Skip to main content
European Commission logo print header

InteGration of pRocess and quAlity Control using multi-agEnt technology

Final Report Summary - GRACE (InteGration of pRocess and quAlity Control using multi-agEnt technology)

Executive Summary:
The main objective of the GRACE project (InteGration of pRocess and quAlity Control using multi-agEnt technology) is to conceive, study, develop, implement and validate a collaborative multi-agent system (MAS) which operates at all stages of a production line, integrating process control with quality control at local and global level.
For this purpose, a society of autonomous and cooperative agents representing the manufacturing components of a production line producing washing machines, was designed, realized and tested on-line.
The GRACE multi-agent system infra-structure was implemented by using the JADE (Java Agent DEvelopment Framework) framework and a specific ontology was developed. Each agent behavior is modeled through Petri nets formalism.

Focusing on the production of home appliances, specifically washing machines, the project team created a collaborative multi-agent architecture and demonstrated it in a real production line, during normal operation at full capacity. It featured mechatronic units with specific functions that cooperate to control the process and quality of products produced.
Four quality controls and three production processes were developed, integrated and tested on a real production environment: all exhibit self-adaptive behaviors at local level and are associated to resource agents.
A variety of self-adaptive quality control stations was developed. Three vision based quality control stations have been developed: they embody self-adaptable illumination for image enhancement, self-adaptable template generation for improved pattern matching and a set of features to improve confidence level on diagnosis upon variable production conditions. A self-adaptable gap inspection station, embodying an innovative conic vision system, was developed, alongside a vibration testing station with self-adaptation of the scanning laser Doppler vibrometer for minimizing uncertainty. Furthermore, a robot vision system able to reposition its camera searching randomly hidden parts has been tested in laboratory condition. Finally, researchers implemented product functional testing with adaptable test sequences.
The GRACE team developed also adaptive control schemes for screwing actions and bearing insertion, common in many production processes. It was also designed an adaptable on-board controller for the washing machine, an innovative concept demonstrating how process and quality control at the local and global level can refine the individual tuning of a product, based on its production history.
Besides systems architectures and technical prototypes, a special focus was set on the engineering methodology, analyzing and defining the steps required to implement multi-agent systems for process and quality control, bringing to evidence the importance of integration of mechatronic-units. As many European research projects manage a pilot implementation, but at the same time seem to fail real industrial applicability, focus on engineering methodology was defined to evaluate methods to improve industrial application.
Furthermore a new methodology for consideration of product quality within manufacturing and manufacturing systems engineering has been developed - the so called MPFQ model, based on its four main elements: Material (M), Production Processes (P); Product Functions / Features (F); Product Quality (Q). This method, through the measurement of performance indices, allows to relate each individual product quality to the complete production process performance, thus integrating control of quality and process at global level.
The project ended in a demonstration of probably the larger MAS ever run in an industrial environment during normal production; this was done in the Whirlpool factory in Naples, where about 1500 agents were set into operation in the real production line, interfaced to all the prototypes of self-adaptive process and quality controls that were developed throughout the project, and interfaced to a specific factory data base developed for the purpose and to the MPFQ model. The on-line demonstration of what developed constitutes the major element for future take-up of these concepts in manufacturing industry.

Project Context and Objectives:
Within European research, the development of flexible and intelligent production processes is often debated to transform European industry from a resource intensive to a sustainable knowledge-based industrial environment.
The main objective of the GRACE project is to conceive, study, develop, implement and validate a collaborative multi-agent system (MAS) which operates at all stages of a production line, integrating process control with quality control at local and global level.
The choice of the multi-agent systems paradigm is motivated by its intrinsic characteristics like decentralization of control, distributed autonomous and cooperative entities, modularity, flexibility, robustness, adaptation and re-configurability. Thus, it is possible to respond promptly and correctly to changes and disturbances in the production environment.

The following specific objectives were set at the beginning of the project:
Objective 1 - Development of a flexible, adaptive and reconfigurable architecture that integrates process and quality control.
Objective 2 - Development of self-adaptation and self-optimization mechanisms for process control at local and global level.
Objective 3 - Development of modular and adaptive quality control systems.
Objective 4 – Enable systematic transferability and industrial applicability of GRACE, by defining a proper engineering methodology.

The overall goal of the GRACE project was to perform an on-line demonstration of all the developments achieved throughout the project. Full success of the project consists therefore in the successful completion of Milestone MS5 “Implementation and validation”, which implies the assessment that “The GRACE multi-agent architecture has been operated on a WM production line and tested against planned variation of variables and unforeseen ones. Production efficiency has been measured and compared to normal performance.” Indeed this challenging goal has been successfully achieved.
The GRACE project, was financed within the Call NMP-2009-3.2-2 “Adaptive control systems for responsive factories”, which was specifically demanding for the “implementation of intelligent factories, able to manage complex and variant production processes”. The Call for Proposals stressed also on a specific feature that projects should have had: “In order to ensure industrial relevance and impact of the research effort, the active participation of industrial partners represents an added value to the activities and this will be reflected in the evaluation”.
The GRACE consortium therefore developed a proposal fully in line with the Call, which included a real demonstration on a real factory. This required strong motivation of all partners to develop the research up to the industrial application; indeed this is a very challenging task, both for academic partners and for industrial partners, because the interaction with a real production system, while it is operating and cannot be stopped, poses severe difficulties for prototype testing, which are not commonly encountered in laboratory activities. These difficulties have been overcome by the GRACE consortium through a continuous and full involvement of industrial partners in the research activities throughout all the project, which has led the consortium partners to become a real “team”. In the mean-time, academic partners where stimulated during all the project to consider and assume the perspective of the industrial OEM, so to develop a sensitivity to the practical problems of a real world implementation, which too often are underestimated by researchers, and erroneously considered too simple and not deserving adequate attention. Indeed we are convinced that this cultural attitude is often the main reason for the still existing gap between research and application. The GRACE partners were motivated to fill this gap and bring the project, as required by the Call, to a real demonstration. An enhanced inter-partner team work has been therefore the key for a successful demonstration phase.
During the first part of the GRACE project the partners addressed the question “How to develop a scientific RTD project up to an effective demonstration on a real production line ? ”. Of course it was not possible to consider all the processes of the line, therefore our approach has been to develop self-adaptive behaviours at local level on a sub-set of selected critical processes across the WM assembly line. They are:
• 3 manufacturing processes:
o Bearing insertion station;
o Screwing stations;
o On-board controller programming (this means that each product is individually tuned);
• 4 types of quality controls:
o Drum Geometry Control Station;
o Vision stations:
- Washing Unit Vision Station;
- Assembly Vision Station;
- Final Assembly Vision Station;
o Functional Testing.
The objective was to realize for each of the above listed processes and quality controls specific self-adaptive behaviours, which could enhance performances. While for process control self-adaptivity was designed to improve process control even in the variable conditions of the production line, for the quality control systems self-adaptivity was implemented in order to manage measurement uncertainty, that means to manage confidence level on the diagnostic output of a quality control station. Therefore quality control stations were designed as self-optimizing measurement systems, i.e. devices that are able to estimate measurement uncertainty in real time and react by modifying influencing parameters, so to reach a desired uncertainty; the level of desired uncertainty is made dependent on the production scenario, therefore realizing also a control at global level, which was one of the objectives set in the Call for Proposals.

Project Results:
The GRACE project has produced many results, some of which have an innovative character by themselves, while others are innovative if considered as an ensemble in terms of industrial application. They all are focused on the concepts of intelligent factory automation, taking into particular consideration the complementary aspects of process control and of quality control, considered here as inherently linked elements of a manufacturing plant, which jointly contribute to product quality and production efficiency.
In order to reach its results, the GRACE project has focused on two main pillars: self-adaptive control at local and global level and multi-agent systems.
The GRACE project has therefore developed and implemented self-adaptive process controls at local level, i.e. at single process level, as well as it has developed global controls at factory level, implementing also in this case self-adaptive behaviors designed to cope with the evolving scenarios of production. Of particular and specific relevance for product development, it has been the implementation of adaptivity also on the on-board controller that controls each washing machine during its lifecycle; it has been demonstrated how this controller can be programmed individually for each machine, taking into account its individual production history, thus realizing an individual tuning of each product, based on the manufacturing plant performance achieved during the production of that specific item being produced as well as its individual characteristics which have been forming through the production process. This realizes an intimate interaction of process control with product quality, thus demonstrating in practice the concepts of quality oriented production processes.
Self-adaptive behaviors have been designed and realized also for quality control systems, with the aim of improving the behavior of measurement systems performing measurements for quality controls. Self-adaptive behaviors at local level, i.e. at the level of a single quality control station, have been designed to keep under control measurement uncertainty, thus improving the confidence level of diagnosis; self-adaptive behaviors have also been integrated into the global control system operating at factory level, therefore adapting the operation of quality control stations depending on the needs emerging from the specific production scenario; one example is to redefine the measurement sequence or to change the algorithms or to vary the thresholds used for diagnosis, depending on the history of production of each individual item flowing on the assembly line. These quality control systems therefore adapt measurement uncertainty and overall performance to the requirements of the whole manufacturing plant emerging during production.
A key role to achieve these results has been played by the multi-agent system, which is indeed the backbone of the whole developments realized in GRACE; indeed the MAS acts as a backbone for the integration of process and quality control levels, providing intelligence and adaptation functionalities at high-control level of the production process.. To each process control and quality control station has been associated an agent; agents have also been associated to each product and integrated in a multi-agent system. The resulting MAS platform guarantees the achievement of several innovative quality properties, showed in the experimental tests; the most significant properties are:
• Modularity and flexibility
• Run-time and on the fly reconfiguration
• Distribution
• Robustness
• Responsiveness
• Scalability
• Smooth migration
At the end, probably one of the major contributions of the GRACE project is the full assessment of the multi-agent systems benefits in real industrial environments, by demonstrating the effective applicability of multi-agent systems in a real world case, considering the case of a factory operating with a production on-demand. The demonstration on-line has indeed been a challenging and difficult task; its successful completion testifies that what developed has a real potential for industrial take-up.
In order to enforce the chances for industrial development of MAS technologies joined to self-adaptive schemes for process and quality control, the GRACE project has devoted a relevant effort to define the engineering methodology needed for such a step.
The following paragraphs provide more details on what developed throughout the GRACE project; more technical information is also available in the Deliverables and the many scientific papers published during the project. Abstracts of all papers and deliverables are available at the web-site of the GRACE project , from which it is also possible to download all deliverables classified as public.

Integration of process and quality control levels by using multi-agent systems (MAS) principles has been reached developing a society of autonomous and cooperative agents representing the manufacturing components disposed along a production line producing washing machines. The following agents were defined in GRACE:
• Product Type Agents (PTA), which represent the catalogue of products that can be produced by the production line.
• Product Agents (PA), which handle the production of product instances along the production line (e.g. washing machines and drums).
• Resource Agents (RA), which represent the physical resources of the production line, such as robots, quality control stations and operators.
• Independent Meta Agents (IMA), which implement global supervisory control and optimized planning and decision-making mechanisms, e.g. defining and adapting global policies for the system.

In this approach, each distributed and autonomous agent only has a partial knowledge of the problem. Thus, interaction among agents is necessary to achieve global objectives. In consequence, the overall system behavior emerges from the cooperation among individual agents, each one contributing with its local behavior. In such distributed systems, the establishment of a common understanding among the agents is of special importance, since the exchange of shared knowledge becomes difficult if each agent has its own knowledge structures. The solution is to use proper techniques that guarantee the common understanding among distributed entities. The use of ontologies addresses this challenge by defining the vocabulary and the semantics of the knowledge used in the communication between distributed agents. For this purpose, an ontology was designed and implemented considering the particularities of the home appliance domain and the integration of process and quality control levels. This ontology formalizes the structure of the knowledge, namely the concepts, the predicates (relation between the concepts), the terms (attributes of each concept), and the meaning of each term (type of each attribute).
The GRACE multi-agent system infra-structure was implemented by using the JADE (Java Agent DEvelopment Framework) framework for the development of agent-based solutions, which provides an integrated environment for the development of such system, with less complexity and reduced effort and time-consumption.
The skeletons of the several GRACE agents were developed and implemented: namely the structure of behaviors for each agent, the ontology schema for the knowledge representation (using the Protégé framework), the interaction patterns supported by FIPA (Foundation for Intelligent Physical Agents) protocols, and the integration with legacy systems, such as LabView™ applications running in quality control stations and the production database.
Several Graphical User Interfaces (GUIs) were also implemented to support an easy interaction with the users in terms of administration, management and monitoring of the system. Each type of agent provides different GUIs, since each type handles a particular set of information, and allows different types of interactions with users. The use of a Java based framework to develop the multi-agent system, offers the possibility of using Swing, a well-established toolkit to implement GUIs for desktop applications. Each type of agent in the GRACE MAS has its own GUI implemented as an extension of the javax.swing.JFrame component.
The developed multi-agent system infrastructure was intensively tested using historical real production data from the GraDaCo database (a data base specifically developed for the GRACE project collecting data from the Whirlpool factory in Naples), aiming to test and correct mistakes and bugs during the development process. One of the main advantages of the GRACE MAS system is the possibility to run also in off-line mode, i.e. using the historical production data stored in GraDaCo to run simulations in a virtual environment. The MAS infra-structure accommodates the adaptation mechanisms, aiming to provide adaptation and optimization based on the integration of the quality and process control.
The design and implementation of the MAS platform were developed in WP1 (described in Deliverables 1.2 1.3 and 1.4). The integration of adaptation mechanisms into the MAS platform was performed in WP5.1 and described in Deliverable 5.1. Note that each agent comprise a local database, a graphical user interface (that can be launched at beginning or not) and a set of proper behaviours (inherited from its agent type), being customized according to a proper XML file.

The goal of the production process modeling is to enable self-adaptation and self-optimization, performed by resource agents along the production line, as well as in the on-board product agent. The main purpose is to improve the quality of operation and thereby the product quality. Two specific assembly processes have been in focus along the production line; the Screwing processes and the Bearing insertion process. Then the individual tuning of the on-board controller of each washing machine has also been developed.
Each agent related to the product will define and control a set of parameters to be stored on board the final product, so to adapt its controller to the individual characteristics of each machine, which are naturally dispersed, and in order to keep its performance within specifications even in case of assembly deviations. In the factory, a number of functional tests (FT) are performed on each machine, e.g. power consumption and water inlet. Naturally there can be differences in the functional test results depending for example on the upstream assembly processes. This can be handled by adaptation of on-board controller parameters based on information from the production line and the factory database.
The factory database contains history of the production processes enabling the use of correlation analysis to adapt specific machine parameters. Example of relevant parameters are prediction of water and energy consumption which have been demonstrated in GRACE.

A) Screwing process
Screwing processes are typical in production of appliances and the most common way of joining parts is either by use of machine or self-tapping screws utilizing automatic screw-drivers. The screwing profile is often rigid and not adapting to flexibility in the material and characteristics of each single screw which again may lead to variations in clamping force and tightening torque.
In GRACE a dynamic and adaptable screw profile generator has been developed which continuously adapts the screw profile based on on-line measurements as velocity and rounds per minute ensuring equal clamping force in each screw operation. Calculation of the adapt parameters is based on a mathematical model of the screw process.
The main task of the resource agent (RA) of the screwing process quality station is to evaluate the quality of the joining of the drum cross piece and the wrapper. The available measurements from this process are the final screw torque and angular rotation; these are the quality indicators of this process.

The screwing station is associated to a RA in the multi-agent system infrastructure that is responsible for managing the process operations performed by the machine. The integration of the screwing processes quality assessment and the MAS takes place according to the following sequence:
1) The virtual resource (timer based) notifies RA that a pallet is in front of the screwing station, passing the 12NC code (type: string) and the serial number (type: integer) of the washing machine
2) RA notifies the PA (message 1), passing the 12NC code (type: string) and the serial number (type: integer) of the washing machine. Also RA agent calculates reference indicators for quality evaluation based on historical data
3) After running an adaptation function, the PA agent sends the adapted operation parameters to the RA agent (message 2). After the execution of the operation, the process results are passed to the RA for evaluation (message 3). The results from this evaluation is then passed to the PA (message 4), namely:
i. Overall result, i.e. OK or KO;
ii. Detailed result, i.e. torque and depth values;
iii. Performance index.
It is important to notice that the direct connection between the RA agent and the physical screwing station is not implemented, because it is not possible in the Whirlpool factory to interact with the screwing station controller and modify it, but instead a connection to the GraDaCo database (via the Virtual Resource component) is used to gather the results from the real production execution.

B) Bearing insertion
The bearing insertion is an essential process in the production of washing machines. The bearings are an important part of the drive train in the washing machine unit, as they have to absorb all rotating forces and retain the drum rotating within the washing tub.
The process of inserting the bearings in the tub is considered since the structure of this process allows the implementation of different feedback loops. The insertion is carried out imposing a constant pressure to the pump; a number of parameters affect the quality of the operation as for example temperature of material and insertion velocity. In GRACE, adaptation of the force and insertion depth is performed within a resource agent.
The adaptation strategy is based on a mathematical model of the process. The main goal of the bearing insertion process model is to provide estimates of internal physical states and to calculate the quality of the assembled product. This information may be used to:
• Adjust the processes downstream of the bearing insertion cell.
• Adjust the functionality of the finished product related to some quality tolerances.
• Check if the tub hubs manufacturing process are precise and if the products are satisfying the design tolerances

C) Also the bearing station is associated to a RA in the multi-agent system infrastructure that is responsible for managing the process operations performed by the machine. It operates according the same scheme described for the screwing operation. On-board controller
Adaptive adjustment of the parameters of the on-board controller allows to reduce significantly dispersion of performance between products, taking into account the results of the tests performed during the assembly and at the end-of-line and to keep its performance within specifications even in case of assembly problems. In this way each final product is set and calibrated for its specific characteristics in order to reach maximum efficiency and optimize compliance to specifications. The MPFQ model developed in GRACE supports the task as well as the integration into the MAS.
For the demonstrator, it has been developed the individual setting of the control board parameter for the inlet flow valve. This parameter is used in the calculation of the time interval the valve is open, and is therefore crucial for controlling the water consumption of the washing machine. This subject has a large industrial relevance, therefore no details are given in this publishable report; detailed information is available in the deliverables of the project.

Quality control systems conceived as quality control agents (QCA) exhibit behaviors which realize self-optimization and self-adaptation. The purpose is to improve reliability on diagnostic information generated by the quality control system and a successful realization of this objective implies implementation of actions aimed to keep measurement uncertainty under control. Four types of quality control systems were realized in GRACE:
• Drum geometry control station;
• Vision inspection stations (4 variants);
• Vibration control station;
• Functional testing.
They are outlined hereafter and described in detail in the deliverables of the GRACE project.

A) Drum geometry control station
The Drum Geometry Quality Control Station is an automatic test bench installed in the Washing Unit (WU) line after the welding of the front and rear tub. It performs a quality control on the assembled Washing Unit (WU) and provides as outputs the measure of the gap between the front tub and the drum, and a Pass/Fail result according to predefined thresholds. The objective of this control station is to measure the final gap existing between the rotating drum and the front tub, in particular its seal, whose thickness depends both on the marriage process and on the geometry of the assembled components (basically the front and rear tubs and the drum). The width of this gap is important because if it is too large cloths may be trapped between the rotating part and the tub, damaging the appliance and the cloths, while if it is too little the drum may touch the rubber seal in the front tub, causing friction, noise, wear and an increase in energy consumption due to friction losses.
An innovative sensing unit has been developed for measuring the geometry of the gap over the whole circumferential direction, without any moving part or scanning device. It is a new camera based on a catadioptric optics, in particular a coaxial conic mirror aligned on-axis in front of the camera objective, which allows a single-shot vision over a cylindrical space over 360°. Once inserted axially inside the drum, the gap profile is measured by image processing.
This particular optical system requires a specific calibration and an accurate positioning inside the cylindrical drum. The station therefore compensates drum position variations by exhibiting a behavior which realigns the conic camera, through feed-back control of geometric parameters of the acquired image. Once the camera position is adapted, then image acquisition and processing provide the gap thickness profile. This station allowed the successful measurement of the gap over 100% production, within the uncertainty limits that were set.
The QCA and the quality control package, implemented as a LabVIEW application, interact by exchanging XML files using socket communication.
Then the system computes the final value of each quality indicator, and the QCA managing the Drum Geometry Control Station calculates a performance index using the data provided by the station and this information is used in the MPFQ model.

B) Vision inspection stations
Three different vision stations have been installed in different positions on the production line of the factory:
1. Washing Unit (WU) Vision Inspection Station: it is positioned at the end of Washing Unit Line before the marriage with the cabinet and performs the following checks: the position of the belt on the motor, belt thickness, the correct mounting of the heater, the presence of the clamp blocking the exhaust pipe.
2. Assembly (WMB) Vision Inspection Station: it is positioned in the assembly line before the assembly of the front panel and performs the following checks: the proper insertion of the door-lock connector, the correct mounting of pipes and clamps.
3. Final Assembly (WMA) Vision Inspection Station: it is positioned after the functional test and performs the following checks: the position of the belt on the pulley, the proper insertion of a connector, the correct mounting of pipes and clamps, the correct connection of the electrical system.
All these stations have been implemented developing self-x strategies and integrated into the MAS system. The vision inspection quality control stations have associated QCA agents (one for each station) in the multi-agent system infrastructure that is responsible for managing the inspection operations performed by the machine.
The integration procedure is similar to that described for the Drum Geometry Control station. Namely, each pair QCA and QCS (implemented as a LabVIEW application) interact by exchanging XML files using socket communication.
The self-x strategies which resulted more suited for the implementation on the on-line vision inspection stations are the lighting color adaptation (RGB illuminator) and the camera exposure time adaptation for the hardware part, and adaptive pre-processing, multi template and test plan customization for the software part. These components have been installed on the on-line vision stations and tested in a real production environment. The reduction of false-positives and wrong diagnosis has been very large.
Vision stations then compute quality indices and interact with the MPFQ model as well. For example, the inspection that gives the major content of information to the MPFQ model is the one related to the position of the belt (and its thickness). The proper position of the belt is important in order to guarantee a good transmission of the movement from the motor to the pulley that rotates the drum. The position of the belt can be calculated both on the WU and the WMA station: in the WU the position of the belt is referred to the motor, while in the WMA the position of the belt is referred to the pulley.
Furthermore, a solution that takes advantage of a robotic manipulator in order to move the camera has been developed and tested but only in laboratory, due to the difficulties to install the robotic manipulator on the production line in operation. Therefore only some of the concepts developed for the laboratory robot vision prototype have been brought to the Whirlpool’s vision inspection stations.

The robotized vision inspection station is designed to verify the presence and correct mounting of parts of the washing unit and of the complete washing machine by image processing and exploits state-of-art matching algorithms. It implements self-adaptive strategies which allow to react to changes in the scenario so to keep image quality at the desired level, and to reduce false diagnosis:
• A six degrees of freedom anthropomorphic robot arm displaces the camera and the illuminator at specific locations with the purpose to re-position the sensor and adapt to changes in product lay-out, position and geometry.
• It implements automatic exposure time to maximize a given image quality indicator.
• It uses controllable illumination systems designed to compensate changes of the environmental light conditions and/or of the characteristics of the surfaces in order to optimize a given image quality indicator by:
o variable intensity/color (diffuse light ring LED RGB illuminators);
o illumination system with capacity to project programmable structured light in space and time (Digital Light Projectors DLP).
• It implements adaptive software algorithms for image processing suited to deal with varying conditions, such as multi-template feature recognition.

C) Vibration control station
The vibration control station is based on a scanning laser Doppler vibrometer which implements optimization strategies for signal-to-noise improvement, therefore minimizing uncertainty in the computation of diagnostic features. This implies searching for a point of maximum light scattering over the rough surface of the target and it is achieved by controlling the position of the laser beam at sub-millimetric scale by scanning mirrors. The same scanning mirrors are also used to reposition the laser beam over the washing machine so to compensate position fluctuations due to the transport system. These behaviors represent a self-adaptation of the measurement system to the varying conditions of the production line and of the models being produced. The actual improvement in measurement uncertainty observed during the tests is remarkable.
Once a machine enters the vibration test station a test sequence starts: the machine accelerates up to its centrifuge speed, then remains in steady state for a few seconds and then stops. Vibrations are measured on the tub, both during transient and steady state. Vibration signal analysis is performed in time domain, in spectral domain and in time-frequency domain. Depending on the production history of the single machine being tested, which is made available to the QCA by MAS, the level of detail in the signal processing is adapted. The concept is that if the machine had no problem during assembly and its quality indicators show that highly probably it is compliant to specifications, then the signal analysis can be simplified, while for machines which may have problems, signal analysis (which requires more time, averaging, etc.) is done more in depth.
This behavior realizes indeed an adaptation which takes into account information at global level.
Similarly to the other quality control stations, the vibration control station is implemented as a LabVIEW application and interacts by exchanging XML files using socket communication and provide quality indices to the MPFQ model.

D) Functional test
The functional test station comprises 6 boxes to perform a set of functional tests to all product instances produced in the production line. The classical test lasts 6 minutes and comprises a fixed plan of tests. During GRACE project this test has been redesigned and made adaptive, i.e. the test sequence can be reprogrammed depending on the production scenario and on the quality indices of the item being tested. Therefore each machine will undergo a specific test sequence, possibly different from others.
The idea is to use the data collected from the individual RA agents, related to the operations executed along the production line for a specific product appliance, to customize the functional test plan accordingly. Particularly, the output of this self-adaptation mechanism embedded in the PA agents will influence the functional test station by customizing the default test plan, namely:
• Changing the sequence of the tests for a specific washing machine.
• Customizing the messages provided to the operator, e.g. highlighting particular points for a more detailed and effective test.
Along the production line, the PA agent, associated to the appliance being produced in the production line, is collecting the data measured during the execution of processing and inspection operations. In particular, the PA receives from each RA agent (i.e. MA or QCA representing respectively processing or quality control stations) a performance index, which is then used to customize the sequence of the tests to be performed by the functional test station. For reason of confidentiality, no detail is given in this publishable report about what is actually implemented; a full description is available in the deliverables of the project.

As many European research projects manage a pilot implementation, but at the same time seem to fail real industrial applicability, focus on the engineering methodology was defined to evaluate methods to improve industrial application and provide a guideline for decentralized manufacturing system engineering.
The GRACE engineering methodology is built up from an engineering process reference model and general engineering activities. The basic dependencies between plant engineering, manufacturing processes and product quality have been investigated. To setup an engineering methodology applicable not only for systems based on the GRACE MAS architecture the idea of functional modularization and mechatronics was used. Mechatronic thinking and multi-agent approaches might be synergized to build functional oriented distributed manufacturing systems. Based on this functional approach, engineering activities identified to be crucial in general and especially for quality oriented and MAS oriented engineering and the engineering process reference model, engineering workflows have been defined. These workflows present a kind of step by step approach describing how to engineer distributed manufacturing systems.
Besides these guidelines a new methodology for consideration of product quality within manufacturing and manufacturing systems engineering has been developed - the so called MPFQ-model. The MPFQ-model is so called based on its four main elements: Material (M), Production Processes (P); Product Functions / Features (F); Product Quality (Q). This model is then used for integrating all information at global level with single product quality, through the action of IMA.

To enable agent communication not only among each other but also with the production database, the agent ontology is including all elements of the MPFQ-model. Thus the concepts of materials, processes, functions and quality, as well as their correlations are built into the ontology.

The MPFQ model as presented in D4.2 – Appendix A allowed the definition of interactions and correlations between components, processes, and functions on the one hand and a list of specific appliance quality characteristics on the other hand.

The MPFQ model was intended to define a metric that would allow highlighting the contribution of any individual assembly processes and quality for the overall appliance quality produced. In particular, eight quality attributes have been selected: Noise (Q1), Energy Saving (Q2), Component Conformity (Q3), Assembly Conformity (Q4), Off-Leakage (Q5), Washing Performance (Q6), Safety (Q7), and Green Footprint (Q8).
The metric has been designed using a correlation table, analyzing the influence of every single Material/Process/Function vs. the Quality Attribute List above mentioned. This way, MPFQ model stores knowledge about the product, its design parameters, as well as about production processes and product-production interrelation at a central point. The correlation is composed of several hundred of information: every single information or weight has been elaborated based on the experience coming from different points of view. This may be is the first time that this approach has been applied, in particular the generation of an overall figure representative for production quality.
As a first step for the quality assurance and improvement, quality has to be known. Therefore, QCAs and Machine Agents (MA) are collecting data from the running production line. MAs are collecting data right from production processes. They can measure e.g. the insertion depth and insertion force during a bearing insertion process. On the other hand QCA are collecting data directly from the product by the quality control stations they control. This measurement raw data is then stored into a production database.
After the data acquisition the MAs and QCAs also elaborate the data to generate quality numbers describing the quality level on which the process has produced. The elaborated quality values can be assumed as normally distributed and are also stored into the production database.
This knowledge is beyond pure technical experience, it includes a very well understanding of the physical processes carried out during a production process. For the bearing insertion it has to be known which characteristics of the bearing and the housing are influencing the insertion process. Additional the physical interaction between the insertion tool and the bearing has to be known, as well as the physical parameters of the insertion process. Such in depth knowledge has been created throughout the GRACE project within work package 2, where different production processes like bearing insertion, screwing and on board-customization have been investigated.
The same has been done within work package 3 with a focus on new quality measures. Within the GRACE projects multiple points within production have been identified, where information about the current product quality is needed. In order to gather these information new concepts for quality measurements have been designed in order to realize visual and vibration measures.

Both work packages provided algorithms for the elaboration of product quality produced or measured. During the manufacturing process it can be assumed that the results from these measurements are normally distributed and that the optimal quality is also the peak of the Gaussian bell curve. By doing the data elaboration, the position of the current proceeded process/measurement can be evaluated. This quality value is also stored into the production database.
Based on the Performance Indicators measured for the single processes, functions and materials measured, the product agent can evaluate the quality of "his" product. Therefore he is using a quality correlation table, which can be calculated from the MPFQ-model and is stored within a quasi-static part of the database.
It can be seen that the agents play a crucial role within this whole process as they are controlling the production and quality control stations, gathering data, elaborating the information and drawing conclusions from the results. In order to integrate these agents smoothly with the production line as well as with the new quality control approach from the MPFQ-model, there has been a strong interaction between IPB, Whirlpool and Siemens (WP1 and WP4) in order to define a MAS not only capable of controlling Whirlpool production line but also capable of realizing the new quality control approach based on the MPFQ model and the knowledge about production and quality control processes from WP2 and WP3.
The quality correlation table is a matrix representation of MPFQ-model. Here, every influence of a process, function or material on a quality feature is documented. Thus it can be seen which processes, materials and functions have an impact on product quality. Correlation tables once defined may be improved by an IMA during production. To create this table multiple ways are possible. The MPFQ-model can be transformed into a matrix easily as every connection between a quality feature and another model element result in a correlation. The question here is how strong this correlation really is. If it is known it can be written into the correlation table e.g. by experts or it may be already included in the model by weighted edges. Thus, the reasonability of the correlation table is highly depending on expert knowhow. This is one reason why a future update of this table might still be needed. During the expert workshops among all partners and especially between Whirlpool and Siemens special attention was paid to this fact. Nevertheless, the stability of the correlation table can only be proven during extensive life tests. The table will then need to be tuned based on statistical correlation methods. These can be done by an IMA during production to improve table reliability.
Additionally IMA may also run statistical analysis during production to detect trends in raw data, quality numbers or quality of products produced and provide the results e.g. for global adaptation to other agents. In this way additional warnings (e.g. for exchanging tools to maintain high product quality) are enabled.

The most important result of the GRACE project has been the successful demonstration on a real production line of the MAS operating during normal production. This result was achieved through an extraordinary effort of all consortium partners, at the factory of Whirlpool in Naples, Italy, which is producing washing machines.

In the installation of the GRACE MAS system in the Naples factory, the agents were distributed on 7 computers. Namely, they were distributed in the following way:
• 1 PC (Intel(R) Xeon(R) CPU W3565 processor @ 3.20 GHz, 8GB RAM, OS: Windows 7, 64-bits) containing: the Jade platform, 1 IMA agent, several PTAs agents (a different agent is launched for each washing machine model, e.g. 9 PTAs corresponds to 9 different washing machine models) and PAs agents (variable number but in a stable production flow more than 400 agents are simultaneously running).
• 1 PC (Siemens SIMATIC Box PC 627B, Core 2 Duo T7400 2.16 GHz, 3GB RAM, OS: Windows XP Professional) located in the production line containing the LabView application for the Gap Control station and the associated QCA agent.
• 1 PC (Siemens SIMATIC Box PC, Core 2 Duo T5500 1.66 GHz, 1GB RAM, OS: Windows XP Professional) located in the production line containing the LabView application for the Washing Unit Inspection station and the associated QCA agent.
• 1 PC (Siemens SIMATIC Box PC, Core 2 Duo T5500 1.66 GHz, 1GB RAM, OS: Windows XP Professional) located in the production line containing the LabView application for the Assembly Vision Check station and the associated QCA agent.
• 1 PC (Siemens SIMATIC Box PC, Core 2 Duo T5500 1.66 GHz, 1GB RAM, OS: Windows XP Professional) located in the production line containing the LabView application for the Vibration station and the associated QCA agent.
• 1 PC (Desktop PC, CPU AMD Athlon 64 X2 3800+ (Dual Core, 2.0GHz) 2GB RAM; OS Windows 7 Professional) located in the production line containing the LabView application for the Final Assembly Visual Check station and the associated QCA agent.
• 1 PC (Intel(R) Xeon(R) CPU W3565 processor @ 3.20 GHz, 8GB RAM, OS: Windows 7, 64-bits) containing the other 12 RA agents for the A-bearing, B-Bearing, Pulley Screwing, Screwing Upper Counterweight, Screwing Front Counterweight, Functional Tests (6 boxes) and On-board Controller stations.

The agents running on the several PCs are inter-connected by TCP/IP over an Ethernet network.
Note that the processing stations located in the production line (namely the bearing and screwing stations) are not directly connected to the agents due to technical restrictions of the factory, which requires an indirect access by the GraDaCo (i.e. the database that is collecting the production data specifically developed for GRACE). Since GraDaCo is being refreshed each 5 minutes, the GRACE system only detects the execution of such operations with a delay of more than 5 minutes. For this purpose, the clock of the PC running the RAs associated to the processing stations is delayed by 10 minutes (to ensure that all operations are found by the agents). For the quality control stations, there is a direct connection between the LabView applications and the agents, and consequently the operations performed by these machines are detected in real-time.
For the validation of the on-board controller customization, experiments were conducted in the factory by utilizing the factory system facilities GraDaCo/MAS and the reliability lab. The practical implementation of the customization scheme was conducted according to the following steps:
• Selection of WM for calibration: at the FT area the data from upstream processes are analyzed in the MAS and suitable WMs are tagged for calibration. At the end of the production line the selected WMs are picked up for calibration in the reliability lab.
• WM Calibration: at the reliability lab an operator calibrates the WM according to information provided by the integration with the MAS, and the MPFQ model is used in order to update the quality features related to the specific WM.
All the quality control stations developed in GRACE have been physically implemented on the line and fully integrated with it; therefore they operate in real time, on the normal production, testing 100% of it, fully integrated with the production system. They are:(gap control station, 3 vision inspection stations, vibration control station and functional testing. For all the quality control stations, there is a direct connection between the LabView applications and the agents, and consequently the operations performed by these machines are detected in real-time and directly available on the MAS.

Validation took place in several steps. A rather large person-month effort was necessary, because all the implementation and validation activities had to be done by partners at Whirlpool premises. In particular 7 working sessions at the Whirlpool factory in Naples with all partners attending, lasting from a minimum of 3 working days, up to a full week, were necessary:
Working Session 1: May 2012
Working Session 2: June 2012
Working Session 3: August 2012
Working Session 4: October 2012
Working Session 5: March 2013
Working Session 6: April 2013
Working Session 7: May 2013.
This implied also an intense and frequent travelling and the transport of a lot of equipment to the factory and an attitude of the partners to be involved into technical implementations at the factory floor.

The GRACE MAS prototype was intensively tested under off-line and on-line mode using the real data of the production line placed in the Naples factory. The resulting MAS platform has proven able to guarantee the achievement of the following innovative quality properties, showed in the experimental tests:
• Modularity and flexibility: the use of the MAS principles simplifies the development of complex computational software applications by dividing the complex problem by simple ones. This allows achieving modularity since the whole specifications are built upon several modules (i.e. the agents). In the prototype implementation, only 4 different types of agents were developed, i.e. PTA, PA, RA and IMA, each one exhibiting a proper behaviour /specified in WP1.2). For the installation in the Naples factory, it was necessary to create several instances for each type of agent, as described in the previous section. Each instance uses the same agent codification, inheriting its behaviour but customized for its particularities according to a XML file.
• Run-time and on the fly reconfiguration, i.e. adding, removing or changing components without the need to stop, re-program and reinitialize the other components: in the prototype, agents can be removed, others can be added or even some modifications can be performed in the behaviour of the agent without changing the system (i.e. the system can continues running without any perturbation). In the demonstration, this is illustrated by shutting down the RA associated to the Screwing Front Counterweight station, changing the algorithm to calculate the Pi index embedded in this agent and start again the RA agent. This is successful tested validating this functionality (note that in a centralized implementation this feature is not possible).
• Distribution: due to the distributed architecture provided by the MAS system, the distribution in large scale systems is easy as agents might be distributed to hardware computational resources with low processing usage (according to the application needs, e.g. geographical dependency or processing capabilities).
• Robustness: the inexistence of a central control node improves the system robustness, since if an individual node breakdowns, the system continues running and does not also breakdowns (in opposite to the traditional centralized structures). In the demonstration, this is illustrated by shutting down some RAs, for example those related to the functional tests stations, or IMA, without affecting the global behaviour of the system.
• Responsiveness: the use of distributed control structures allows the run-time adaptation, i.e. applying local self-adaptive concepts to adapt the system behaviour according to the unplanned changes i.e. a better response to changes/failures. In the demonstration this was illustrated by the adaptation of the parameters to customize the sequence plan of the functional test and to customize the flow rate parameter of the on-board controller of each washing machine.
• Scalability: the prototype experimentation showed that the GRACE MAS system is scalable. The main drawback is usually associated to the agent middleware (in this case JADE) and it is related to possible delays or congestion in the communication infra-structure due to the growth of the exchanged messages. Note that the increase of the number of agents implies an increase of exchanged messages, not in a linear way, but in a more exponential way. From the achieved experimental results, the increase from several agents to approximately 600 agents (some of them exhibiting weighted GUIs) didn’t provoke visible degradation at this level. As an example, during the on-line demonstration the average of CPU consumption was 40% and memory usage of 30%. The same values were achieved for other distributed off-line demos (in some situations running more than 1.500 agents).
• Smooth migration: the use of the MAS technology allows the smooth migration from old technologies/system to new ones. This is illustrated in the demonstration with the consideration of only 12 stations to be controlled by the GRACE MAS; a slowly and smoothly integration of the remaining stations along the line can be performed gradually in the time.
• At the end, and probably one of the major contributions of the project, is the full assessment of the multi-agent systems benefits in real industrial environments, by demonstrating the effective applicability of multi-agent systems in a real world case, considering the case of a factory operating with a production on-demand.

A computational platform exhibiting such properties is suitable to support improvements on the production process (e.g. since on-board controller parameter are customized to each machine) and the product quality (e.g. since most effective quality control procedures are performed). In fact, the use of the GRACE MAS system allowed implementing some trend analysis mechanisms, namely performed by the IMA agent at local and global level. Namely:
• Analysis of the evolution of the Qi indexes for each product along the line, being compared with the average values for the same indexes for its model. This allows to detect if the quality of this particular product is being above or below the average standard values.
• Analysis of the Qi indexes for each product and generation of warnings in any point of the production line (i.e. at any point of the production) in case the desired quality is not possible to be achieved anymore. Note that a yellow warning is generated if the achievement of the desired quality is in risk, and a red warning is generated when the system detects that it is not possible anymore to achieved the desired quality (even if the remaining operations will be performed with Pi equal to 1).
• Analysis of the evolution of the Qi at a specific station, detecting the trend, average value and standard deviation for each model produced.
• Warning the adjustment of the process parameters.

The experience gathered from the installation of the MAS system in a real industrial production line allowed to learn some lessons:
• The integration with legacy systems, e.g. LabView applications and production databases, should be performed very carefully. At the design level, the integration should be tested intensively to detect and correct exceptions.
• The physical security of computers running the MAS system is a critical issue. For example electric breakdowns will shut down the system. Thus avoid the access of unauthorized personnel and the use of UPS is mandatory.
• The current communication infrastructure, where the MAS operates, must be carefully understood. For example, if a computer has two network cards, it is necessary to ensure that the correct one is the one used.
• The offline tests are important to ensure that errors are corrected, but do not replace the use of online tests. There are situations that arise only in the online environment, and particularly in the industrial environment. For this reason, it is important to properly balance the negative impacts of occupying the production line for testing and the benefits of avoiding the occurrence of errors with the system deployed and functioning.
• The correct feeding of the GraDaCo database is required for the correct operation of the system solution. The use of a single point along the line to carry out data collection emphasizes the importance of constant monitoring that operation. Without data, the MAS system will continue functioning, due to its robustness, but will not provide the desirable outputs (since there is no data available).
• The Swing elements in the agents’ GUIs must consider their capabilities to represent data. For instance, when feeding a pop-up menu it is necessary to consider an upper limit for the number of elements represented. In case of abnormal functioning, GUIs can freeze (although the agent continue running properly), compromising the smooth functioning of the system.

Overall there are five main contributions which have been integrated, as described above, within GRACE and led to a successful realization of the project:
1. A new Multi-Agent System architecture for decentralized production systems. This architecture builds the basis for the technical, especially information technological integration of the project results and provides a basis for execution of “intelligent” production and process control by agents.
2. A quality focused engineering methodology for decentralized production systems. This methodology and especially the MPFQ model developed within provides the basis for the semantically integration of results as the semantic correlations are gathered. Additionally this semantic is also stored within the ontology of the MAS, in order to be available for production control. Besides the gathering of interrelations in product design and production control, the engineering methodology also provides an engineering handbook as a kind of red-line or “how to?” for the engineering of decentralized production systems based on GRACE architecture.
3. An in depth analysis of production processes and process control, as well as process control evaluations. This step is very crucial as it enables the MPFQ-based control strategy by providing the ability to assess production processes. In order to evaluate processes a detailed description of its physical processes is needed. Thus, crucial process parameters could be identified and used to evaluate the result of a production step only from its own feedback data. Hence, no additional quality control is needed and quality estimation at every time in production is enabled.
4. New Quality control and measurement systems have been developed and realized. Although additional information is gathered during the production processes itself, there is still a gap in quality data, as not all processes might feedback all data needed or physical measurement of crucial quality influences is not easily possible. Thus, quality control station are designed to measure these information and close the quality information gap. Analogue to the production processes, a detailed knowledge about the product and the processes are needed.
5. All developments and methodologies defined within GRACE project have been implemented at a real and running industrial production line. The single implementation of the GRACE MAS itself can be found to be maybe the largest implementation in industrial environment. But this was only possible due to the additional yields gained from integration of process and quality control. New quality control stations developed are integrated into the production line and ensure product quality of future washing machines produced. Process control algorithms especially for the on-board controller customization lead to high increase in energy and water efficiency of future washing machines. Finally the engineering methodology has been tested during the engineering of new quality control stations and the MPFQ model servers a new basis for quality based process control and quality assessment of washing machines produced.

The framework developed in GRACE is a new approach for process and production control using MASs. Results obtained throughout the project and demonstrated by real on-line tests in a production line during full operation are very promising as not only product quality can be improved, but also other important effects are seen; such as:
• Monitoring of product quality at any time within production, as quality cannot only be measured at the end of the production line but also after each production step.
• Improved Quality of products due to quality oriented influence on production processes based on knowledge of current state of product quality.
• Improved efficiency of the production line due to ability to elaborate specific problems in product quality and adapting functional test plans accordingly; thus leading to reduced testing times.
• Cost reduction within production as quality can be evaluated at any time. Products may be scrapped when evaluated with bad and not improvable quality results. Improved certainty about product quality, as 100% of the manufactured products can be evaluated on a basic level regarding their quality without any additional test / control stations.
The scientific technical experience gathered from the installation of the GRACE system in a real industrial production line allowed to learn some lessons:
• Calculation of products and processes quality indicators can be used to monitor the production and take corrective actions.
• Improvement of the products performances is possible by analysing the data collected along the production line.
• Quality control stations should be more flexible and should be able to adapt more easily to products and production processes variations. This can be achieved implementing self-X strategies, exploiting past information and using data coming from other stations of the production line.
• Manufacturing process design can be oriented to quality adopting the MPFQ Engineering methodology.
• The integration with legacy systems, e.g. process and quality control machines and production databases, is a very delicate phase. At the design level, the integration should be tested intensively to detect and correct exceptions.
• The offline tests are important to ensure that errors are corrected, but do not replace the use of online tests. There are situations that arise only in the online environment, and particularly in the industrial environment. For this reason, it is important to properly balance the negative impacts of occupying the production line for testing and the benefits of avoiding the occurrence of errors with the system deployed and functioning.
• Multi-Agent technology is mature enough to be profitably adopted also in an industrial environment. It is very scalable (several hundred agents have been deployed) and requires just standard equipment (PCs and network connection).

Potential Impact:
The implementation of the GRACE multi-agent system had an impact at industrial level on the following issues:
• improvement of production efficiency;
• reduction of production-line down-time;
• reduction of non-conformities by improved quality control of products;
• improvement on product energetic efficiency.

A quantitative estimate of the industrial impact of an action on production efficiency has been done with reference to WM manufacturing. Typical figures of production efficiency in a modern production line of WMs are about 90%. The adaptation of the functional inspection test plan could increase the efficiency of the functional testing area of about 8% (if 12% of the production are tested with a shorter functional test based on the high quality indicators values).
Moreover, appliance industry are facing relevant costs for down-time due to frequent change of models and costs of non quality (production inefficiency cost, service cost, etc.). The self-adaptation of the test plan parameters permits to reduce the down-time of the stations when new models are produced. Only for the 20% of new models the human intervention is still required (the expected down-time reduction has been estimated to be about 10%).
On the side of product quality, leading producers of home appliances are developing highly innovative and sophisticated products. In particular, consumer demand in Europe is asking for energy efficient and silent appliances: Europeans are indeed aware that efficient models are less costly to operate in the long term and that using WMs during the night, when energy costs are often lower, is possible only if WMs exhibit good vibro-acoustic performance. Quality management is the only strategy to improve European customer satisfaction and, as a consequence, defend European manufacturing competitiveness and continue production in Europe. Production of high quality products requires improved quality control; GRACE project has an impact on improvement of quality control systems, thus on the reduction of costs of non-quality. The introduction of self-optimizing quality controls reduces non-conformities of about 1,5% .
GRACE has an important and measurable impact on product energetic efficiency. In fact, GRACE multi-agent architecture developed adaptable Machine Agents, which allow to adapt the parameters of each microcontroller on-board of each WM with reference to the control flow valve parameters. This reduced the scatter in performance among WMs compliant to specifications by 50%; at present, scatter in performance determines lower energy efficiency of WMs.
About 15% of electric energy consumption of an average European family is due to WM and Dish Washer. WM energetic efficiency depends on compliance to design specifications, therefore GRACE impacts significantly on the sustainability of domestic appliance use. The impact in terms of energy and water saving is estimated to be 5% for each WM. This will be perceived by customers as improved quality and in general will help reducing energy consumption of families.

In the dissemination activities, three different levels can be defined: awareness, understanding and action.
First, it is important to create awareness on the project, its objectives and its results. Scientific publications and participation to conferences and other events are the first steps moved by the consortium in the first part of the project.
The next step will be to select the target interested in understanding the project, its development and its results, improving the audience and the network of figures interested in it, gaining important feedbacks from the conversations about the project.
Then, it will be possible to reach the third level of dissemination that means action, change of practice. This important step involves, for example, industries that will perceive the importance of the project and its outcomes in order to improve their production processes.
According to the first level of the dissemination activities, from the beginning of the project (July 2010) different documents to allow the initial communication of the project and the tools useful for the activities of the partners (i.e. website reserved area and document templates) have been realized.
From 2012, when the first results achieved with the project will be available, more focused and interesting dissemination activities will begin, according to the 2nd level of dissemination.
Target groups of the project as industry, research centres, academia and decision makers at international level will be precisely defined, in order to address the third level of dissemination.
The dissemination activities have been divided in Internal and External Dissemination.

Internal dissemination has been devoted to the preparation of instruments and activities addressed to share knowledge and information between the consortium partners, such as:
- document templates
- reserved area
- internal technical meetings
- deliverables
External dissemination has been devoted to the development of instruments and activities able to widespread awareness on the project among scientific, industrial and public community, such as:
• website
• brochure & poster & video
• seminars, fairs & events
• workshops
• publications
• scientific journals
• press
• in-house partners’ communication
• student internships and thesis
• link with other EU research projects
• handbook
All these activities have been carried out during the 36 months period of the project.

The main activities performed for the dissemination can be summarized as following:
• Organization and update of the public website in order to disseminate technical results, publications, technical brochures about the results of the project, deliverables and to publicize workshops and other activities;
• Preparation of a video about the main objective of the GRACE project;
• To link the GRACE project with other European research project;
• Presentation of GRACE at international farirs;
• Preparation of publications in scientific, industrial-oriented and International journals and conferences, tradeshows, company organized events and trade magazines;
• Preparation of an Engineering Handbook dedicated to industrial engineers who want to apply the GRACE MAS platform in their manufacturing facilities (it will describe prerequisites, engineering methodology and best practices).

A) Web-site
One of the most efficient ways for dissemination is considered to be the public website. It is the best tool to spread information about the GRACE project activities and it is designed to improve both ease of use and information completeness.
The URL is and it is reported in all dissemination materials produced for GRACE, so that people can be redirected to the website to find further details about the project.
The web-site contains a download area, open to the public, where all public deliverables are available, as well as abstracts of all confidential deliverables and of all publications. About one post per month, throughout the project, has been posted, describing updated information on the project progress.
B) Video
A video of the GRACE project has been realized. The first part shows the project partners and their main competences, and then the concept and achievements of the GRACE project are illustrated.
The video is an excellent dissemination tool, it has been distributed to all partners and it has been shown at many important fairs like Automatica and Industrial Technologies. The GRACE video is also available on the Youtube channel: .
C) Brochures
The main results of the GRACE project have been summarized in ten technical brochures distributed in paper copy during fairs and meetings to people belonging to the public, scientific and industrial communities. These brochures can be downloaded from the GRACE web-site as pdf documents.
D) Fairs and Events
The GRACE project has been presented with a dedicated stand in three international fairs and one conference. The use of stand is fundamental to show the results developed, to attract people and to distribute communication material.
D1) AUTOMATICA, Munich (Germany), 22-25 May 2012
GRACE project attended AUTOMATICA, the International Trade Fair for Automation and Mechatronics, with its own stand: Hall B3, stand 529. In the same hall the major research centres were hosted. In the GRACE stand a demo representing an individual agent in the overall Multi Agent System (MAS) architecture has been presented: the self-adaptive robotized vision system for online quality control. The project and the solution presented raised a lot of interests and, as a matter of fact, we received the visit of two media representatives: the official radio of the fair and the Italian journalist of ‘Automazione Industriale’ (Industrial Automation), to take interviews on the Grace project contents. During the fair, representatives of almost all the project partners have been at the booth, alternating in welcoming visitors and explaining the project contents. Automatica brought a lot of interesting contacts from industry and research.
D2) Industrial Technologies, Aarhus (Denmark), 19-21 June 2012
This exhibition highlights opportunities in the fields of nano, advanced materials and new production technologies and focuses on raw materials, factories of the future, sustainable solutions for energy and resource efficient process industries.
In the 2012 edition GRACE project had a stand in the Technology Area and raised a lot of interest. The participation to the exhibition was fundamental to enrich the project research network: the coordinator of Self Learning EU project and the dissemination responsible of IDEAS EU project were met and common activities were discussed.
D3) AIVELA, Ancona (Italy), 27-29 June 2012
AIVELA is a Conference on Vibration Measurements by Laser and Noncontact Techniques, organized by the Italian Association of Laser Velocimetry and non invasive diagnostics. AIVELA represents a forum for delegates from Industry, Academia and Research Institutes worldwide and for the GRACE project has been the occasion to present some of the results of WP3 concerning the development of self-adapting quality control agents. A stand has been prepared where two demos representing individual agents in the overall Multi Agent System (MAS) architecture has been presented: a self-adaptive robotized vision system for online quality control and a self-adaptive laser vibrometry station for on-line diagnostics. Being the exhibition located at the premises of University Politecnica delle Marche, a large number of students of the faculty of engineering had the occasion to visit the GRACE stand, attracted by the interdisciplinary content of the exhibit at the stand. Also companies and academic institutions visited the stand and expressed their interest in the project developments: Polytec, Bosch, Denso, ST microelectronics and LMS on the industrial side, and English, Polish, Belgian, Italian and Malaysian Universities on the academic side.
D4) SPS IPC Drives Italia, Parma (Italy), 21-23 May 2013
This fair originates from the German homonymous one which is an important event of the industrial automation in Germany and Europe. It brings together suppliers and producers of the automation sector organizing also conferences for researchers and students.
GRACE video and brochures were displayed inside the Loccioni stand and it has been a very good opportunity because most part of visitors came from production industries.

E) Workshops
In the second half of the project 3 workshops have been organized: two sessions during IECON 2012 in Canada and the third one during ISIE 2013 in Taiwan. An important aspect of these workshops is that in two occasions, they have been a moment of close collaboration and knowledge exchange with other two European research funded projects, as IDEAS and SELF LEARNING.
E1) Special session at IECON’12 (Montreal, 2012) - “SS25: Research and Development Projects on Industrial Agents”
• Cooperation between GRACE and IDEAS projects
• 6 papers accepted (3 from GRACE and 3 from IDEAS)
E2) Workshop at IECON’12 (Montreal, 2012) - “Towards Industrial Implementation of Cyber-Physical Systems Merging Service-oriented and Multi-agent System Infra-structures”
• 3 papers accepted
E3) Workshop at ISIE13 (Taiwan, 2013) - “Concept and technologies for Factory of the Future”
• Cooperation between GRACE, SELF LEARNING and IDEAS projects
• 6 papers accepted (2 from GRACE, 1 from IDEAS, 1 from SELF LEARNING and 2 from others)

F) Publications
The scientific production has been rich of results and contents and quantitatively relevant for a project strongly industrially oriented.
At the date of preparation of this report the total count is: 2 journal papers, 30 conference proceedings, 1 book chapter; a number of papers submitted to journals is waiting for review.

G) Students internships and thesis
Several thesis have been completed within GRACE: 7 bachelors, 7 masters and, moreover, 3 PhD thesis have been discussed focusing on different topics and covering most part of the project issues. In particular the 3 PhD Thesis are:
• 2012 | UNIVPM | Lorenzo Stroppa | Self-Adapting Robot Vision System
• 2012 | UNIVPM | Stefano Serafini| Self-Adapting Vibration System
• 2013 | Siemens | Matthias Foehr | Engineering in Factory Automation
Internships are also important instruments to improve both dissemination, research activities and possible future tak-up of technologies; during the project 7 students have been involved in internships.
• Sindre Pedersen, from Norway, moved from NTNU to SINTEF for 3 months in 2010, working on the MAS Architecture
• Samar El-Baharawi, from Egypt, moved from German University at Cairo to UNIVPM for 3 months in 2011, developing her research in VISION SYSTEMS
• Signe Moe, from Norway, moved from NTNU to SINTEF for 3 months in 2011, focusing on the activities about the SCREWING STATION
• Lorenzo Stroppa, from Italy, moved from UNIVPM to IPB for 3 months in 2012 combining the study on MAS with the VISION research
• Alexej Steblau, from Germany, moved from OvGU to Siemens for 5 months in 2012, implementing the FT model
• Franziska Fichtner, from Germany, moved from Siemens to WHI for 5 months in 2012/13, studying the MPFQ mode
• Esteban Arroyo, from Costa Rica, moved from IPB to UNIVPM for 1 MONTH in 2012, working on IMAGE PROCESSING.

H) Press release
Several press release have been done and some more are planned for the next future.
2013 - Automazione
SPS Italia: non c'è due senza tre.Partecipazione del Gruppo Loccioni alla fiera dell'automazione dedicata al food & beverage
2013 – Innovation News
GRACE bringt Produktqualitaet in die Fertigung
2012 - Automazione Oggi
Verso la fabbrica del futuro.Il progetto di ricerca Europeo Grace migliora l'efficienza dei processi produttivi e la qualità dei prodotti
2012 - Ancona
Imprese: Univpm e Loccioni verso la fabbrica del futuroProgetto GRACE: un progetto europeo per migliorare l’efficienza dei processi produttivi e la qualità dei prodotti
2012 - Automazione Industriale
La ricerca ad Automatica con il progetto europeo GraceProgetto GRACE: un progetto europeo per migliorare l’efficienza dei processi produttivi e la qualità dei prodotti
2012 -
Verso la fabbrica del futuro: un progetto europeo per migliorare l'efficienza dei processi produttivi e la qualità dei prodottiProgetto GRACE: progetto di ricerca europeo coordinato da Univpm e Loccioni
2012 -
Towards the Factory of the Future: a European Research Project for the improvement of process and quality control at factory levelGRACE Project: European research project coordinated by Univpm and Loccioni
2012 -
Grace Project at Automatica 2012Grace,a European Project coordinated by Università Politecnica delle Marche in collaboration with Loccioni Group, partecipates to AUTOMATICA
2011 – Tutto Misure
GMMT: Integration of process and quality control using multi-agent technology

Partners are committed to exploitation; detailed plans are not described in this publishable summary, because they are confidential. details are reported in a confidential annex to the deliverable on exploitation.
The main objectives of the Exploitation Task are to identify and coordinate the exploitation of the technologies and of the solutions coming out of the project. The following list has come out after the analysis of all project outputs.
1. Multi-agent architecture for line-production system, integrating process and quality control: the innovation relies on the use of a distributed structure composed of intelligent entities (agents) that are associated to individual physical products, process machines and quality control stations. The communication, the share of knowledge and the cooperation between agents permits to respond quickly to perturbations in the production (planned or unplanned) providing to the system dynamic and fast self-adaptation and self-optimization capabilities. The advantages of this approach are higher production performances and product quality, better re-configurability and flexibility, faster development and maintenance of the system.
2. Engineering process and methodology for implementation of the GRACE system: the exploitation of the engineering process and methodology defined in GRACE project permits to design an effective quality-driven production process and to implement it through a decentralized system.
3. Self-adaptation / optimization mechanisms for the process control of manufacturing / assembly:
3-a) Adaptive screwing profiles: a mathematical model has been created for describing screw assembly operations. The model can be exploited for designing screwing profiles, extracting features, monitoring and controlling the screwing process.
3-b) Bearing insertion: a mathematical model has been created for describing bearing insertion assembly operations. The model can be exploited for extracting features and monitoring the bearing insertion process.
4. Self-adaptation / optimization mechanisms for the parameters of the final product: the values of the parameters contained in the control board are adjusted according to the information collected during the production process about the behaviour of sensors and actuators in the specific washing machine and the measurements taken by the production stations in particular assembly and quality check operations.
5. Algorithms for self-adaptation / optimization at global level of the manufacturing process: mechanisms to analyze and correlate the production data collected by the multi-agent system permit to discover pattern and trends and to elaborate new strategy production policies in order to adapt and optimize the manufacturing process.
6. Algorithms for feature extraction and classification: the interpretation of complex signals (e.g. vibrations and images), in order to classify the product under test, requires the involvement of very skilled persons and is a very time consuming task. The algorithms developed in GRACE project permits to autonomously perform this task resulting in a simplified use of the quality control systems and an increase of the quality of the controlled production.
8. Drum geometry control station: a complete quality control station equipped with a vision system conceived for the measurement of the width of the gap between the tub and the rotating drum of a washing machine has been developed and validated in the production line.
9. Catadioptric vision system based on cone mirror: this vision system permits to perform dimensional measurements all over a cylindrical surface with a single shot and without any scanning devices. This solution reduces significantly the measurement time and improves the system reliability because has no moving parts.
10. Spatially controllable illuminator: the possibility to control independently the spatial distribution of light intensity and color through a single device can be exploited in machine vision applications with difficult light conditions and/or non-optically cooperative surfaces.
11. Algorithms for adaptive robotized vision system: a robotized vision system is more flexible of a fixed constellation of cameras. Variations in the product under test or in its configuration may lead to a manual repositioning of fixed cameras, while in case of a robotized vision station these variations can be managed by modifying (also autonomously) the robot path.
12. Algorithms for self adaptation of vibrometer optical signal: active control of laser beam position and real time data processing of optical signal amplitude lead to an improved signal to noise ratio and consequently to a low measurement uncertainty and a high diagnostic reliability.
13. Functional testing stations with adaptable characteristics: the test plan can be adapted according to the information collected during the production process by the multi-agent system. This permits to achieve an higher efficiency and quality of the production and consequently a reduction of costs.
14. Working prototype system installed in the WM production line: the installation of the prototype permits to fully assess the benefits of the multi-agent technology in real industrial environments. In the Washing Machine production case study the GRACE multi-agent system has led to an increase of energy and washing performances and a reduction of production and warranty costs.
15. MPFQ Model: it a dependency model for extensive description of quality dependencies to product components, product functions and production processes. Strong integration of process and quality control, product design and plant engineering. The model can be also used for production control/quality assurance during the production in order to increase the products quality and to reduce the scraps.
Ownership of background and foreground has been established. No patents have been applied for.
Based on this list for each of the results a table that specifies the possible market, price and possible competitors has been prepared and updated.

List of Websites:
Project web-site:

Contact Persons:
Prof. Nicola Paone -
Dr.Cristina Cristalli -

Project video on youtube:
Dowload video from: