher Scientific, Dublin, Ireland at 103°C for 24h). According to the references, moisture content of spray dried skimmed milk powder should be lower than 4%. Therefore, further adjustments need to be made to simulate the industrial manufacturing of skimmed milk powder. Estimation of particle size and shape of spray-dried powder using the PureFormula particle characteriser was performed. Particle size and shape were determined by image analysis with the algorithm scripts developed by DIT. Measured and real sizes were correlated to obtain a calibration curve (R2=0.998). The estimated particle size of the spray dried powder was 44 μm. Therefore, particle size of the obtained spray dried powder was considerably smaller than commercial PIF (approx. 125 μm – estimated in report 3.3).
The ability of the PureFormula prototype to image through glass was also assessed. Images of standard spheres of several sizes and PIF were taken through the glass of the spray-dryer collecting vessel. Image analysis was carried out using the R algorithms to estimate the particle size and shape. Measured sizes of the standard spheres were correlated to the real values to obtain a calibration curve for size estimation when measuring through glass. Finally, the calibration curve (R2=0.99) was applied on the PIF size measured by the R algorithms (88.8 μm).

The PureFormula prototype was then designed and built based on the learnings from the test rig assessments. Some elements of the design are developments of parts used in the test rig. However, the optics of the camera system for particle-sizing required a significant re-design to enable full analysis of the particle measurement range from 10 to 5000um. For this reason, a Function Means Analysis was performed on the camera optical design, to decide the optimum configuration.
Three alternate optical designs were investigated to achieve the inline particle-sizing requirements. In all cases, a beam-splitter cube is used to combine the optical paths to the macro and micro cameras, so that their fields of view overlap. Zemax modelling software was used to generate models of the microscope portion of the alternate optical designs.
Each of the optical designs was fully built and tested on an optics bench to obtain a complete analysis of the image quality that is achieved in reality. In all cases the microscope image quality was good, and the macro imaging resolution was also good. However, the macro images generated by Design 2 and 3 were affected by Barrel and Pincushion distortion respectively, which is a comparison of images obtained using the 1951 USAF resolution test target. Therefore, Design 1 was selected as the best option in terms of maximum resolution and quality for both images. The micro camera can resolve down to the smallest feature of the target which was available to IRIS at this time, which is Group 7, element 6, which has a spatial frequency of 228 line-pairs per mm. The contrast at this frequency is 0. 11 of the maximum contrast obtained in the image for large features (14/130 grey levels). This can be compared to the results of the modelling carried out with Zemax.

Depth of Field
The depth of field was calculated for the micro and macro optics, using the smallest resolved features of the USAF test target, obtained over a range of working distances. The acceptable resolution or circle of confusion for the micro camera is 5um, so that 10um particles can be identified. At this resolution, the depth of field is 300um.
The acceptable resolution for the macro camera is 40um, which is achieved over a working distance of 8mm. For each camera, the Depth of Field is larger than the diameter of the largest particles which the camera is designed to image, which are 250um and 5mm respectively.
Block Diagram
Based on the specifications which were developed using the Quality Functional Deployment (QFD) system, a block diagram of the prototype was developed as the first step of the design process to map out the individual components and their interconnections.

Optical Sub-Systems
The prototype brings together two optical metrology techniques for monitoring industrial processes in Powdered Infant Formula production. Dual process monitoring of physical and chemical properties can be achieved either at the same location or else separated by up to 10 meters. The two techniques are multipoint NIR spectroscopy and particle-sizing by direct imaging.
Multipoint NIR
Inline multipoint NIR spectroscopy is realised using the same Multieye spectrometer as developed for the test rig but with a series of additional optical fittings developed for industrial monitoring.
Fibre optic probes
Custom fibre optic probes were ordered with SMA connectors on the ends of the probes for coupling to the collimator lenses. The total probe length is 2 metres.
A probe for long distance monitoring was ordered with total length 10 metres, which will enable simultaneous process monitoring at the two locations identified as most critical.
Short fibre cables were designed to fit inside the body of the particlesizer. These contain the same arrangement of six fibres around one that matches with the sampling end of the probe fibre, so that the probe length can be extended through the body of the particlesizer.
Collimator lenses
Custom designed adjustable collimator lenses are fitted to the ends of the fibre optic probes to extend the working distance to the range of 2 - 5cm. For deployment of the fibre probes in this way, it is necessary to integrate collimator lenses within the front wall of the particlesizer. The short fibre cables run from the collimators within the front wall to SMA adapters that are integrated in the rear wall, allowing coupling to the Y-branched fibre probes.
The full optical system for particle-sizing plus the optomechanical assembly to house the optical elements is described as follows.
Camera Optomechanical assembly
The construction of the dual camera imaging system is based on modular lens tubes and optic mounts plus a plate and mounts for the cameras that are custom made. This provides a compact, stable and enclosed optical path for the dual camera system. The longer optical path of the micro camera is bent to make the most compact design. The assembly rests on an optics base plate which is fixed to the platform of a translation stage that provides 1 inch of travel for focus adjust of the optical system.
Micro and Macro Cameras
Several optics parts are re-used from the benchtop test rig, including the two cameras and the LED ring.
LED ring: The light source consists of 16 white-light emitting LEDs in a ring, angled at 75° to the optical axis to overlap their beams at a distance approximately 1cm in front of the ring holder.

The mechanics of the particle-sizer are almost entirely redesigned for the inline version.
The LED ring is re-used from the benchtop version of the test rig, and this forms the basis of the front wall. A steel flange extends around the LED ring, which is compatible with the tri-clover clamp system, for a 6” pipe. Behind the tri-clover flange, the front wall profile is a cylindrical surface of outer diameter 167mm and length of 18mm, which can also be used as a location for attaching the prototype using a simple circular clamp. The external cover fits inside this cylindrical section of the front wall, helping to ensure ingress protection against dust and liquids.
The Internal framework includes a base plate of 5mm thick aluminium which carries the optomechanical assembly including the translation stage which enables focusing. The base plate supports the front and rear walls with L-shaped brackets. A side wall of 1mm thick steel supports the PCBs that provide power and control to all the electronic parts. This wall also provides additional support to the front and rear walls.
The rear wall is a steel disc containing the electronic cable connector. The cover is a stainless steel cylinder which can be removed by sliding towards the rear of the prototype over the rear wall. A mounting block with a series of M5 and M6 tapped holes is welded to the exterior of the cylindrical cover.

The enclosure rating for the mechanical assembly is IP54, providing protection against the ingress of dust and splashing water. This is achieved by using a single continuous steel cylinder as the cover, with rubber seals at the front and rear where it connects with the front and rear walls. The materials used in the exterior are all GMP compatible, including stainless steel, glass and Teflon. The enclosure can be cleaned using detergents and solvents.
Enclosure Rating
The internal electronics from the benchtop test rig are re-used for the inline version, as they don’t have any impact on inline capability. In addition, for the inline prototype, we have added a controller PCB for the motorised translation stage. This is an off-the-shelf control box which has been adapted for the prototype by removing the enclosure and associated switches, so that just the control board itself is fitted inside the prototype.

There are 6 Individual cables from the exterior to different parts of the prototype. Because of the multiple cables, some of which include adapters for converting to USB, the inline prototype will combine all of the cables in a single 32-wire cable with a robust connector. This requires the building of an external electronics box to receive the four communication cables from the computer and power from the mains supply, and to convert these to dedicated pins in the large cable. The electronics box will also enclose the two cable adapters and a 24V DC transformer. Though the box is an additional item, it can be mounted close to the prototype pc and the NIR multipoint spectrometer. The particle-sizer which is mounted inline onto the granulation machine will have a single electronics cable leading to the operator and pc.

The computer used is a Dell Latitude E6440, with i5 dual-core processor. The operating system is Windows 7 32-bit, which is compatible with the drivers available for the NIR spectrometer and particle-sizer electronics. The screen is Full High Definition, with 1920 x 1080 pixels, to optimise the resolution of the camera images.

The Software for the PureFormula system included:
• The development of the general software for the operation of the PURE-FORMULA system, including the user application and the analytical software that will process the data obtained by the system. This analytical software will use the algorithms and spectral databases developed.
• The design and development of the user application.
• The integration of the system hardware, software and User Interface
The PUREFORMULA application is composed of the following parts:
I. The PUREFORMULA user interface which allows the user to control the whole prototype. The user interface is developed using Windows Presentation Foundation (WPF).
II. The particle-sizer software which connects the controls of the user interface with the settings of the LEDs, cameras and translation stage to adjust focus. The LEDs controller is an executable provided by VTT which is called directly from the PUREFORMULA control software. On the other hand, for the cameras and the translation stage, the respective manufacturers provide libraries which are used to create custom classes in C# to control them both.
III. The Multieye software that controls the NIR spectrometer device. The control and acquisition software is written using LabVIEW and is embedded as an executable file. This is called from the PUREFORMULA control software, and an interface software written in C# is programmed to enable communication between them.
IV. The particle-sizer analysis software that analyses the particle images for both cameras. This is based on the algorithms provided by DIT to analyze the particles images and includes combined statistics developed by IRIS. This analysis is implemented in C++ and an executable file is generated in order to be called from the PUREFORMULA control software.
V. The analysis routines provided by DIT to analyze the spectra, based on Partial Least Squares Regression. This analysis is done in R, which can be called from a C# project using the functions provided by the RDotNet library.

The particle-sizer uses two Basler Ace cameras, each with a gigabit Ethernet connector for video transfer and a Hirose 6-pin connector for power and triggering. In order to control the Basler cameras the Pylon 4 software suite is installed,
Frame-transfer from the cameras to the laptop is through an Ethernet connection. Since the laptop used in the prototype only has one Ethernet port, for one of the cameras there is the need to use an Ethernet 2.0 to USB converter (Delock Gigabit-Ethernet-to-USB Adapter, for USB 2.0) with appropriate drivers, which allows use of an available USB port in the computer. Therefore, the cameras used in the PUREFORMULA prototype have an IP address, so to access them, the computer network is set with an IP address that belongs to the same subnet as the cameras.
Translation Stage
The translation stage is an MTS series from Thorlabs, with a TDC001 “T-cube” controller. The connection with the translation stage is done directly through USB. In order to control the translation motor, the following suite and drivers are installed:
• ThorlabsUI, called “APTUser”.
• Drivers to control by USB
To control the LEDs, the User Interface executes a program provided by VTT called: “LED_controller.exe”.
The USB connection with the LEDs needs a driver that Windows O.S. downloads automatically once an internet connection is available.
A control and acquisition software package for the Multieye NIR spectrometer has been written by VTT using LabVIEW. This is re-used for PUREFORMULA, and an interface software is written to enable communication from the main software application, written in C#. The LabVIEW Runtime Engine, which is a free software, is required to enable the LabVIEW commands. The next software and drivers are needed to control the Multieye:
• Driver LCP134X
• NI-DAQmx: National Instruments data acquisition software.
• NI-VISA: National Instruments Virtual Instruments Software
• NI-IVI: National instrument drivers that perform the actual communication and control of the instrument hardware in the system
• LabVIEW Run-Time Engine 2013

The chemometric analysis software developed by DIT is written in R, which is a high-level mathematical language. For the PUREFORMULA application, analysis is performed by calling R when required. R is installed on the PUREFORMULA computer with the packages containing the necessary functions.
• # R software:
• # R.NET NuGet:
• # R packages:
o signal

Control and Analysis software
The PUREFORMULA software is written in C# using Microsoft Visual Studio 2013. For the acquisition, LabVIEW and Multieye software are needed in case of the Multipoint NIR, while for the ParticleSizer cameras, the Pylon 4 software suite from Basler is used. In order to do the analysis using R for the Multipoint NIR, R needs to be installed on the computer as well as R runtime and some R packages that are listed on the section 3 Drivers and Software needed.

The visual studio solution contains two projects, a C# project called PUREFORMULA_Labview, that is the main project and controls the hardware (both particle-sizer and NIR spectrometer), contains the user interface and performs the Multieye NIR analysis; and a win32 project called “PFparticleSizer” that is written in C++ which generates the program that analyses both camera images.
This mentioned main project is divided in the following different parts.
Visual Studio solution structure):
• Basler Cameras.
• Database.
• Icons. Contains the images used in the user interface.
• LED controller. Contains the class and the executable file needed to control the LEDs.
• Multieye structs. Contains the structure of the data used to store the Multieye data.
• Multieye controller. Contains the programs to control the Multieye.
• NIR analysis. Contains the class for the Multieye NIR analysis.
• ParticleSizer_analysis. Contains different classes for the particle sizer analysis and for the graphics (trendlines and histogram for the particles and NIR analysis).
• R files. Contains the Rcode provided by DIT and modified to match the needs of the user interface as well as the fields with the coefficients used by the R code.
• Translation stage. Contains the custom class that controls the translation stage.
Multipoint NIR
The following section describes the parts of the software used for the control and analysis of the data from the Multieye device
NIR Data acquisition.
The Multieye device is managed by an application developed in LabVIEW by VTT which includes the libraries and drivers needed for the control of the hardware. In order to control the Multieye device from PUREFORMULA UI, the shared variable engine offered by National Instruments is used. This technique allows the transfer of data and commands between LabVIEW and the PUREFORMULA control software, which is written in C#.
In order to control Multieye from C#, some variables had to be created to be shared between the control software and LabVIEW.
NIR Analysis
The data acquired by the Multieye device are submitted to an analysis to obtain the result of the measurement, which is the percentage chemical composition. This analysis uses R code developed by DIT.
The DIT functions read the required data from a file, which are the 4 absorbance spectra most recently acquired by the Multieye. So the acquisition software creates automatically a file called “raw.txt” with a header containing the wavelengths of the acquired spectra, and four different lines, each of which is the absorbance spectrum of one of the 4 different probes that are used. Each of the four probes is connected to a different channel of the spectrometer. The spectrometer has in total 8 channels in a rectangular array, and four of these are used, so the probes are designated with the numbers 2, 3, 6 and 7. The other channels are not enabled for PUREFORMULA.
To be able to run the R code from the UI software the following actions have to be taken:
1. Load the R library for C#.
using RDotNet;
2. Declare a variable of type Rengine.
private REngine engine;
3. Set the environment variables.
4. Get a reference to the R engine, creating and initializing it if necessary.
engine = Rengine.GetInstance();
5. Set the input arguments.
6. Include the file with the script and the needed functions:
7. Include the R workspace with all the needed functions and models:
engine. Evaluate("load(\"Rfiles\\\\data.RData\")");
8. Execute the desired function from the R engine.
DynamicVector probe_result = engine.Evaluate("functionName(parameters)").AsVector();
The analysis has the capability to calculate the percentage composition of four different parameters, which are moisture, lactose, protein and fat, in the milk powder that is flowing by a probe. The composition of a parameter is calculated using a model by applying a set of coefficients to the absorbance spectrum, using the function shown below:
fat=(new_spectral_data%*%Fat_coefficients[-1])+Fat_coefficients[1]) where fat is the desired component being obtained, from a spectrum called “new_spectral_data”.
The set of coefficients is generated offline using R, following the calibration analysis procedure developed by DIT, and the absorbance spectra are pre-processed with the appropriate treatments, including Standard Normal Variate, Savitsky-Golay smoothing, as described in D3.4 Ability to obtain particle information from laboratory scale spray dryers.
The set of coefficients for each parameter to be analysed is stored as a matrix, in the form of a .txt file. These files are stored in the folder: “C:\PureFormula\Rfiles\”.
PLS analysis models of this kind are generated for very specific acquisition conditions. So for example, a model to calculate the percentage of moisture in powdered in fant formula (PIF) will change depending on the production plant, the type of PIF, the location in the production process, or the type of spectrometer used to acquire the spectra. For this reason, the option is available to store several (up to 5) versions of models for each of the parameters being monitored (moisture, lactose, protein and fat) depending on the conditions under which the models were built. Models that apply to the same conditions are given a common number as part of their filename: parameter_coefs_number.txt. Models for all four parameters for that set of conditions are combined in a common R-function with the same number in its filename: NIR_
The appropriate models to use for a certain set of monitoring conditions are selected by the user with the “type of analysis” combo-box. The combo-box index = 0 means that no analysis should be done on the spectra.

Three elements are controlled in order to acquire the images needed to do the particle size analysis: the LEDs and the two cameras.
LEDs control
The LEDs are controlled by an executable called “LED_controller.exe”. The following parameters can be modified from the PUREFORMULA user interface:
• LEDs’ intensity
• Pulse delay
• Pulse frequency
The pulse frequency is directly related with the camera acquisition rate because the signal that starts the LED pulse is hardware connected to the internal trigger of the cameras.
The C# class LEDS.cs works as an interface between the LED software and the PUREFORMULA control software.

The class that controls the cameras is found in cameraProvider.cs. It initializes the cameras, configures them, and provides the functions to grab frames and the routines to change its parameters. There are implemented two different grabbing functionalities: oneShot and continuousShot. This first one is used during single shot operation and only one frame is grabbed and saved into an internal image class. On the other hand, the continuousShot is used during Continuous operation or Live View, and it is continuously grabbing frames and saving them into a buffer of images. Once a frame is grabbed, a signal is sent and collected by the main thread which updates the User Interface with the last grabbed image.
Furthermore, the following parameters of the cameras can be modified from the PUREFORMULA user interface:
• Gain
• Integration time (#s)
The user interface application creates as many camera Provider objects as cameras are connected to the laptop and it is periodically checking if the number of cameras configured by the application is the same as attached cameras to the computer.
Image Analysis
The analysis of the images with the particles is carried out by the PFparticleSizer project that generates a win32 exe file, PFparticleSizer.exe which is called by the main project in order to do the analysis.
The analysis is done simultaneously for both images, and the results, particle average diameter, are combined in a histogram. The class psAnalysis that can be found on the file ParticleSizer.cs contains the code that calls and passes the parameters to the executable file.
Image analysis win32 project
The image analysis algorithms were developed and tested against standard samples, by DIT in R. This project closely follows the analysis that DIT has developed in R, but using C++ source code in place of the R language functions which were used previously, to increase processing speed. Some differences were introduced to assist in speeding up the process:
• A different approach has been taken to calculate the centres and the diameter of the particles. Instead of using the R morphology function, a function called distmap is used, which gives different intensity values depending on the distance to the edges of the particle. The centre is the pixel with highest intensity and the edge is defined by the pixels with the lowest intensity. The diameter assigned to each particle is the mean of all the distances from the centre point to all the edge points.
The steps in the image analysis algorithm implemented in C++ are:
i. Acquire grey-scale images from cameras.
ii. Apply a median filter, to eliminate dead and hot pixels (single pixels with extremely high or low values due to a nonlinear optoelectronic response in the pixel itself).
iii. Set a threshold of particles against background.
iv. Binarise images (black and white).
v. Segment particles using watershed algorithm:
a. Includes generation of distmap.
vi. Remove particles in contact with border of image.
vii. Label the particles numerically.
viii. Particle size estimation based on distmap.

Particle sizer histogram and statistics
The analysis of the images from each camera is carried out separately, but the results should be merged to provide a complete picture of the particle size range of the product, in the form of a histogram, with a corresponding mean and standard deviation of particle size. In order to merge the results, algorithms are coded in C++ and the results are displayed on the Analysis window of the User Interface.
Histogram bins
Bin widths have selected to avoid empty bins and to generate a picture that conveys the distribution of particle sizes in a meaningful way.

Histogram frequency values scaled according to Bin width
To combine the two populations in a single histogram, the frequency value of each macro camera bin is divided by 4 (macro bin width=100/ micro bin width=25). Then the height of each 100um wide macro bin is the same as it would be if it were divided into 4x 25um wide micro bins. There are two other bins that need their frequency values to be scaled. The 250-300 bin frequency needs to be divided by 2 as the bin width is 50 in this case, which is half of the normal width 100. The 10-25um bin value needs to be divided by 0.6 (25-10=15 divided by 25 = 0.6).

The micro and macro cameras count particles in sample volumes that are different sizes. To compare counts between the two cameras, the numbers counted are converted to the equivalent number of particles that would be found in 1cm3.

Histogram Count Units - Volume of solids per cm3
The analysis software generates a simple numerical count of particles. To compare the results across a large particle size range, it makes more sense to compare the volume of particles. In this way, 1000 particles of 100um diameter is the same volume of PIF as 1 particle of 1000um diameter. The volume of the particles in a bin is calculated using the mid-diameter for that bin, so for the bin from 150 to 175um, the particles are said to be 162um in diameter.
Volume of solids per cm3
= Total Volume of all particles detected in a cubic centimeter
= number of particles in a cubic centimetre x volume of each particle
= number of particles in 1 cm3 x (4/3)#(Diam/2)3 where Diam is the diameter value of the histogram bin.
Size Statistics – Mean and Standard Deviation
The mean and standard deviation particle diameter are calculated based on the volume of particles of different sizes. The following algorithms are coded in C++ and the results are displayed on the Analysis window of the User Interface.
The mean is calculated with this formula: =##
With the following steps.
• Divide the full histogram into 25um intervals (include the first narrow bin [10-25um] as one of these). Generate a table with the first column being the mid-value of the 25um intervals and the second column being the Solid Volume per cm3 per 25um interval, for each of these intervals.
The standard deviation is calculated with the following formula: =##[(#)2#]
With the following steps.
• Diameter minus Mean Diameter for each entry

The User Interface consists of two windows, a Main window that appears once the program is started and an Analysis window that appears when any analysis has finished and the results have to be shown. Once the program has been started the main window will appear in the middle of the screen. The Analysis window can be opened using the “Analysis” button in the main window, and also opens automatically once data has been acquired and analysed to show the updated results

The PureFormula main window is divided into four parts according to function: the particle sizer, the Multipoint NIR and the common part, with a fourth part made up by the header with the prototype name. The main window is the part of the program that controls the devices and allows the user to configure different parameters for both devices (ParticleSizer and Multipoint NIR).

Common bar.
The common bar has buttons and functionalities that affect both systems, the particle sizer and the multipoint NIR Single Shot. Button that performs a single acquisition of spectra and images, analyzing them if the analysis is selected.
• Continuous. Button that selects to perform a continuous acquisition of spectra and images, and to analyse the images if it is selected to do so.
• Interval. Time interval between successive shots, to acquire and analyse spectra and images when the continuous button is pressed.
• Batch. A text box where the batch number of the sample to analyse can be introduced.
• Date. Current date and time.
• Analysis. This button will open the analysis window if it is currently closed and bring it to the front if it is already opened.
• Log. This button will open a dialog to select a file to open. The file can be of type .nir or .psz which contain the results of previous analyses of particle size and chemical composition respectively, and it will show these results in the analysis window. Several files can be selected to load at the same time.
Particle Sizer
This is the section of the user interface where the particle sizer is controlled and configured. Here can be changed the configuration of the cameras, the LEDs and the image (display, area of interest, pixel size etc.).
Camera Displays
There are two displays showing the current images captured by each camera, the Macro and Micro. There is the possibility of enlarging the display image of either of these by clicking on one of the arrows. This action will open a new window where the image at its original size was opened, using a scale of 1 display pixel per image pixel. This window has been created to fit in the analysis window, while still allowing the user to have access to the camera controls, including gain, exposure and focus. This large window can be further enlarged to a window that is the original width of the image, and retains the 1:1 pixel representation. This is done by clicking on the same type of dual arrows in the top right hand corner of the large image display.
There are two options to close this new window:
• Right click on the new window arrow.
• Click on the arrows shown on the new window.
The triangle symbol on the left part of the displays is a play symbol, which when pressed causes the displays to show the images that are being acquired by the camera. The symbol then changes to a stop symbol while the images are being acquired and refreshed. And when this is pressed, the display will stop refreshing with new images acquired by the camera.
An area can be selected on the Camera Display for analysis, either with the mouse or by introducing the desired coordinates on the AOI (area of interest) section. If an area of interest is selected a white edged rectangle corresponding to it was seen on the camera display.
LEDs pulses configuration
The LEDs on the ParticleSizer device can be configured to obtain better image quality for the analysis. This part of the User Interface allows the user to change the configuration of the LED pulses.
To do so there are some sliders that allow configuration of the LED pulses by moving the sliders or by clicking the arrows on the text-boxes arrows or inserting directly the value in the text-boxes. The configuration is done in terms of:
• Delay – time interval between beginning of camera image acquisition and LED pulse. The LED pulse should occur within the image acquisition of the camera.
• Width – duration of the LED pulse
• Interval – time between successive LED pulses
This configuration is done in terms of microseconds for the delay and the width, and milliseconds for the interval.
LEDs power configuration
This slider allows configuration of the LED power of the system. As was the case for the LED pulse configuration, the values can be changed either by moving the sliders or the arrows or by introducing the desired value directly in the text-box. This configuration is done in terms of percentage.
Pixel Size
These sliders allow configuration of the scale of the pixels in the camera sensor arrays. This configuration is done in terms of micrometers, and represents the real-world size of a sample object that fills one pixel of the macro or micro camera. This value is used to convert the size of objects calculated by the image analysis software as number of pixels into real lengths in micrometers. These values should only be adjusted if the imaging optics or the cameras themselves are changed.
Inside the area of interest section are included different controllers to control the cameras and the displays. The main functionality is the area of interest:
• This selector allows configuration of the AOI (Area of Interest) that has to be analysed. The AOI is a rectangle which can be configured by setting the position of the upper left corner and the width and the height. A white edged rectangle will show the selected area in the camera display.
The AOI can also be selected using the mouse, by clicking a point on the display and dragging diagonally to extend the white rectangle.
• When the Area of Interest is defined, the analysis is applied only to the particles contained within this area.
• The other configurations placed in this section are:
• Camera selector combobox. This combobox allows the user to select the camera that is going to be configured.
• Do analysis checkbox. The do analysis check-box allows the user to select if the particle sizer analysis should be carried out when the single shot or continuous buttons are pressed. If not selected, the analysis won't be done.
• Save Image. To select if we want to save the images for a posterior analysis.
• Com Port combobox. Communication port where the led controller is connected.
Camera gain configuration
These sliders allow configuration of the gain and integration time of the selected camera (selected in the combobox. The gain is a dimensionless number, and the integration time represents the duration in microseconds during which the sensor acquires a single image. The values can be changed either by moving the sliders or by introducing the desired value directly in the text-box
The focal position of the optics is adjusted by movement of a motorised translation stage inside the prototype. The lenses themselves are fixed relative to the cameras, but the focal plan is moved relative to the front of the particle-sizer, similar to an optical microscope stage. This allows the user to ‘see’ particles clearly at different depths within a process vessel. The range of the stage is 20 mm, and the minimum step is 1 #m. The position of the stage can be adjusted by pressing the forward (FWD) or reverse (REV) arrows, either continuously, which will cause the stage to move at its fastest speed, or with a single click, which will move the stage by the minimum, 1 #m. The stage can also be moved by entering the desired position in the text-box, in millimeters. The position of the stage is indicated by the colouring of the bar at the right hand side and by the number in the text box
Multipoint NIR
This is the main area used to show the Multipoint NIR spectra and to allow the user to change the configuration, control the Multieye device, select the different analyses to perform and change the type of spectra displayed (intensity, absorbance, bright reference etc.).
Type of analysis
This is the section of the Multipoint NIR where the user can select the type of chemometric analysis to apply and the spectral information to display.
The section is divided in three parts:
• Analysis type selection. With 5 different types of analysis to select, corresponding to 5 different process conditions, plus a No analysis option to allow the user not to perform any analysis.
• Probes. To select the probes to display or hide from the Multieye display. The Multipoint NIR spectrometer has four channels and it is possible to attach up to 4 optical fibre probes at the same time. If some of the probes are not attached or their data is not required at the present time, they can be removed from the display. For legacy reasons related to the physical fibre optic cable ports on the Multieye hardware, the numbering of the probe channels is 2, 3, 6, 7, which has not been changed.
Display mode. Select the information displayed on the Multieye display from the list below:
• Intensity.
• Absorbance.
• Reflectance.
• Bright Reference.
• Dark Reference.
To be able to generate Absorbance and Reflectance spectra, a bright reference and dark reference have to be acquired as well as at least one spectrum measurement from a sample.
If an analysis is selected and the “Continuous” button is pressed the acquired spectra was shown on the Multipoint NIR display, and the selected analysis was applied to the measured spectra. When the analysis has finished the Analysis Window will appear with the results, if No Analysis is selected, then the spectra will still be displayed but no analysis was carried out and no results will appear in the Analysis window.
Every time that an analysis is done the results as well as the spectra (bright reference, dark reference and spectra intensity) are saved in a file that is date-time-stamped.
Wavelength configuration
These selectors allow configuration of the Wavelength range and steps of the spectra to be acquired. The configuration is done in terms of:
• Range: initial and final values of the range
• Step
This configuration is done in terms of nanometers for the range and steps.
Sampling configuration
Inside this section we have different features about the Multipoint NIR:
• Samples/Wavelength. This selector allow configuration of the number of samples per wavelength. A spectrum is acquired by the multipoint NIR sequentially as a series of acquisitions at each wavelength defined by the range and step size. To improve Signal to Noise Ratio, at each wavelength step the signal is sampled up to several thousand times and the average is calculated.
• Sampling Frequency. This selector allows configuration of the sampling frequency, which is related to the sampling time.
• Insert Comment.
• Offline Analysis
Offline Analysis.
Any previously saved data can be re-loaded for analysis when the instrument is not acquiring data. Clicking this button will open a dialog where the user can select the file to apply the offline analysis to. In the case of the chemometric analysis make sure to select the analysis to perform.
• For the particle analysis the user can select image files saved as bmp, png and jpg.
• For the chemometric analysis the user can select a text file with the structure:
• Header with the wavelengths
• Lines with the absorbance for every probe.
• Or a .nir file.
In this section six different buttons with their corresponding indicators regarding the spectral acquisition can be found.
The LED indicators on the right of every button indicate if the spectra to which the button refers has been acquired or not, where red means not acquired, green acquired
and pink acquiring.
• Dark Ref. The button will acquire the dark reference for the four probes.
• Bright Ref. The button will acquire the bright reference for the four probes.
• Scan status. Scan status will acquire the measurement spectra for the four probes.
• Close Com-Open Com. This will open and close the connection with the Multipoint NIR. In this case the LED was red for closed and green for open.
• Load Bright. This will open a dialog where a .nir file can be selected, and the first bright reference stored in this .nir file was loaded by the program.
• Load Dark. This will open a dialog where a .nir file can be selected, and the first dark reference stored in the file was loaded by the program.
Multipoint NIR display
This display shows the spectra that are being acquired by the Multipoint NIR probes, and the spectra that are stored in a Multipoint NIR file if this type of file has been loaded using the load button on the common section of the program. The type of spectra that are displayed are selected using the “Display mode” combobox.
Analysis Window.
In the Analysis window the results from the analysis are shown, either for the particle sizer or the Multipoint NIR or both systems at the same time. The Analysis window is divided in two sections – one for the ParticleSizer and the other for the Multipoint NIR.
On the top part of both sections can be found some information about the results that are being displayed in the corresponding section. This information is the batch and the date and time when the results were obtained.
The batch and the date can be different in each of the two sections since the user can load independently .psz files and .nir files. If the results displayed are of the analysis that is being carried out at the moment, then the batch and date will match, although the time cannot match as several chemometrical analyses will usually be taken during the time of one particle analysis as the image processing takes much longer.
Interactive trendlines
A function that is available in both sections is that when the user double clicks on the trendline the results matching the clicked point was displayed. On the particle sizer display this will mean that the histogram will change to show the one corresponding to the clicked point, and on the NIR display this will mean that the percentage values of constituents will update to the values corresponding to the clicked point. When doing this the batch and date will update as well, to match the selected point.
Particle sizer Analysis
The images acquired during a measurement session were analysed automatically in real-time as they are acquired. The default analysis routine will generate and display on the particle sizer side of the analysis window the following information:
• A trend-line of the mean particle diameter and the standard deviation added and subtracted from the mean.
• A histogram of the diameters of particles from the most recently acquired images.
• Diameter Statistics for the latest images, including:
• Number of particles in the micro and macro images.
• A statistical analysis of the diameters of the particles, showing Mean, Standard Deviation, maximum and minimum.
• Coordinates of the centre coordinates of each particle.
• The diameter in microns of each particle.
Multipoint NIR Analysis
The results of the analysed spectra were shown on the Chemometric results section of the analysis window. Four separate displays show a trendline graph for each of the four probes, displaying the results of the analysis for every constituent as a percentage value.
Below this are shown the precise percentage values for the constituents, analysed for each probe from the last acquired spectra.
The PureFormula software saves the analysis results automatically in a separate file for each type of measurement, which are created once the Single Shot or Continuous buttons were clicked and the type of analysis has been selected, and which are filled after the analysis has been completed.
These files are stored in the path: C: \ PureFormula\Results. The application generates two types of files, *.nir files for Multipoint NIR data and *.psz files for ParticleSizer data. Both are text files that can be opened with a text editor.
The file name is the date and time when the file was created with the format: yyyyMMdd-hhmmss, preceded by the batch name-number that the user provides in the batch text box on the user interface.
In order to facilitate the task of finding the desired stored data, every file is saved in the above mentioned directory in a hierarchical folder system based on time. The software automatically creates extra folders with the year, year-month, year-month-day structure, like the example: C:\ PureFormula\Results \2014\2014-11\2014-11-01\
If a continuous acquisition has been selected, then all of the NIR data and analysis results for that acquisition are stored in a single .nir file, which is time-stamped for thestart of the acquisition.
For particle-sizing on the other hand, a separate .psz file is generated for each image acquired during the continuous acquisition.
Multipoint NIR file structure Each row of the *.nir files includes the following fields separated by a semi-colon ”;”
• sw_version: Indicates the software version.
• probe_id: Indicates the probe identifier.
• batch_id: Indicates the batch identifier.
• datetime: Indicates the date and time of measurement.
• moisture_content: Indicates the moisture content predicted.
• Lactose_content: Indicates de Lactose content predicted.
• Protein_content: Indicates de Protein content predicted.
• Fat_content: Indicates de Fat content predicted
• range_start: Indicates the start of the wavelength range.
• range_end: Indicates the end of the wavelength range.
• range_step: Indicates the step of the wavelength range.
• length_spectrum: Indicates the length of the intensity vector.
• spectra[0];...spectra[N]; Indicates the intensity values.
• length_dark_ref: Indicates the length of the dark reference vector.
• dark_ref[0];...dark_ref[N]; Indicates the dark reference values.
• length_bright_ref: Indicates the length of the bright reference vector.
• bright_ref[0];...bright_ref[N]; Indicates the bright reference values.

Particle Sizer file structure
The internal structure of a .psz file is divided in three blocks. First of all, can be found the header (first row) with the next information separated by a semi-colon:
• sw_version: Indicates the software version.
• sw_code: Indicates a software code
• datetime: Indicates the date and time of measurement.
• batch_id: Indicates the batch identifier.
• num_of_particles: Indicates the number of particles detected.
• pixel_size: Indicates the pixel size set in the acquisition

The second block belongs to information related with the particles detected. Its first row (the header) indicates the parameters calculated, and the next rows belongs to each detected particle. Each row shows all of the next calculated values separated by a semi-colon:
• area: Indicates the area calculated of the particle
• perimeter: Indicates the perimeter calculated of the particle
• diameter_mean: Indicates the mean of the diameter calculation.
• diameter_sd: Indicates the standard deviation of the diameter calculation.
• diameter_min: Indicates the minimum value of the diameter calculation.
• diameter_max: Indicates the maximum value of the diameter calculation.
• center_x: Indicates x coordinate of the center of the particle.
• center_y: Indicates y coordinate of the center of the particle

The third and last block is the image itself. The first line contains:
• keyword: It used to identify the start of the image block.
• Aoi_height: Indicates the height of the area of interest selected.
• Aoi_width: Indicates the width of the area of interest selected.
The next line of this third block contains the RGB information of each pixel from the source image encoded as a byte.

The prototype system including hardware and software were built and integrated and then tested in industrial environments. The objective of testing in industrial environments was to challenge the PureFormula technology to determine it’s ability to work as it is designed to do. Industrial trials were carried out at both Dairygold and Carbery. The PureFormula technology was integrated in-line in the manufacturing process after the spray dryer. These sites and processes were chosen as they offered the best opportunity to really challenge the technology in an industrial environment.
Dairygold supplies dairy nutritionals made from milk whey and casein to the global marketplace, including infant formulae and protein powders. Their already nutrient rich raw milk is converted into top quality ingredients that will nourish the growing world. Their mission is to nourish people around the globe with their naturally-sourced gold standard dairy ingredients from the most fertile farmlands in the world.
Carbery are global leaders in the manufacture of value added food ingredients, flavours and cheese. Carbery are a member of the Dairy Processing Technology Centre (DPTC). DPTC is an industry–academic collaborative research centre, hosted by the University of Limerick, Ireland, with a research agenda driven by the long-term growth opportunities for the dairy sector created by the removal of milk quotas in 2015.Initial testing was performed off-line in the laboratory before being tested in-line in industrial settings. NIR spectra, particle size and shape were determined by the Pureformula NIR-particle characterizer PureFormula Technology prototype, which is equipped with two cameras, microcamera and macrocamera and a multipoint NIR spectrophotometer with 2 measurement channels and 2 NIR probes attached to collimators. These features allow the prototype to measure particles of size ranging 2 – 3000μm as well as their chemical properties, such as preteins, fats, carbohydrates and moisture content at the same time.
Bench-top analysis
Tests were initially performed to assess the ability of the prototype to quantify and characterise particle properties. The bench-top unit will be employed to image particles of defined chemical characteristics. Well characterised milk powders will be employed to evaluate the system under controlled static bench-top conditions.
Whole milk powder and skimmed milk powder of different gross composition were acquired from Dairygold.
Results generated from the activities
Bench-top analysis
Well characterised milk powders were employed to evaluate the system under controlled static bench-top conditions. A number of different batches of skimmed milk powder with different gross composition were supplied by Dairygold for bench-top analysis of the performance of the PureFormula prototype.
Samples were analysed with NIR. Triplicates of each bag were used for the analysis. Samples were placed in a single-use aluminum dish and its surface was flattened and placed directly against the protective glass of the PureFormula Technology in order to mimic the most possible real scenario in a production line where powders in free fall would be in direct contact with the protective glass. Samples were scanned at 10 different locations in triplicate in the range from 1520 nm to 2140 nm with a step size of 5 nm. Before analysis, the system was calibrated with dark current by occluding the light coming from the lamp with the help of a shutter; followed by white reference measurements by recording the spectrum of a 99 % reflectance standard (FW-WCVisNIR-O2, Avian technologies LLC, Sunapee). Raw data was pre-processed by Standard Normal Variate (SNV) and Savitzky-Golay smoothing was also applied – refer below spectra. From the plot, the following absorption bands are identified: band at 1728 nm related to fat, 1940nm related to water and 2055 nm related to proteins (Khodabux et al.,2007, Corredor et al., 2011, Wu et al., 2008, Gombás et al., 2003, Siesler et al., 2008). As expected, these characteristic bands change with composition of the different samples.
Partial Least Square (PLS) was applied on the full spectral data to build a PLS model to predict fat, moisture and protein content in the mixtures. After PLS treatment, cross validation with “leaving-one-out” option was applied. The cross-validated root mean squared of error prediction (RMSEP) and the determination coefficient (R2) was obtained as a function of number of factors or components for the calibration model. Variations in the RMSEP were small for number of factors n=8. Coefficients of determination R2 of 0.918 0.859 and 0.806 were obtained for fat, moisture and protein content respectively, indicating the models provided a good fit to the data analysed.
Particle size analysis
Well characterised milk powders were employed to evaluate the system under controlled static bench-top conditions. A sample of whole milk powder (WMP) was supplied by Dairygold for bench-top analysis of the performance of the Pureformula Technology prototype. The samples were tested in triplicate for particle size analysis. The whole milk powder was spread evenly in a thin layer into a black surface in order to act as a background to better differentiate individual particles. The sample was placed directly against the protective glass of the PureFormula Technology. Before analysis, the macro camera and micro camera gain and integration time were individually adjusted in order to obtain a focused image from both cameras.
An image with the macro camera and the micro camera was taken per triplicate and image processing was performed with the help of scripts built and executed in R statistical computing package by the DIT team. Results of the image analysis performed to determine the particle size of the powders using the macro and micro images separately. The mean particle size diameter calculated from the micro images was 48.8 μm, whereas the mean particle size diameter calculated from the macro images was 228.8 μm.
System set up in DFI
The PureFormula Technology was installed in the partner facilities of DFI (Ireland) directly after spray drying of concentrated skim milk powders. The PureFormula Technology was mounted to a wall to ensure secure fixation and attached in horizontal position to plastic tubing connected directly after the spray drying process, monitoring the skimmed milk powder produced on free fall. A constant vibration was applied to facilitate the skimmed milk powder to move through the whole process in the production line.
In-line testing in DFI
The process commenced in continuous mode and ran for approximately 14 hours. The PureFormula technology ran without issue for the duration of the process run. The measurement was stopped in line with the completion of the process run. This run demonstrated the robustness of both the hardware and the software as it is expected that the technology would have continued to perform over a much longer duration.
Image acquisition was configured in order to obtain a macro and micro image every 1.16 seconds approximately, producing a total of 14275 images in each case (macro and micro). Before analysis, the macro camera and micro camera gain and integration time were individually adjusted in order to set the focus of both cameras directly against the protective glass of the Pureformula prototype.
The NIR component of the PureFormula technology was configured in order to obtain a spectrum every 10 seconds in the range from 1520 nm to 2140 nm with a step size of 5 nm, producing a total of 5085 spectra per probe (probe no 2 and probe no 7 were active). Before analysis, the system was calibrated with dark current by occluding the light coming from the lamp with the help of a shutter; followed by white reference measurements by recording the spectrum of a 99 % reflectance standard (FWWCVisNIR-O2, Avian technologies LLC, Sunapee). Raw data was pre-processed by Standard Normal Variate (SNV) and Savitzky-Golay smoothing was also applied.
Parallel trials involving off-line analytical methods were performed every 1.5 hours in order to monitor any changes in the composition of the in-process material. The results obtained in terms of protein, fat, moisture and milk solids-no fat (MSNF) over the duration of the process run were recordrd. The composition of the powders over time was very stable and no major changes were detected.
NIR analysis
In order to compare the proximate analysis results obtained over time with the data collected by the prototype, NIR data was analysed on different timeframes every 1.5 hours. From the spectral signatures, the following absorption bands are studied: band at 1728 nm related to fat, 1940nm related to water and 2055 nm related to proteins (Khodabux et al., 2007, Corredor et al., 2011, Wu et al., 2008, Gombás et al., 2003, Siesler et al., 2008). However, only the moisture band at 1940 nm is prominently visible. No response is seen for the other components, fat, protein or MSNF, whose bands are less responsive that water.This suggests that the skim milk powder in free fall was not close enough to the protective glass and focus point of the NIR probes. The NIR system was therefore unable to detect any response for the protein, fat and MSNF levels. Analysis of the water band in the NIR spectra demonstrated no major changes in the NIR spectral response over the duration of the process run. This is consistent as no major changes in water content is observed. A resolution to this issue could be to integrate a deflection method to deflect a representative quantity of the in-process material closer to the NIR probe or to integrate the probe closer to the material flow.
The calibration models built in bench-top conditions (static) for moisture, protein and fat content were used to test the ability of the technology to predict the composition of the powders in free fall. Prediction vs measured plots of moisture, protein and fat content of the powders analysed in the DFI facilities were generated. As expected, based on the NIR response the results show poor predictions for the components, predicting higher values of moisture and lower values of protein and fat. These over- and under- estimation of the predicted values could be explained as only the moisture band at 1940 nm could be observed, whereas fat and protein bands were not visible. Results therefore show that calibration should ideally be produced in the same conditions as the intended validation. In this case it was not possible as calibration requires the use of samples with a different known compositional range in order to build a robust calibration and to take into consideration all limits in moisture, protein and fat content.
Particle size analysis
A selection of macro and micro images taken over the duration of the trial were compared. The images show both cameras were not well focused to the powders in free fall, obtaining on its majority images of poor quality. Dark images also suggest the pictures were either momentarily out of synchronisation with the LED lights, (poss
ibly due to the continuous vibration from the equipment during the trial) or more likely as there as no powder present in front of the camera when the image was captured. Due to the lack of good quality images, two of the best quality micro images were selected for image processing using the scripts built and executed in R statistical computing package by the DIT team. The images show a large range of particle sizes and shapes as well as agglomeration of multiple particles which would lead to a large range particle size distribution. Indeed individual particle size ranged from 1 μm to around 300 μm which was consistent with the range obtained from static measurements.
System set up in Carbery Food Ingredients Ltd.
The prototype was installed on a production line at Carbery (Ireland). The technology was integrated onto the production line after spray drying at the end of production of whey protein concentrate. The prototype was mounted to a tripod to ensure secure fixation and attached in horizontal position to a pipe connected directly after the spray drying process, monitoring the whey protein concentrate produced in free fall. A constant air flow was applied to facilitate the whey protein concentrate to move through the whole process in the production line.
In-line testing
Before analysis, the macro camera and micro camera gain and integration time were individually adjusted to optimise the focus of both cameras by directly placing against the protective glass of the prototype.
Image acquisition was configured in order to obtain a macro and micro image every 10 seconds approximately, producing a total of 663 images in each case (macro and micro). The NIR system was calibrated with dark current by occluding the light coming from the lamp with the help of a shutter; followed by white reference measurements by recording the spectrum of a 99 % reflectance standard (FW-WCVisNIR-O2, Avian technologies LLC
NIR analysis
Raw data was pre-processed by Standard Normal Variate (SNV) and Savitzky-Golay smoothing was also applied. NIR was also configured to obtain spectra at 10 second intervals in the range 1520 nm to 2140 nm with a step size of 5 nm. A total of 663 spectra per probe (probe 2 and probe 3 were active) were produced. The trial started in continuous mode for over 2.5 hours.
NIR data was analysed on different timeframes every 30 minutes. The NIR spectral signature of the whey protein concentrate powders analysed by probe 2 and probe 3 respectively were plotted. For comparison, the NIR spectral signature of whey protein concentrate powders analysed under bench top conditions were plotted. From the plots, the following absorption bands are visibly pronounced: lactose (1515 - 1600 nm), fat (1700 – 1765 nm) and protein (1980 - 2050 nm). However, no characteristic bands and spectral signature over time were observed, suggesting the whey protein concentrate powders on free fall were not close enough to the NIR probes to be detected. Another reason for the NIR results obtained could be that the NIR-particle characterizer PureFormula Technology was not in direct contact with the powders as the prototype was fixed to a glass window port that allowed the visualization of the powders in free fall, therefore the spectra observed could be the glass window being reflective to NIR.
Particle size analysis
A selection of macro and micro images taken over the duration of the trial were reviewed. The images show both cameras were not always well focused to the powders in free fall, obtaining on occasions images of poor quality, especially with the micro camera. In other occasions the powders accumulated in the glass impeding full visualization of the particles.
Nine of the better quality micro images were selected for image analysis with the help of scripts built and executed in R statistical computing package by the DIT team.
The images show a large range of particle sizes and shapes as well as agglomeration of individual particles which would lead to large particle size distribution.
The images also ranged from widely dispersed particles to highly populated images in order to take in to consideration all kind of scenarios. The number of particles detected per image ranged from approximately 50 to 350 particles, giving a good overview of the total average particle size for the whole production process. The particle size distribution in terms of both particle numbers and % volume for the selected 9 images respectively. Overall results obtained from the nine images processed showed the whey protein concentrate powders exhibited an average particle size of 0.000162 + 0.0001 mm3.
The testing of the Pureformula prototype in static conditions gave excellent results and showed potential to be used as an in-line measuring system. Good models were obtained for compositional analysis of skimmed milk powders in static conditions. Also, the Pureformula prototype data followed by image analysis was capable of estimating the particle size of the powder as well as capturing images through glass in static conditions. The testing of the particle characterizer prototype in in-line conditions showed potential to be implemented as a particle size measuring system. Image analysis was capable of determining the particle size of the powder as well as capturing images through glass in in-line conditions over the duration of the trial.
Further tests are required for process optimization, consistency and relevance of data generated and its potential use as a tool for quality improvements of powder characteristics. A calibration should be performed ideally in the same conditions as the ones encountered in in-line conditions in order to be able to validate further analysis
performed in-line.
It is suggested when possible, to place the prototype in a location in direct contact with the powders in order to increase the sensitivity of the NIR probes, allowing a better detection of the powders’ chemical attributes. Adjustments in the focusing of both cameras as well as the collimators to reach the powders at a further distance than the one established by the protective glass could be beneficial to obtain an improved NIR spectra as well as micro and macro images.

Potential Impact:
The results of this project will represent a significant advance of the state-of-the-art in analytical technology to characterise the physical and chemical properties of PIF agglomerates during processing and indeed the developed hybrid physical imaging and NIR chemical measurement device will be a worldwide breakthrough for the PIF, and indeed food granulation sectors. It is envisaged that by overcoming the limitations of the existing state-of-the-art in agglomeration control that a number of key innovations and technological progress will emerge from the research, which will have a major impact in equipping PIF processors with the knowledge they need in real-time in relation to the effects of the processing conditions on protein denaturation and thermal damage, to ensure these conditions can be better controlled to arrive at an optimum nutritional PIF profile. This will safeguard baby and infant health and wellbeing.
Powders are characterized in terms of size, shape, and their functionality, while there is a lack of knowledge about their behavior under varying temperatures and moisture contents.
The bulk properties of food powders are a function of physical and chemical properties of the material, the geometry, size, and surface characteristics of the individual particles, as well as the whole system. Parameters that determine the properties of agglomerates include those related to primary particles and agglomerates. Thus, the measurement of powder property is important because these properties intrinsically affect powder behavior during storage, handling, and processing11.
The current mechanism for determination of agglomeration end point is typically based on a pre-validated time period with fixed air flow rates, air temperatures and liquid feed rates. This defined process is on occasion supported by in-line fixed point NIR analysis to monitor moisture content but more usually by off-line moisture Loss on Drying instruments and off-line physical analysis using sieve analysis or laser based analytical instruments. The paradigm shift, offered by PURE-FORMULA, towards in-line determination of agglomeration endpoint through characterisation of physical and chemical parameters and enabling advanced process understanding will ensure greater process robustness, leaner manufacturing processes (saving time, energy and money) and ultimately patient safety. The PIF processor will gather significant understanding of the agglomeration process to facilitate new product development and formulation improvement. This is important as companies engaged in production of baby foods face cut throat competition, particularly on the innovation platform. Each company is focused on gaining a competitive edge over rivals by periodically introducing new products into the market. Improved understanding of the physical and chemical properties will enable a significantly reduced scale-up time during the commercialisation phase of the PIF products thereby introducing products to the market in a timelier manner.
In terms of a contribution to the technology progress, as an emerging technique, multipoint measurement provides both spatial and spectral information from an object. With the combination of the chemical selectivity of vibrational spectroscopy and the power of image visualisation, this technique enables a more complete description of ingredient concentration and distribution in heterogeneous systems.
The PUREFORMULA technology will provide an array of quantitative parameters of PIF, namely particle size distributions (meave, average, D50, D90 etc), particle shape factors (degree of Sphericity), blend homogeneity values (RSD) & moisture content values. Also, the system will employ data fusion of this physical and chemical information and correlate to the above discussed critical PIF parameters (dissolution, wetability, etc.). Fusion of the identified critical data under real-time process conditions will allow a data rich library of process variability and product quality criteria to be built up by processors.
VTT has developed several robust cost efficient and miniaturized spectrograph technology platforms for spectroscopic measurements, such as the following:
1. VTT patented miniature piezo-actuated Fabry-Perot tunable filter technology: this technology is commercialized by RIKOLA, gives more freedom for wavelength selection and still offers good wavelength resolution, low drift and high measurement speed. The benefits of these new devices compared to for example Acousto-Optic Tuneable filter (AOTF) or Liquid Crystal Tuneable Filter (LCTF) devices are their small size and weight, speed of wavelength tuning, high optical throughput, independence of polarization state of incoming light and the capability to record three wavelengths simultaneously.
2. By combining multi-channel detector technology with tunable Fabry-Perot filter VTT has developed and patented a quasi-imaging technique, to simultaneously increase the number of wavelength channels from multi-channel systems and still get the highest possible measurement speed. Advantages of quasi imaging, known wavelength channels, include fast data acquisition, flexibility of measurement, robustness, small size and low cost.
In all cases robustness of the measurement is extremely critical and wavelength drift must be minimized with hermetically sealed detector solution integrated to the PURE-FORMULA device.
The advance in knowledge that will be brought to the PIF sector and indeed to the wider powdered food and beverage sectors as a result of this development will be significant in relation to having real-time information about the impacts of the processing conditions on the nutritional and quality characteristics of the final product. Knowledge and understanding will enable optimal control and ultimately higher quality products that offer the desired nutritional value. Such an advance for PIF processing is priceless considering that baby and infant health and wellbeing way well be totally dependent.

The expected impact from the dissemination strategy was to generate a public awareness of the project, its progress, its results and potential benefits.

The project communication leveraged the commercial opportunities and efficiency gains provided by a novel and unique technology that will enhance process development, technical transfer and manufacturing processes.

It is also important to emphasise during the communication activities the significant opportunity provided by mechanisms such as FP7, H2020 and the research programs for SMEs.

Dissemination strategies

The strategy for dissemination of the results, findings and benefits of the Pure Formula project followed an agreed pre-determined communication plan. The unique and proprietary nature of the technologies involved in the design and building of the detection device have to be protected until full exploitation can be realised.

To ensure protection of the early research results, external communication focused on the project website and industry visits and consultations with a specific focus on those companies in the PIF and related industries who have a genuine interest and need for the inclusion of PAT in typical PIF processes.

The project leaflet and poster were preliminary designed with the printing of the leaflet targeted to include some images and non-proprietary results for WP4, “design and building of the Pure Formula system.

Note: As the project moved through each of the development phases associated with the construction and testing of a pre-competitive prototype portable test rig, communication and dissemination of the results and the strategy associated with the communication plan, was agreed by all members of the consortium to protect and maximise the opportunity to exploit the Pure Formula technology.

Completed activities

Development and upkeep of the Pure Formula website.

Develop a link to the PureFormula website on the Innopharmalabs technology website

Development of PureFormula powerpoint presentation template.

Generation of poster presentations X 3

Generation of PureFormula project Leaflet

Manuscript Published in Talanta which is a peer-reviewed scientific journal in pure and applied analytical chemistry. It was established in 1958 and is published by Elsevier, with 15 issues per year. In addition to original research articles, Talanta also publishes review articles and short communications.

Dissemination to a wider audience at the Alimentaria international food and drink exhibition in Gran Vía Venues in Barcelona. The Alimentaria international food and drink exhibition is the landmark international event for all professionals in the food, drink and food service industry and represents an unmissable date with innovation, the latest trends and the internationalization of the sector. The tradeshow brings together the most prominent aspects in the industry, attracting the main operators in the market and displaying the latest trends and innovations in RDI in the food industry, as well as promoting commercial networking and work meetings for international buyers in order to generate business opportunities.
Its trade fair model is based on permanent innovation and the creation of disruptive spaces that provide knowledge and ideas for the challenges currently facing food, hospitality and gastronomic tourism companies.

Discussions with the key contributors to the dairy industry centre of excellence in Ireland. The dairy food centre of excellence was formally launched in Dublin in January 2015, with the responsibility for PAT evaluation lying with UCD (University College Dublin) and lead by Dr Colm O’Donnell.

Periodic communication through the SETBIR network.

Place Activity The Date of Publication Audience Size Audience Type
SETBİR Board Meeting 19-Dec-13 Dissemination of project developments Industry
SETBİR Board Meeting 12-Feb-14 Dissemination of project developments Industry
Rekolte Dergisi (Magazin) Publication 01-Feb-14 4000 Article published in the popular press* Industry
Tarım Türk Dergisi (Magazin) Publication 01-Feb-14 5000 Article published in the popular press** Industry
Dünya Gazetesi (Newspaper) Publication 01-Feb-14 10000 Article published in the popular press Civil Society
Career Days of Hacettepe
University Leaflet 06-Mar-14 20 Giving Leaflet about the project Civil Society
Hacettepe University, Ankara Mini conference 7-Mar-14 50 Mini Conference about EU Projects Civil Society
Drinktech Dergisi (Magazin) Publication 01-Apr-14 5000 Article published in the popular press Industry

• Conference talk at the IFPAC, USA – Jan 2015 Titled “Moisture & content determination of powdered infant formula in static and dynamic conditions using Multipoint NIR Spectroscopy”.

• Poster presentation at ICEF12 conference Quebec, June 2015 titled; Moisture content determination of powdered infant formula in static and dynamic conditions utilising a novel multiprobe near-infrared spectroscopy system

• Visits and communications to PIF and related industries, Academic and Development facilities across Europe and the US. This activity commenced with the generation of the industrial specifications, Deliverable 1.2. This included consultation with industry experts in Ireland where there is a government funded program in place to establish a centre of excellence for powdered milk and infant formula production. This centre of excellence focuses on adopting in line monitoring systems to provide real time data on the key attributes of powdered milk products. The site visits consisted of visits to the following companies:
Teagasc Moorpark: Research, Pilot and Commercial process applications both spray drying and Fluid Bed Granulation processes. Visited by Innopharmalabs.
Bahçivan: Powdered milk production using spray drying technologies. Visited by all consortium members.
DairyGold: Visited by Innopharmalabs.
UCD: Centre for food and agri-science facility. Visited by IRIS, DIT and Innopharmalabs.
Trinity College: Agri-science facility, laboratory scale spray drying systems. Visited by Innopharmalabs and DIT.
Alimentaria International food and drinks exhibition in Barcelona. Presentation of project and understanding of end user requirements: IRIS.
Survey circulated: key powdered formula users and contributors to the dairy centre of excellence, there were 6 respondents in total. While this does not seem like a significant number, there are only about a dozen producers globally of PIF so the 6 respondents is definitely representative of the industry. These respondents included DANONE and Nestle two of the bigger PIF producers.

• In conjunction with these activities members of the consortium attended a number of conferences/forum’s/workshop’s and exhibitions. These events offered a unique opportunity for the team to meet face to face with specific target groups in order to relate the findings, benefits and results of the PureFormula technology. As appropriate, the consortium members took these opportunities to deliver talks/presentations, exhibit and/or network with the specific target groups.

The following are the conferences and exhibitions that members attended:

• The IFPAC (International Foundation of Process Analytical Chemistry) conference took place in Washington DC, USA, in Jan 2015, 2016 & 2017. IFPAC has emerged as the pharmaceutical and food industry's premier event relating to the techniques of Industrial Process Analysis, Process Knowledge and Quality Control. IFPAC attendees come to learn more about how they can boost productivity, streamline implementation processes, enhance operational efficiency, and maximize their resources and therefore fit perfectly into the target audience category for PureFormula. Both Innopharma Labs and VTT exhibited at these conferences and DIT were in attendance. PureFormula awareness was generated through direct communication to attendees both at the exhibition booths & networking throughout the conference and through the distribution of project leaflets.

• The Pittcon, Conference & Expo, took place in New Orleans, LA USA on March 8-10th 2015. Pittcon is the world’s largest annual premier conference and exposition on laboratory science. This dynamic global event offers a unique opportunity to get a hands-on look at the latest innovations and to find solutions to all your laboratory challenges. The robust technical program offers the latest research in more than 2,000 technical presentations covering a diverse selection of methodologies and applications. Pittcon attracts more than 16,000 attendees from industry, academia and government from over 90 countries worldwide. PureFormula awareness was generated through direct communication with attendees both at the exhibition booth & networking throughout the conference and through the distribution of project leaflets.

• The ACHEMA Forum took place in Frankfurt, Germany from June 15-19th 2015. ACHEMA boasts attendance figures in excess of 166,000 with approximately 4,000 exhibitors. ACHEMA 2015 focused on three themes: Industrial water technology, Process analytical technology and Bio-based production. The extensive lecture programme provided information on new technological developments and trends. Developers and providers as well as users and plant carriers had opportunities to discuss focal themes and exchange ideas both during the exhibition and the events. Innopharma Labs attended this conference. PureFormula awareness was generated through direct communication with attendees both at the event & networking throughout the conference and through the distribution of project leaflets.

• The 7th 8th & 9th International Granulation workshops took place in Sheffield, UK. Granulation encompasses a number of processes vital to the food, pharmaceutical, chemical, mineral, metallurgical, fertiliser, waste water treatment and catalysts industries, among others. Growing interest in the field has led to the development of the International Granulation Workshop, which has become one of the world’s largest forum for industry and academia to share and discuss recent developments in this complex and evolving field. Innopharma Labs exhibited at these conferences and included information on the PureFormula project at their exhibition booth.

• POWTECH took place in April 2016. POWTECH, the leading exhibition for experts in powder and bulk solids technology. This event provides a presentation platform for innovations and advancements in processes for the manufacture of quality products made and processed from powder, granules, bulk solids and liquids – also for the environmental and recycling sector. Innopharma Labs exhibited at these conferences and included information on the PureFormula project at their exhibition booth.

• Powder and Solids handling Solutions conference took place in CPI, Durham, UK in July 2017. This event explored innovative solutions to current industry problems which arise during powder and solids handling. Delegates heard from a range of speakers who are working to overcome these issues as well as took a tour of CPI’s formulation facilities. Innopharma Labs delivered a talk at this conference.

• CRS annual meeting and Exposition took place in Boston in July 2017. CRS was incorporated in 1978 as a not-for-profit organization devoted to the science and technology of controlled release. Over time, interest in controlled release has grown and broadened in scope, due to the realization of the economic, therapeutic and social benefits that can be derived from controlled release technology. While 70% of attendees are from the pharmaceutical industry 30% are non-pharma, including food. Innopharma Labs delivered a talk at this conference.

List of Websites: