Final Report Summary - FRAME (Fast ramp-up and adaptive manufacturing environment)
The aim of the FRAME project is a paradigm shift from the conventional human-driven ramp-up and system integration process to fully automated, self-learning and self-aware production systems. In particular, the project investigates new methods to represent structure, capability and behaviour not only on machine level, but from a holistic point-of-view. This overall view allows the production system to understand the effects of processes other components, thereby allowing the development of novel methods to automatically detect bottlenecks, errors and potential for system optimisation with regards to the fine-tuning of processes. Furthermore, the project integrates time-compressed simulation sandpits with a self-learning manufacturing environment that allows automatic strategy proposal to fine-tune processes during ramp-up or in response to changes or disruptive events. The outcomes of the project facilitate the development of self-learning production systems that are easy to deploy and can react to fluctuations and disruptive events. This in turn will decrease system ramp-up times and down-times for European industrial sectors, thereby increasing productivity and yield. For system integrators, FRAME lays the foundation for successful system-to-service transformation from static capital-intensive production lines towards using dynamic manufacturing services on demand.
Vision:
FRAME aims to create a new solution for highly adaptive, self-aware assembly systems, which will use automated self-learning, dynamic knowledge sharing, highly integrated sensor networks, and innovative human-machine interaction mechanisms. These next generation assembly systems equipped with FRAME technology will be able to proactively support ramp-up, error-recovery, and operational performance improvement. This will lead to a dramatic cost and time reduction of deploying and maintaining complex assembly systems on demand and improve their effectiveness.
Project context and objectives:
When a manufacturing device is moved, deployed or constructed it goes through a phase known as ramp-up. This is the task of fine-tuning and revising a device from the moment the first item comes off the production line, to the point of maximum output of the machine. This process is traditionally extremely person-centred, that is, a manual operator, or team of operators, using a combination of their own experience and trial and error, to improve the system performance.
FRAME is an European Union (EU) Seventh Framework Programme (FP7) project aimed at introducing automation and software tools to the ramp-up process that will accelerate it and maximise re-use and storage of expert knowledge. Through these actions FRAME aims to reduce the time to market for automation solutions, reduce system integrators' costs for deployment and mitigate the risks associated with the inherently brittle reliance on human expertise. In summary, the project aims to reduce time-to-market and time-to-volume for new and reconfigured systems by more than 30 % through steeper learning curves of self aware and more adaptive machines.
The FRAME consortium brings together ten European partners from five countries:
- the University of Nottingham, United Kingdom (UK), Project coordinator, Higher education institute
- DIAD, Italy, Small and medium-sized enterprise (SME)
- Mikron, Switzerland, Industrial end-user
- IPA - Fraunhofer, Germany, Research institute
- Rolls-Royce, UK, Industrial end-user
- University of Sheffield, UK, Higher education institute
- Bosch, Germany, Industrial end-user
- IDEKO, Spain, Research institute,
- Heitec, Germany, Enterprise,
- ADP, Germany, SME.
This summary explains the aims of the FRAME project in more detail, explores the science and technology both used by and created for, the FRAME project before concluding on the impact of FRAME through exploitable technology and industrial case studies.
Project context:
One of the key production processes in high labour cost areas such as Europe is assembly of final products in sectors such as automotive, aerospace, pharmaceutical and medical industries. They all require systems that can be installed quickly, can achieve high volumes in very short times, can perform with minimum interruptions and can be reconfigured for new products with minimum cost. As a result high value manufacturing is being transformed from using high capital-intensive assembly lines towards using dynamic assembly services on demand. The assembly system to service transformation is dictated by a number of factors:
(1) increased demand for rapid ramp-up and downscale production systems;
(2) increased demand for assembly systems which can react to disruptive events and fluctuations during the production process; and
(3) increased drive towards after sales service grant agreements for maintenance and upgrade of manufacturing systems.
Consequently, there is a need for a radically new approach towards self-learning and self adapting assembly systems utilising accumulated knowledge that would allow extremely short ramp-up times combined with low cost of maintenance, system reconfiguration and capability upgrade.
FRAME: Aims and objectives
As noted in the introduction, the project aims to reduce time-to-market and time-to-volume for new and reconfigured systems by more than 30 % through steeper learning curves of self aware and more adaptive machines. Similarly, time-to-recover from unpredictable events will be halved compared to the current-state-of-the-art using the knowledge a system has accumulated over its operational life. The development of a holistic, continuous learning system operation and optimisation environment is expected to raise the overall achievable performance and availability of assembly systems. Overall, this will result in significantly more efficient manufacturing systems which will gain increased competitive advantage of European in the medium to long term.
Historical data shows that on average 60 % of the total time taken in ramp-up of assembly systems is spent in error identification, location and recovery. The initial integration and rectification actions require only around 40 % of the time. Furthermore, the ramp-up effort contributes to around 65 % of the cost of the system. By introducing systems that can recognise and identify faults and learn the most appropriate solutions to bring a system up to and beyond full volume production, the ramp-up time can be significantly reduced. Indeed, by offering capabilities which reduce fault detection times by 50 % will result in the total ramp-up time being reduced by 30 %.
In summary, initial projections from the industrial partners indicate that FRAME will be able deliver major quantifiable improvements in the medium to long term:
- reduced ramp-up times by 30 %;
- reduction of ramp-up delay factors by 50 %;
- 30 % reduction to downtime based on improved system reaction to disruptive events (e.g. station and sensor failures) through application of self-adaption and self-learning methods;
- system monitoring, optimisation and verification leading to reduction of rework by 30 %;
- reduced time for system reconfiguration by 30 %;
- increased system flexibility leading to increased level of product customisation by 40 %;
- back tracing relevant data of all processes within the system and respective products leading to improved process validation and quality assurance.
In addition to the impact that FRAME intends to deliver to the European manufacturing industry, the project is focused on producing a coherent set of tools, which interact together and provide a similar look and feel to the users. The tools will not only achieve the aims of FRAME, but do so by fitting into the existing production life cycle for production machines, thus minimising the need for process modification on the part of system integrators. The technical aims and objectives of the project are expanded in the section covering the main science and technology (S&T) results.
Project results:
S&T objectives
The aim of the FRAME project is a paradigm shift from a conventional, resource-intensive, and largely human driven ramp-up and system integration process to human-centred automated assembly systems with self-awareness and self-learning capabilities enabled by the development of an integrated methodology and tools. The project aim is supported by the following key S&T objectives:
- developing new methods to achieve self-aware assembly systems;
- developing methods for human-centred automated ramp-up, system optimisation and adaptation;
- developing new ways to re-use ramp-up strategies and experience in a self-learning production environment;
- supporting the step-change from purely human-driven ramp-up and system integration to proactive self-adaptive human-machine environments;
- integrating and enhancing state-of-the-art machines and production systems with sensor capabilities and metrology as a basis for real-time system awareness;
- developing behaviour models at station and system levels enhancing self-awareness;
- supporting the system-to-service transformation in assembly by automating the deployment of manufacturing services into the line.
As has been widely reported, current manufacturing systems rely on intensive human interactions and expert knowledge during the ramp-up process and in the event of fluctuations in the production process. Furthermore, self-learning mechanisms and re-use of experience are not extensively applied to speed-up system integration or for product-process optimisation. FRAME addresses these gaps through the development of a self-adaptive assembly system that supports the shift from manual ramp-up and rigid volume production to automated, self-learning and self-aware production systems that adapt autonomously to disruptions.
FRAME concept
The FRAME concept comprises three cores focused on self-awareness, self-learning and self-adaptation which reside not only in the individual assembly stations but also as part of the whole assembly system. A collaborative environment, supported by the FRAME architecture, is created where each station is aware of its own capabilities, behaviour and objectives and also shares this understanding across the wider assembly system environment, thus enabling self-awareness on both station and system level.
1. The first cornerstone of the FRAME concept is the enhancement of station / system self-awareness based on synergetic interaction and collaboration between human operators and machines. The objective of the FRAME self-awareness core is to capture human actions, machine actions, and sensory responses and correlate them in a common semantic framework.
2. The second cornerstone of the FRAME concept is extending the functionality of assembly stations and systems with the capability to learn from experience. The objective of the FRAME self-learning core is to make the inherent knowledge of the ramp-up, operation, and error detection and recovery process available and accessible for future decision-making.
3. The third cornerstone of the FRAME concept aims to enable assembly stations and systems to proactively participate in the solution finding process for any given problem scenario. The objective of the FRAME self-adaptation core is to utilise the knowledge of its ramp-up operations to find better and more reliable solutions to solve problems more quickly.
Industrial focus
The industrial focus of the project is realised through three demonstrator platforms aimed at strategic EU sectors:
(1) medical devices manufacture - high volume assembly of smart drug delivery devices (Mikron);
(2) automotive: sub-assembly of automotive sensors (Bosch);
(3) aerospace: assembly and drilling (AMRC, University of Sheffield - Rolls Royce in advisory role).
Summary of the results
Following the FRAME concept description above, this subsection provides an overview of the main scientific and technological results, grouped by subsystem.
Self-learning
Self-learning allows a knowledge base of common errors and corrections to be formed automatically. This will greatly shorten ramp up but also go on to reduce down time and the time spent by maintenance engineers in the field.
In particular, self-learning can:
- learn from small training data samples. The knowledge object algorithm provides the best answer that is feasible from the very beginning.
- deal with varying input vector dimensionality. Affinity matching allows similar, but different dimensionalities to be operated upon.
- learn and apply knowledge continuously. As part of the FRAME system, the knowledge base can be continuously used and updated.
Smart tools
Smart tools provide a technology to offer end-users an interface for tracking manual adjustments performed during the ramp-up phase. This interface allows:
- recognition of a selected manual parameter through its tag (barcode, QR, etc.);
- checking for previous values of a particular parameter (store and present historical data);
- interacting with users to support the decision of a fixed value;
- recommending a value for a specific parameter based on previous experience.
In particular:
- smart tools technology is strongly tied to the FRAME architecture, interacting with internal components to empower FRAME's solution of storing changes and recommending values based on experience;
- the value-added product is provided to floor operators in order to support their activity during ramp-up. Its support will continue for end-users providing an interface with FRAME intelligence;
- just by selecting the parameters' tag-id, smart tools allows the setting up of changes, providing a complete overview of the process;
- one of the most important issues for manual adjustment activities is to be ubiquitous: the software component is deployed on a mobile device (PDA), equipped with sensors to recognise different kinds of tag-ids.
Common semantic model
- The semantic models contain a graphical representation of the assembly systems which can be translated into a formal semantic language.
- This allows further utilisation of the model with other gathered data entries (such as operator information, machine-derived data alarms, etc.).
- The modelling of human intrinsic knowledge with regards to system behaviour and interactions between different system components is possible and can be connected to all other information sources.
- The models can be specified at an early phase as information from the actual assembly system specification can be used to create the model.
- This can be completed either manually or semi-automatically by mapping relevant entities from the assembly system specification (parameters, sensors, etc.) to the corresponding CSM concept.
- After initial specification, the model can and should be continuously refined during the ramp-up phase.
- The models can either work as standalone semantic-based databases or can be integrated with other tools / components (like self-learning etc.).
- User group engineers and operators are targeted. The close integration of the CSM and the behavioural model allows deep insights into the assembly system model and the semantically derived system knowledge.
- Pre-defined queries can be executed during the actual ramp-up of the machine. They are very useful for shop floor engineers.
- For further utilisation by shop floor operators customized GUIs for data entry are highly recommended in order to minimise the required effort for the knowledge formulisation.
Time to event transformation (T2ET)
T2ET connects the FRAME System to PLCs. More precisely it translates the time continuous PLC signals of sensors and actors to FRAME events. These events are then utilised by the self-learning mechanisms.
Data explorer
The data explorer is a common visualisation tool suited to visualise historical data from FRAME events. There are several unique features and snap-ons:
- cycle exploration, suited for experts tuning the dynamics of the machine;
- system analysis, performing virtual material tracking (suited for experts who wish to relate data from different stations for root cause analysis).
Behavioural model
The model can be used as a 'sandpit' for the self-adaption core. The operator is able to test scenarios in the model before making real life changes to the process unit. The behavioural model poses the added advantage that:
- the model can perform significantly more evaluations in a time period than an operator with a real machine;
- the model can produce events that can be used as the best initial guess in the self adaptation core whilst the experience base is being populated.
Architecture design
The aim of the FRAME architecture is to define and formalise a common vision of FRAME in terms of architecture requirements and guidelines that act as a common reference model for all the S&T developments. This has been realised in close collaboration with all project partners and with a clear focus on the industrial requirements. The architecture caters for all FRAME use cases, specifies how components interact using the communication infrastructure and is structured according to the distinction between self-learning methods for fast ramp-up at the station and system levels.
Communication is enabled by a common communication language defined in a semantic model which provides the semantic structure of the knowledge captured and utilised in FRAME. The information is passed in the form of 'Events' - broadcast messages that are sent to all components and processed by any component that deems the information relevant to their task, and 'Queries' - targeted messages that request specific information from a specific component. As the name suggests, 'Queries' can be replied to, allowing two way communications between the components.
The system level architecture for FRAME has a series of additional, bespoke components that operate at the system level. Typically, this means that there is an additional experience recognition, self learning and self adaptation component which monitor information from all sources generating FRAME events within the system. This change in scale facilitates FRAME in identifying changes that may have a locally beneficial effect on the station, but actually cause a globally detrimental effect on the production line in general.
The specific station and system components from the FRAME Architecture are described in the subsequent sub-sections.
Time to event transformation
Demonstrator: Bosch, General
This component takes information from Bosch's AT system and runs an analysis on it in order to convert the continuous feed of time-based information from the machine into a series of FRAME events that can be processed by the other components.
Experience recognition (station)
Demonstrator: Bosch, Mikron, General
The station-level experience recognition component monitors adjustment events and generates objects representing experiences. Experiences are considered to be descriptions of changes made to the system, the typical state of the system prior to the change, the typical state of the system after the change and the quality (KPI) of those states. This makes up part of the self adaptation and self learning use case. When an experience is generated, an event is broadcast. The station-level self learning component reacts to this by generating a query that allows the experience to be converted / absorbed into the knowledge base.
Experience recognition (system)
Demonstrator: Sheffield, Mikron, General
At the system level the experience recognition component has no concept of time, as all information arrives in terms of system state and product. As the station level experience recognition does temporal matching, the same technique cannot be used at the system level. Instead, the experience recognition explores the system state when each product was made and correlates that with the quality of the product as reported by the system level KPI.
Common semantic model (CSM)
Demonstrator: All and general
The CSM component uses the FRAME query interface to allow components to collect the data gathered by the CSM. This pertains not only to the events recorded by the CSM, but also to the structure of the device and the components that make up the item being assembled.
Smart tools
Demonstrator: Sheffield, General
The smart tools component allows the user to scan physical markers on the machine and use them to interact with the FRAME system. Smart tools can operate at the system and station level simultaneously as its primary function is to generate associated events when presented with a physical tag. How those events are then interpreted by the rest of the system is dependent on the user requirements. For the Sheffield demonstrator, those events logged the plate being examined for either drilling or measuring, allowing the user to easily connect the data associated with the drilling process or the measuring process to the data produced.
Self-learning (station)
Demonstrator: Mikron, Bosch, General
The self-learning component (station) has two processes, learning and applying knowledge. Learning occurs when the station-level experience recognition system triggers a 'experience generated' event. This causes the self learning system to query experience recognition for details about the experience and incorporate that experience into the knowledge base.
Self-learning (system)
Demonstrators: Mikron, Sheffield, General
At the system level the self learning process is slightly simpler. As the system level has no concept of time, it is not possible to rely on the user being able to supply a measure of the 'current' system state. Instead the self learning component simply estimates the expected rewards of each adjustment and returns that information, regardless of context.
Self-adaptation
Demonstrators: All and general
The self-adaptation component spans the station and system levels. It receives requests from the self learning HMI, though these requests can be from any FRAME-enabled component. These requests include a flag which specifies whether or not the adaptations are to be considered at the system level or the station level. At the station level, the component interrogates the CSM for the relevant station and includes this information when interrogating the station level self learning component. At the system level this information is not required. The adaptations returned are sent to the sender of the query for presentation to the user.
KPI (station)
Demonstrators: Mikron, Bosch, General - requires demonstrator specific configuration
The KPI component constantly calculates an index of the current performance of the station based on events about product quality or cycle times. This is broadcast as an event.
KPI (system)
Demonstrators: Mikron, Sheffield, General - requires demonstrator specific configuration
The KPI component constantly calculates an index of the current performance of the system based on events about product quality. This is broadcast as an event. For Mikron, the checking station specifies the quality of each item on the pallet. For Sheffield, events are generated by the drill interface which contain the relevant information.
Behavioural model
Demonstrators: Mikron, Bosch, General
The behavioural model attempts to predict the effect of specific, parametric adjustments. These adjustments are entered into the self learning HMI either manually or via a request to the self adaptation system. The system state to be simulated is combined with the system configuration stored within the CSM and a model is built within the Witness software system. The KPI is calculated and returned.
Self-learning HMI
Demonstrators: Mikron, Bosch, General
The self-learning HMI provides interfaces into the FRAME system for the user.
Faro gauge importer
Demonstrators: Sheffield
This component provides an interface for information from the Faro gauge into the system level FRAME components. It broadcasts the plate being measured and the data produced by the Faro gauge measurement device.
Drill file interface
Demonstrators: Sheffield
The drill file interface component produces FRAME events about the drilling process and the drilling recipe to the system level experience recognition system. This information is correlated, using the plate ID, to the quality information from the faro gauge component and used to create a picture of how the recipe affects the quality of the holes drilled.
Mikron loading station
Demonstrators: Mikron
The loading station component provides an interface for the loading user to specify how the pallet is loaded (which of the three pallet locations are occupied), the loading user (so that how the user effects quality can be tracked) and the pallet number (to allow correlation later by the experience recognition system). This information is sent to the checking station, so the pallet can be tracked across the system.
Mikron checking station
Demonstrators: Mikron
The checking station allows a human to perform a final quality check of the items from the Mikron demonstrator. This final check provides the KPI component with the information necessary to assess how adaptations affect the system on a global level, rather than looking at individual station outputs.
MOOS interface
Demonstrators: Mikron
The MOOS interface receives events native to MOOS system and relates them to FRAME events using a relationship table specified in the input file. This allows MOOS events to be externalised to FRAME.
The system and station paradigm
Part of the vision of FRAME is to realise learning and adaptation as a multi-level process. Consider the example of a production device which locally is achieving high-throughput, but is doing so by cutting corners which cause failure in another part of the process. In this situation the station-level KPIs would look extremely promising, but the system-level KPIs would suffer. Being able to link changes at both the system and station level enables FRAME to make sensible decisions that have globally effective results.
The definition of a system
The definition of a system within FRAME is 'A collection of two or more processing units that only share data that relates to the product'. Thus, if a station does not need any information about the parameters of the preceding processing unit, then these two stations will be classed as part of a system
Example
When the high pressure rotor in a gas turbine has been assembled the following stage is only passed the details of the assembly. These details include serial numbers and balance factors. The actual time, temperature and number of attempts are not recorded. The following processing unit is oblivious to the actual method by which the components were assembled and thus they are classed as part of a system.
The University of Sheffield demonstrator has two distinct processing units. Firstly, there is the cell that produces the holes in the test coupons. When all the required holes have been completed the coupon is passed to the second cell where the holes are verified. There is no process data flow between the two cells but the second cell quantifies the performance of the first cell.
The definition of a station
A station for the purpose of the FRAME project is 'A single or collection of processing units whose parameters are linked by a common piece of hardware or software that controls the manufacturing process'. Thus, if any two processing units share the same PLC or drive shaft then they are classed as a station as this PLC or driveshaft is a common point of information or failure.
Example
The Bosch demonstrator has a collection of conveyor belts that link three separate work areas. These areas are gluing, dispensing and vision verification. All three of these processing units do not operate independently as all three cycle times are dependent on the conveyor belt PLC. Thus, for the purpose of the FRAME project this is classed as a station.
The original Mikron station demonstrator has multiple areas within the machine where processing is undertaken. All of these areas are driven from the same motor and drive shaft thus they are all inexorably linked and the compete machine considered a station. The addition of manual workbenches that are physically separate from the machine and operate independently leads to this collection being defined as a system.
Ramifications of system / station definitions for self-learning
The self-learning strategies devised for the station level assume that relating an adjustment to its reward can be performed temporally, i.e. it is assumed that there is a trivial cause and effect relationship between an adjustment being made and its effects being felt by the process.
However, at the system level, naive temporal correlation is not appropriate. Adjustments made at the early stages of an assembly process may have effects that are not detectable until much further along. As a result, learning on a system level requires a different strategy.
A system wide KPI looks at the quality, performance and throughput of the system from the perspective of the pieces coming off the end of the production line. In some cases, the KPI of individual pieces can be trivially separated. For example, in the Sheffield demonstrator, the output is a number of easily index-able holes which can be individually inspected for quality. In this case, it is possible to keep track of the system state during the production of every piece. This provides the opportunity to correlate system-wide quality to station/system level adjustments. For example, the KPI of individual pieces has been measured and stored as a graph. Above it, the system state when each piece was made is also recorded. By merging the information contained in these graphs, it is possible to produce a third graph, relating system state to system KPI.
Finding the relationship between a system state, (or the adjustments made to put the system into that state) and a system-wide KPI closes the feedback loop for the learning algorithms and allows them to make deductions about what adjustments are good for the system and what adjustments are bad. In this case, the adjustments that produced system state 4 are the best, regardless of their effects on a station level.
In systems where it is not possible to recover individual system states and quality measures from pieces, the process becomes more challenging. For example, in the mass production sector, it is not always possible to tell which specific piece was created when. However, it is possible to trace the various system states that contributed to a specific batch. In this case, the relationship between batch quality and the throughput of the system states that contributed to that batch will allow the system KPI to be estimated. The more batches that are created in a given system state, the more accurate that estimate will become.
FRAME as a process
FRAME is not simply a collection of functional components. It is a coherent solution that allows ramp-up problems and disruptions to production to be corrected quickly and efficiently, through integration into the standard production life-cycle for a production device.
At the specification stage, the CSM is able to provide a common way of representing the device to be created, thus formalising the specification process and mitigating the risk of miscommunication of requirements. Similarly the specification of the KPIs allows all parties to clearly define the criteria for success.
At the design and implementation phases, FRAME introduces some minimal overhead, where users must define the interfaces between the machine and the FRAME toolkit and implement them. Communication to the FRAME system is simplified by the chosen component architecture, allowing standard API calls to generate events and queries to and from FRAME.
During ramp-up, FRAME adds significant value; manual adjustments and their effects are recorded, KPIs provide engineering and management users to view the production performance in real-time and previous experience allows the system to recommend improvements based on the current system and station context.
During the support phase, the system continues to gather experience and produce useful feedback about system performance. Otherwise difficult to notice drops in quality or volume can be viewed immediately. Disruptions can be presented to the self adaptation component, which will suggest solutions if similar problems have been encountered in the past.
Proving FRAME works: Industrial case studies
Three industrial case studies were explored throughout the lifetime of the FRAME project:
- an aerospace device, in conjunction with the Advanced Manufacturing Research Centre and the University of Sheffield;
- a medical device, in conjunction with Mikron S.A.; and
- an automotive device, in conjunction with Bosch GmbH.
Automotive demonstrator
The Bosch demonstrator illustrates the station-level operation of FRAME, i.e. FRAME working in real-time on a single production device. After training FRAME on the device, the adjustments recommended by FRAME clearly demonstrated an element of self learning and made positive changes to the functionality of the machine. In addition to the recommendations based on experienced changes, the behavioural model estimated that the time required for station one to complete its task was sufficiently slow as to justify a parallel station, in order to maximise utility of the stations further down the line.
Self adaptation demonstrated an ability to make use of real-world data superior to pure simulation, by disagreeing with the theoretical results from the behavioral model in terms of optimum conveyor speed. This was a result of collisions and stoppages that were caused by conveyor speeds being too fast, that the simulator was unable to predict. This ability to observe the effects of non-trivial interactions and incorporate them into the advice given to the user highlights the immense value of the FRAME project.
Aerospace demonstrator
The Sheffield demonstrator is integrated into FRAME on a purely system level. This means that there is no real-time link to the FRAME components, instead FRAME is used as a post-processing analysis tool. All processing is done in batches, offline. Two files are used to communicate information about the drilling system to FRAME. Firstly, a file containing the drill recipe (control parameters) and drill process information (data recorded during the drilling) is associated, using the smart-tools, to the physical plates being drilled by the machine. Secondly, a file containing the hole parameters measured using the Faro gauge, is associated with the plates (again using smart tools).
This allows the files to be scanned by the experience recognition in such a way to relate drill recipe information with quality information about the time taken to drill the holes and the quality of the produced holes.
The resulting assessment is able to identify which of the presented recipe files is the superior. The superiority of that configuration of the machine is based on the cycle time (i.e. how long it took to drill the holes) and the quality of the result (a hybrid measure based on a high precision measurement of the resulting hole).
The information collated by FRAME is inline with the expert knowledge of the system. The changing of the recipe information is stored as a macroscopic change. For more detailed analysis of the recipe information in the future, it would be necessary to analyse the recipes at a lower level.
Medical demonstrator
The Mikron demonstrator demonstrates FRAME running simultaneously on both the system and station level. The demonstrator comprises two stations that generate FRAME events and a processing station that is a fully integrated FRAME station.
The station level FRAME deployed on the Mikron demonstrator has been demonstrated to provide sensible predictions for all test case scenarios. The system level, (which takes into account data from the integrated station, the operator, the pen layout being used and a manual checking station at the end) has also been demonstrated to be functional. After running the station / system implementation of the FRAME software it was clear that the theoretical definition of the boundary between station and system chosen as the basis for the technical specifications is not perfectly aligned with the Mikron operator's working strategy. Rather than modelling the entire machine as a station, it was concluded that future projects should further break large stations such as that into substations, with the subdivisions being instigated at any point where an automated or manual check is made. This fits with the operator's working strategy of repeatedly trying a single arm over and over again, without using any other part of the machine until they are happy with the results. Such a strategy would not only fit better with the Mikron procedures, but also would yield more portable knowledge between stations with similar mechanical features. It is of note that such a strategy would require significantly more processing power and the CSM would require changes to handle the new partitions. However, this concrete understanding of the system / station boundary is an important result of the FRAME project.
The behavioural model is of limited use to the Mikron demonstrator. Whilst a fully working, coherent model of the Mikron demonstrator has been built successfully, the synchronisation of the Mikron machine via the central shaft, means that the discrete event simulation is unable to recommend advice in terms of cell utilisation. However, despite these challenges, it is possible to run through self learning scenarios and demonstrate a level of learning from the Mikron machine. In addition a new scenario, designed to illustrate system / station interaction has been run where an operator is found to be the root cause of failures elsewhere in the system. During the FRAME deployment to the Mikron demonstrator, it was identified that changes to the FRAME enabled station should be passed up to the system level. These changes have now been incorporated into the FRAME mechanic.
Potential impact:
From a strategic standpoint, the FRAME project contributes to the thematic area of adaptive production systems, in particular to NMP-2008-3.2.2 Self-learning production systems, where the expected impact of self-learning production systems is expressed as: 'Factories with self-learning capability are expected to contribute towards increased competitive advantage of European manufacturers by 10 to 30 % in the medium to long term'. FRAME is an essential step towards achieving this target.
Results achieved and validated exercises performed on the FRAME demonstrators using specific test case scenarios indicate support towards the following strategic impacts:
i. Reduction in development and commissioning time for factory assets:
- Support for rapid adaptation and integration of systems, enabling fast system deployment and thus shortened commissioning times by 30 %.
- Reduction in ramp-up time by up to 50 % through a self-aware system that is able to self-optimise.
ii. Reduction in downtime:
- Production environment that pro-actively predicts potential conflict situations based on the continuous analysis of the real-time behaviour at station and system level. By comparing these production scenarios with previously accumulated knowledge, mitigation strategies can be applied to avoid the occurrence of conflict situations.
- Reduction in initial ramp-up times during system commissioning and support to reduce downtime during reconfiguration.
iii. Improvement of product quality:
- Real time tracking of both the assembly processes and the product outputs allow the system to adapt to optimise production, thus increasing product quality.
- Storage of specific process sequence information regarding individual products, thus allowing back-tracing of faulty products to a specific process and system learning to prevent it from happening in future products.
iv. Increase in station availability:
- Pro-actively address production scenarios and simulate optimisation strategies that can be applied to avoid system failure, thus increasing availability.
- Use of continuous real time data analysis for predictive maintenance strategies to support the maintenance-on-demand approach.
v. Improved efficiency of complex production systems:
- Self-awareness offering from the station level to system level thus promoting collaboration between all processes at different levels and increasing their efficiency.
- Provide self-learning system which enables faster configuration and reconfiguration and cross-learning between different stations and production lines for a steeper learning curve.
In summary, the following targets are set for FRAME enabled devices:
- 30 % reduction in downtime;
- 50 % reduction of ramp-up delay factors;
- 30 % reduction of rework;
- 30 % reduction of time for system reconfiguration;
- 30 % reduction of ramp up times;
- 40 % increased level of production customisation through increased system flexibility.
These targets reflect the philosophy of FRAME to maximise knowledge reuse and accelerate both ramp-up and repair.
The time to event transformation (T2ET) has been identified of particular interest to automation specialists, who estimated it could reduce the cycle time of feeding components by between 5 and 10 %. In addition, the context-aware HMI has been assessed as being capable of reducing recovery time from disruptive events by 10 to 50 % and commissioning time by 5-10 %. This analysis reflects well on the adherence of FRAME enabled devised to the set targets.
During the test case investigations, industrial members of the consortium benefitted directly from their interaction with FRAME. Not only were the intended improvements in ramp-up and downtime reported, but additional improvements, such as product quality, were achieved, through the simple introduction of quality as a KPI instead of purely temporal measures. To this end, the aerospace demonstrator reported the following improvements in quality:
- more than 30 % failure decrease;
- 31 % failure decrease for non compliance;
- 5eduction of 23 % in rejected outputs;
- increase of 64 % in ideal outputs;
- 50 % improvement in quality of acceptable products (according to in-house metrics);
- 12 % relative reduction of cycle time.
Not only have the benefits been seen at the demonstrator level, but technology from the FRAME project has already been implemented by one of the industrial partners into their everyday processes, with an estimated cost saving of EUR 6 million in the first year.
Exploitable results
A number of tools and features encompassed within the FRAME Architecture, have been both developed and investigated during the project. Six clear exploitable results have been identified for further dissemination within the industrial world:
1. multivariate system analysis on PLC cycle, time to event transformation;
2. proactive HMI;
3. self-awareness using CSMs;
4. Self-learning;
5. Self-adaptation;
6. fast ramp-up and adaptive manufacturing environment.
A description of the noted exploitable results is detailed below:
1. Advanced data processing - Interactive high volume data visualisation and multivariate system analysis. Sophisticated mathematical features and contextual selection support. User Plug Ins for PLC based time to event transformation, virtual material tracking, cycle exploration and online reaction.
2. Heitec - Pro-active HMI for capturing context-specific operator actions at the point of occurrence. These includes the HMI Framework, the three-dimensional (3D) machine model, the innovative pro-active intent-activity-system which can be used standalone and as functions customized for FRAME.
3. IPA Fraunhofer - The CSM captures relevant information about a production line, its processes, products, mechanical entities, process and quality parameters. This knowledge is available at the line for the life-time of the line, starting with the initial ramp-up activities. It can be browsed by human engineers and operators and can be used to track all important changes and adjustments.
Additionally, this knowledge is available for software reasoning, as well, and therefore forms a 'self awareness' capability of the production line. The self-awareness aims to provide all important (base) information for further software components, which subsequently perform maintenance, adjustments and associated analysis.
4. University of Nottingham - Self-learning (comprising the self-learning core and the experience recognition system) provides assembly systems with the ability to identify the quality of adjustments made during the ramp up process. This knowledge forms a 'knowledge base' which can be used to inform decision making (see self adaptation).
At the centre of this process is the 'knowledge object algorithm', a novel technique designed to address the unique challenges presented by the ramp up problem. This technique allows the system to learn even when the physical structure of the machine is being changed and allows the seamless integration of inferred and expert knowledge.
5. University of Nottingham - The self adaptation is a mechanism which gathers information from various sources, (CSM, knowledge base) and uses it to recommend adjustments that will improve the quality and efficiency of the device being monitored. Such a system is invaluable for handling knowledge acquired automatically.
6. University of Nottingham / IPA Fraunhofer - Integrated environment to capture and analyse ramp-up experience of complex assembly stations and systems liking data capture with latest machine learning techniques.
Ramp up business case
In the case of assembly systems, different phases of ramp-up can be described when considering the entire production life.
With reference to the phases:
1. A period of ramp-up is conducted at the machine constructor's site, from the commencement of the build phase to the fulfilment of the factory acceptance testing (FAT) by the customer. This ramp-up impacts on the lead time to market of the assembly system. The system is then delivered and installed at the customer site, where a second period of ramp-up is completed for machine set-up at the production site.
2. This phase leads to the site acceptance testing (SAT) and commencement of actual production.
3. Finally, during the production phase, systems may require reconfigurating to fulfil new user requirements or to overcome failures. A new period of ramp-up is therefore required to cover system update or failure recovery.
Assembly system manufacturers, aim to decrease their lead time to market when the system is at the manufacturer's site, and reduce the ramp up phases once the system is at the customer site. For both the Bosch and Mikron business cases, the aim is to decrease time to market and improve production ramp-up at the customer site. At present, FRAME is focussed on improving the ramp-up period at the manufacturer's site. However, there is potential to extend the results to the customer site as well as applying the FRAME system to different production systems.
With reference to the customer site, it is appropriate to consider the application of FRAME for predictive maintenance. This will require close collaboration and exchange of sufficient data between both the manufacturer and customer during the 'functional life' of the production system whilst taking into consideration data ownership. From a commercial perspective, one viable solution would be to not equip production lines with the FRAME systems, but to supply the production capability as a service. Ramp-up and maintenance would therefore be the sole responsibility of the manufacturer. In some sectors this may not be feasible due to the subcontracting of some production to specialist third parties.
Based on the validation results from the deployment of FRAME to the industrial test cases, the end users have outlined that FRAME also supports decision making in the case of reduced data volume capture and is able to represent and filter the data, thus simplifying decisions: 'The FRAME system captures the logic behind decision making and, at a certain point, starts to suggest solutions'. In particular, the self awareness component displays the status of the machines and the evolution of their status in a clear and objective manner.
Given the positive performance of the FRAME system when deployed to the industrial demonstrators four groups of potential customers for the FRAME technology have been identified. The potential customers of the FRAME technology are predicted to have a similar profile to the end- users in the FRAME project. Given the fundamental value and customisable features of FRAME, potential customers are likely to be manufacturing companies dealing with varying phases in the production value chain. A profile of potential customers is detailed below.
- Component manufacturers - Provide components for building systems, automation components producer
- System integrators - Receive systems from component manufactures and assemble, in charge of the ramp-up period, receive requirements from customer(s) and transform the requirements into solutions.
- End-users - System integrator customers - Operate the machines and form the final part of the chain, interested in the learning curve as profit from the final learning phase.
- Consultants - Assist the whole manufacturing chain, starting from the system integrators, apply the tools for developing consultancy business.
Project coordinator: Professor Svetan Ratchev, the University of Nottingham
Address: Head of Manufacturing, Faculty of Engineering, Coates Building, University Park, Nottingham, NG7 2RD, United Kingdom (UK)
Email: Svetan.Ratchev@nottingham.ac.uk
Vision:
FRAME aims to create a new solution for highly adaptive, self-aware assembly systems, which will use automated self-learning, dynamic knowledge sharing, highly integrated sensor networks, and innovative human-machine interaction mechanisms. These next generation assembly systems equipped with FRAME technology will be able to proactively support ramp-up, error-recovery, and operational performance improvement. This will lead to a dramatic cost and time reduction of deploying and maintaining complex assembly systems on demand and improve their effectiveness.
Project context and objectives:
When a manufacturing device is moved, deployed or constructed it goes through a phase known as ramp-up. This is the task of fine-tuning and revising a device from the moment the first item comes off the production line, to the point of maximum output of the machine. This process is traditionally extremely person-centred, that is, a manual operator, or team of operators, using a combination of their own experience and trial and error, to improve the system performance.
FRAME is an European Union (EU) Seventh Framework Programme (FP7) project aimed at introducing automation and software tools to the ramp-up process that will accelerate it and maximise re-use and storage of expert knowledge. Through these actions FRAME aims to reduce the time to market for automation solutions, reduce system integrators' costs for deployment and mitigate the risks associated with the inherently brittle reliance on human expertise. In summary, the project aims to reduce time-to-market and time-to-volume for new and reconfigured systems by more than 30 % through steeper learning curves of self aware and more adaptive machines.
The FRAME consortium brings together ten European partners from five countries:
- the University of Nottingham, United Kingdom (UK), Project coordinator, Higher education institute
- DIAD, Italy, Small and medium-sized enterprise (SME)
- Mikron, Switzerland, Industrial end-user
- IPA - Fraunhofer, Germany, Research institute
- Rolls-Royce, UK, Industrial end-user
- University of Sheffield, UK, Higher education institute
- Bosch, Germany, Industrial end-user
- IDEKO, Spain, Research institute,
- Heitec, Germany, Enterprise,
- ADP, Germany, SME.
This summary explains the aims of the FRAME project in more detail, explores the science and technology both used by and created for, the FRAME project before concluding on the impact of FRAME through exploitable technology and industrial case studies.
Project context:
One of the key production processes in high labour cost areas such as Europe is assembly of final products in sectors such as automotive, aerospace, pharmaceutical and medical industries. They all require systems that can be installed quickly, can achieve high volumes in very short times, can perform with minimum interruptions and can be reconfigured for new products with minimum cost. As a result high value manufacturing is being transformed from using high capital-intensive assembly lines towards using dynamic assembly services on demand. The assembly system to service transformation is dictated by a number of factors:
(1) increased demand for rapid ramp-up and downscale production systems;
(2) increased demand for assembly systems which can react to disruptive events and fluctuations during the production process; and
(3) increased drive towards after sales service grant agreements for maintenance and upgrade of manufacturing systems.
Consequently, there is a need for a radically new approach towards self-learning and self adapting assembly systems utilising accumulated knowledge that would allow extremely short ramp-up times combined with low cost of maintenance, system reconfiguration and capability upgrade.
FRAME: Aims and objectives
As noted in the introduction, the project aims to reduce time-to-market and time-to-volume for new and reconfigured systems by more than 30 % through steeper learning curves of self aware and more adaptive machines. Similarly, time-to-recover from unpredictable events will be halved compared to the current-state-of-the-art using the knowledge a system has accumulated over its operational life. The development of a holistic, continuous learning system operation and optimisation environment is expected to raise the overall achievable performance and availability of assembly systems. Overall, this will result in significantly more efficient manufacturing systems which will gain increased competitive advantage of European in the medium to long term.
Historical data shows that on average 60 % of the total time taken in ramp-up of assembly systems is spent in error identification, location and recovery. The initial integration and rectification actions require only around 40 % of the time. Furthermore, the ramp-up effort contributes to around 65 % of the cost of the system. By introducing systems that can recognise and identify faults and learn the most appropriate solutions to bring a system up to and beyond full volume production, the ramp-up time can be significantly reduced. Indeed, by offering capabilities which reduce fault detection times by 50 % will result in the total ramp-up time being reduced by 30 %.
In summary, initial projections from the industrial partners indicate that FRAME will be able deliver major quantifiable improvements in the medium to long term:
- reduced ramp-up times by 30 %;
- reduction of ramp-up delay factors by 50 %;
- 30 % reduction to downtime based on improved system reaction to disruptive events (e.g. station and sensor failures) through application of self-adaption and self-learning methods;
- system monitoring, optimisation and verification leading to reduction of rework by 30 %;
- reduced time for system reconfiguration by 30 %;
- increased system flexibility leading to increased level of product customisation by 40 %;
- back tracing relevant data of all processes within the system and respective products leading to improved process validation and quality assurance.
In addition to the impact that FRAME intends to deliver to the European manufacturing industry, the project is focused on producing a coherent set of tools, which interact together and provide a similar look and feel to the users. The tools will not only achieve the aims of FRAME, but do so by fitting into the existing production life cycle for production machines, thus minimising the need for process modification on the part of system integrators. The technical aims and objectives of the project are expanded in the section covering the main science and technology (S&T) results.
Project results:
S&T objectives
The aim of the FRAME project is a paradigm shift from a conventional, resource-intensive, and largely human driven ramp-up and system integration process to human-centred automated assembly systems with self-awareness and self-learning capabilities enabled by the development of an integrated methodology and tools. The project aim is supported by the following key S&T objectives:
- developing new methods to achieve self-aware assembly systems;
- developing methods for human-centred automated ramp-up, system optimisation and adaptation;
- developing new ways to re-use ramp-up strategies and experience in a self-learning production environment;
- supporting the step-change from purely human-driven ramp-up and system integration to proactive self-adaptive human-machine environments;
- integrating and enhancing state-of-the-art machines and production systems with sensor capabilities and metrology as a basis for real-time system awareness;
- developing behaviour models at station and system levels enhancing self-awareness;
- supporting the system-to-service transformation in assembly by automating the deployment of manufacturing services into the line.
As has been widely reported, current manufacturing systems rely on intensive human interactions and expert knowledge during the ramp-up process and in the event of fluctuations in the production process. Furthermore, self-learning mechanisms and re-use of experience are not extensively applied to speed-up system integration or for product-process optimisation. FRAME addresses these gaps through the development of a self-adaptive assembly system that supports the shift from manual ramp-up and rigid volume production to automated, self-learning and self-aware production systems that adapt autonomously to disruptions.
FRAME concept
The FRAME concept comprises three cores focused on self-awareness, self-learning and self-adaptation which reside not only in the individual assembly stations but also as part of the whole assembly system. A collaborative environment, supported by the FRAME architecture, is created where each station is aware of its own capabilities, behaviour and objectives and also shares this understanding across the wider assembly system environment, thus enabling self-awareness on both station and system level.
1. The first cornerstone of the FRAME concept is the enhancement of station / system self-awareness based on synergetic interaction and collaboration between human operators and machines. The objective of the FRAME self-awareness core is to capture human actions, machine actions, and sensory responses and correlate them in a common semantic framework.
2. The second cornerstone of the FRAME concept is extending the functionality of assembly stations and systems with the capability to learn from experience. The objective of the FRAME self-learning core is to make the inherent knowledge of the ramp-up, operation, and error detection and recovery process available and accessible for future decision-making.
3. The third cornerstone of the FRAME concept aims to enable assembly stations and systems to proactively participate in the solution finding process for any given problem scenario. The objective of the FRAME self-adaptation core is to utilise the knowledge of its ramp-up operations to find better and more reliable solutions to solve problems more quickly.
Industrial focus
The industrial focus of the project is realised through three demonstrator platforms aimed at strategic EU sectors:
(1) medical devices manufacture - high volume assembly of smart drug delivery devices (Mikron);
(2) automotive: sub-assembly of automotive sensors (Bosch);
(3) aerospace: assembly and drilling (AMRC, University of Sheffield - Rolls Royce in advisory role).
Summary of the results
Following the FRAME concept description above, this subsection provides an overview of the main scientific and technological results, grouped by subsystem.
Self-learning
Self-learning allows a knowledge base of common errors and corrections to be formed automatically. This will greatly shorten ramp up but also go on to reduce down time and the time spent by maintenance engineers in the field.
In particular, self-learning can:
- learn from small training data samples. The knowledge object algorithm provides the best answer that is feasible from the very beginning.
- deal with varying input vector dimensionality. Affinity matching allows similar, but different dimensionalities to be operated upon.
- learn and apply knowledge continuously. As part of the FRAME system, the knowledge base can be continuously used and updated.
Smart tools
Smart tools provide a technology to offer end-users an interface for tracking manual adjustments performed during the ramp-up phase. This interface allows:
- recognition of a selected manual parameter through its tag (barcode, QR, etc.);
- checking for previous values of a particular parameter (store and present historical data);
- interacting with users to support the decision of a fixed value;
- recommending a value for a specific parameter based on previous experience.
In particular:
- smart tools technology is strongly tied to the FRAME architecture, interacting with internal components to empower FRAME's solution of storing changes and recommending values based on experience;
- the value-added product is provided to floor operators in order to support their activity during ramp-up. Its support will continue for end-users providing an interface with FRAME intelligence;
- just by selecting the parameters' tag-id, smart tools allows the setting up of changes, providing a complete overview of the process;
- one of the most important issues for manual adjustment activities is to be ubiquitous: the software component is deployed on a mobile device (PDA), equipped with sensors to recognise different kinds of tag-ids.
Common semantic model
- The semantic models contain a graphical representation of the assembly systems which can be translated into a formal semantic language.
- This allows further utilisation of the model with other gathered data entries (such as operator information, machine-derived data alarms, etc.).
- The modelling of human intrinsic knowledge with regards to system behaviour and interactions between different system components is possible and can be connected to all other information sources.
- The models can be specified at an early phase as information from the actual assembly system specification can be used to create the model.
- This can be completed either manually or semi-automatically by mapping relevant entities from the assembly system specification (parameters, sensors, etc.) to the corresponding CSM concept.
- After initial specification, the model can and should be continuously refined during the ramp-up phase.
- The models can either work as standalone semantic-based databases or can be integrated with other tools / components (like self-learning etc.).
- User group engineers and operators are targeted. The close integration of the CSM and the behavioural model allows deep insights into the assembly system model and the semantically derived system knowledge.
- Pre-defined queries can be executed during the actual ramp-up of the machine. They are very useful for shop floor engineers.
- For further utilisation by shop floor operators customized GUIs for data entry are highly recommended in order to minimise the required effort for the knowledge formulisation.
Time to event transformation (T2ET)
T2ET connects the FRAME System to PLCs. More precisely it translates the time continuous PLC signals of sensors and actors to FRAME events. These events are then utilised by the self-learning mechanisms.
Data explorer
The data explorer is a common visualisation tool suited to visualise historical data from FRAME events. There are several unique features and snap-ons:
- cycle exploration, suited for experts tuning the dynamics of the machine;
- system analysis, performing virtual material tracking (suited for experts who wish to relate data from different stations for root cause analysis).
Behavioural model
The model can be used as a 'sandpit' for the self-adaption core. The operator is able to test scenarios in the model before making real life changes to the process unit. The behavioural model poses the added advantage that:
- the model can perform significantly more evaluations in a time period than an operator with a real machine;
- the model can produce events that can be used as the best initial guess in the self adaptation core whilst the experience base is being populated.
Architecture design
The aim of the FRAME architecture is to define and formalise a common vision of FRAME in terms of architecture requirements and guidelines that act as a common reference model for all the S&T developments. This has been realised in close collaboration with all project partners and with a clear focus on the industrial requirements. The architecture caters for all FRAME use cases, specifies how components interact using the communication infrastructure and is structured according to the distinction between self-learning methods for fast ramp-up at the station and system levels.
Communication is enabled by a common communication language defined in a semantic model which provides the semantic structure of the knowledge captured and utilised in FRAME. The information is passed in the form of 'Events' - broadcast messages that are sent to all components and processed by any component that deems the information relevant to their task, and 'Queries' - targeted messages that request specific information from a specific component. As the name suggests, 'Queries' can be replied to, allowing two way communications between the components.
The system level architecture for FRAME has a series of additional, bespoke components that operate at the system level. Typically, this means that there is an additional experience recognition, self learning and self adaptation component which monitor information from all sources generating FRAME events within the system. This change in scale facilitates FRAME in identifying changes that may have a locally beneficial effect on the station, but actually cause a globally detrimental effect on the production line in general.
The specific station and system components from the FRAME Architecture are described in the subsequent sub-sections.
Time to event transformation
Demonstrator: Bosch, General
This component takes information from Bosch's AT system and runs an analysis on it in order to convert the continuous feed of time-based information from the machine into a series of FRAME events that can be processed by the other components.
Experience recognition (station)
Demonstrator: Bosch, Mikron, General
The station-level experience recognition component monitors adjustment events and generates objects representing experiences. Experiences are considered to be descriptions of changes made to the system, the typical state of the system prior to the change, the typical state of the system after the change and the quality (KPI) of those states. This makes up part of the self adaptation and self learning use case. When an experience is generated, an event is broadcast. The station-level self learning component reacts to this by generating a query that allows the experience to be converted / absorbed into the knowledge base.
Experience recognition (system)
Demonstrator: Sheffield, Mikron, General
At the system level the experience recognition component has no concept of time, as all information arrives in terms of system state and product. As the station level experience recognition does temporal matching, the same technique cannot be used at the system level. Instead, the experience recognition explores the system state when each product was made and correlates that with the quality of the product as reported by the system level KPI.
Common semantic model (CSM)
Demonstrator: All and general
The CSM component uses the FRAME query interface to allow components to collect the data gathered by the CSM. This pertains not only to the events recorded by the CSM, but also to the structure of the device and the components that make up the item being assembled.
Smart tools
Demonstrator: Sheffield, General
The smart tools component allows the user to scan physical markers on the machine and use them to interact with the FRAME system. Smart tools can operate at the system and station level simultaneously as its primary function is to generate associated events when presented with a physical tag. How those events are then interpreted by the rest of the system is dependent on the user requirements. For the Sheffield demonstrator, those events logged the plate being examined for either drilling or measuring, allowing the user to easily connect the data associated with the drilling process or the measuring process to the data produced.
Self-learning (station)
Demonstrator: Mikron, Bosch, General
The self-learning component (station) has two processes, learning and applying knowledge. Learning occurs when the station-level experience recognition system triggers a 'experience generated' event. This causes the self learning system to query experience recognition for details about the experience and incorporate that experience into the knowledge base.
Self-learning (system)
Demonstrators: Mikron, Sheffield, General
At the system level the self learning process is slightly simpler. As the system level has no concept of time, it is not possible to rely on the user being able to supply a measure of the 'current' system state. Instead the self learning component simply estimates the expected rewards of each adjustment and returns that information, regardless of context.
Self-adaptation
Demonstrators: All and general
The self-adaptation component spans the station and system levels. It receives requests from the self learning HMI, though these requests can be from any FRAME-enabled component. These requests include a flag which specifies whether or not the adaptations are to be considered at the system level or the station level. At the station level, the component interrogates the CSM for the relevant station and includes this information when interrogating the station level self learning component. At the system level this information is not required. The adaptations returned are sent to the sender of the query for presentation to the user.
KPI (station)
Demonstrators: Mikron, Bosch, General - requires demonstrator specific configuration
The KPI component constantly calculates an index of the current performance of the station based on events about product quality or cycle times. This is broadcast as an event.
KPI (system)
Demonstrators: Mikron, Sheffield, General - requires demonstrator specific configuration
The KPI component constantly calculates an index of the current performance of the system based on events about product quality. This is broadcast as an event. For Mikron, the checking station specifies the quality of each item on the pallet. For Sheffield, events are generated by the drill interface which contain the relevant information.
Behavioural model
Demonstrators: Mikron, Bosch, General
The behavioural model attempts to predict the effect of specific, parametric adjustments. These adjustments are entered into the self learning HMI either manually or via a request to the self adaptation system. The system state to be simulated is combined with the system configuration stored within the CSM and a model is built within the Witness software system. The KPI is calculated and returned.
Self-learning HMI
Demonstrators: Mikron, Bosch, General
The self-learning HMI provides interfaces into the FRAME system for the user.
Faro gauge importer
Demonstrators: Sheffield
This component provides an interface for information from the Faro gauge into the system level FRAME components. It broadcasts the plate being measured and the data produced by the Faro gauge measurement device.
Drill file interface
Demonstrators: Sheffield
The drill file interface component produces FRAME events about the drilling process and the drilling recipe to the system level experience recognition system. This information is correlated, using the plate ID, to the quality information from the faro gauge component and used to create a picture of how the recipe affects the quality of the holes drilled.
Mikron loading station
Demonstrators: Mikron
The loading station component provides an interface for the loading user to specify how the pallet is loaded (which of the three pallet locations are occupied), the loading user (so that how the user effects quality can be tracked) and the pallet number (to allow correlation later by the experience recognition system). This information is sent to the checking station, so the pallet can be tracked across the system.
Mikron checking station
Demonstrators: Mikron
The checking station allows a human to perform a final quality check of the items from the Mikron demonstrator. This final check provides the KPI component with the information necessary to assess how adaptations affect the system on a global level, rather than looking at individual station outputs.
MOOS interface
Demonstrators: Mikron
The MOOS interface receives events native to MOOS system and relates them to FRAME events using a relationship table specified in the input file. This allows MOOS events to be externalised to FRAME.
The system and station paradigm
Part of the vision of FRAME is to realise learning and adaptation as a multi-level process. Consider the example of a production device which locally is achieving high-throughput, but is doing so by cutting corners which cause failure in another part of the process. In this situation the station-level KPIs would look extremely promising, but the system-level KPIs would suffer. Being able to link changes at both the system and station level enables FRAME to make sensible decisions that have globally effective results.
The definition of a system
The definition of a system within FRAME is 'A collection of two or more processing units that only share data that relates to the product'. Thus, if a station does not need any information about the parameters of the preceding processing unit, then these two stations will be classed as part of a system
Example
When the high pressure rotor in a gas turbine has been assembled the following stage is only passed the details of the assembly. These details include serial numbers and balance factors. The actual time, temperature and number of attempts are not recorded. The following processing unit is oblivious to the actual method by which the components were assembled and thus they are classed as part of a system.
The University of Sheffield demonstrator has two distinct processing units. Firstly, there is the cell that produces the holes in the test coupons. When all the required holes have been completed the coupon is passed to the second cell where the holes are verified. There is no process data flow between the two cells but the second cell quantifies the performance of the first cell.
The definition of a station
A station for the purpose of the FRAME project is 'A single or collection of processing units whose parameters are linked by a common piece of hardware or software that controls the manufacturing process'. Thus, if any two processing units share the same PLC or drive shaft then they are classed as a station as this PLC or driveshaft is a common point of information or failure.
Example
The Bosch demonstrator has a collection of conveyor belts that link three separate work areas. These areas are gluing, dispensing and vision verification. All three of these processing units do not operate independently as all three cycle times are dependent on the conveyor belt PLC. Thus, for the purpose of the FRAME project this is classed as a station.
The original Mikron station demonstrator has multiple areas within the machine where processing is undertaken. All of these areas are driven from the same motor and drive shaft thus they are all inexorably linked and the compete machine considered a station. The addition of manual workbenches that are physically separate from the machine and operate independently leads to this collection being defined as a system.
Ramifications of system / station definitions for self-learning
The self-learning strategies devised for the station level assume that relating an adjustment to its reward can be performed temporally, i.e. it is assumed that there is a trivial cause and effect relationship between an adjustment being made and its effects being felt by the process.
However, at the system level, naive temporal correlation is not appropriate. Adjustments made at the early stages of an assembly process may have effects that are not detectable until much further along. As a result, learning on a system level requires a different strategy.
A system wide KPI looks at the quality, performance and throughput of the system from the perspective of the pieces coming off the end of the production line. In some cases, the KPI of individual pieces can be trivially separated. For example, in the Sheffield demonstrator, the output is a number of easily index-able holes which can be individually inspected for quality. In this case, it is possible to keep track of the system state during the production of every piece. This provides the opportunity to correlate system-wide quality to station/system level adjustments. For example, the KPI of individual pieces has been measured and stored as a graph. Above it, the system state when each piece was made is also recorded. By merging the information contained in these graphs, it is possible to produce a third graph, relating system state to system KPI.
Finding the relationship between a system state, (or the adjustments made to put the system into that state) and a system-wide KPI closes the feedback loop for the learning algorithms and allows them to make deductions about what adjustments are good for the system and what adjustments are bad. In this case, the adjustments that produced system state 4 are the best, regardless of their effects on a station level.
In systems where it is not possible to recover individual system states and quality measures from pieces, the process becomes more challenging. For example, in the mass production sector, it is not always possible to tell which specific piece was created when. However, it is possible to trace the various system states that contributed to a specific batch. In this case, the relationship between batch quality and the throughput of the system states that contributed to that batch will allow the system KPI to be estimated. The more batches that are created in a given system state, the more accurate that estimate will become.
FRAME as a process
FRAME is not simply a collection of functional components. It is a coherent solution that allows ramp-up problems and disruptions to production to be corrected quickly and efficiently, through integration into the standard production life-cycle for a production device.
At the specification stage, the CSM is able to provide a common way of representing the device to be created, thus formalising the specification process and mitigating the risk of miscommunication of requirements. Similarly the specification of the KPIs allows all parties to clearly define the criteria for success.
At the design and implementation phases, FRAME introduces some minimal overhead, where users must define the interfaces between the machine and the FRAME toolkit and implement them. Communication to the FRAME system is simplified by the chosen component architecture, allowing standard API calls to generate events and queries to and from FRAME.
During ramp-up, FRAME adds significant value; manual adjustments and their effects are recorded, KPIs provide engineering and management users to view the production performance in real-time and previous experience allows the system to recommend improvements based on the current system and station context.
During the support phase, the system continues to gather experience and produce useful feedback about system performance. Otherwise difficult to notice drops in quality or volume can be viewed immediately. Disruptions can be presented to the self adaptation component, which will suggest solutions if similar problems have been encountered in the past.
Proving FRAME works: Industrial case studies
Three industrial case studies were explored throughout the lifetime of the FRAME project:
- an aerospace device, in conjunction with the Advanced Manufacturing Research Centre and the University of Sheffield;
- a medical device, in conjunction with Mikron S.A.; and
- an automotive device, in conjunction with Bosch GmbH.
Automotive demonstrator
The Bosch demonstrator illustrates the station-level operation of FRAME, i.e. FRAME working in real-time on a single production device. After training FRAME on the device, the adjustments recommended by FRAME clearly demonstrated an element of self learning and made positive changes to the functionality of the machine. In addition to the recommendations based on experienced changes, the behavioural model estimated that the time required for station one to complete its task was sufficiently slow as to justify a parallel station, in order to maximise utility of the stations further down the line.
Self adaptation demonstrated an ability to make use of real-world data superior to pure simulation, by disagreeing with the theoretical results from the behavioral model in terms of optimum conveyor speed. This was a result of collisions and stoppages that were caused by conveyor speeds being too fast, that the simulator was unable to predict. This ability to observe the effects of non-trivial interactions and incorporate them into the advice given to the user highlights the immense value of the FRAME project.
Aerospace demonstrator
The Sheffield demonstrator is integrated into FRAME on a purely system level. This means that there is no real-time link to the FRAME components, instead FRAME is used as a post-processing analysis tool. All processing is done in batches, offline. Two files are used to communicate information about the drilling system to FRAME. Firstly, a file containing the drill recipe (control parameters) and drill process information (data recorded during the drilling) is associated, using the smart-tools, to the physical plates being drilled by the machine. Secondly, a file containing the hole parameters measured using the Faro gauge, is associated with the plates (again using smart tools).
This allows the files to be scanned by the experience recognition in such a way to relate drill recipe information with quality information about the time taken to drill the holes and the quality of the produced holes.
The resulting assessment is able to identify which of the presented recipe files is the superior. The superiority of that configuration of the machine is based on the cycle time (i.e. how long it took to drill the holes) and the quality of the result (a hybrid measure based on a high precision measurement of the resulting hole).
The information collated by FRAME is inline with the expert knowledge of the system. The changing of the recipe information is stored as a macroscopic change. For more detailed analysis of the recipe information in the future, it would be necessary to analyse the recipes at a lower level.
Medical demonstrator
The Mikron demonstrator demonstrates FRAME running simultaneously on both the system and station level. The demonstrator comprises two stations that generate FRAME events and a processing station that is a fully integrated FRAME station.
The station level FRAME deployed on the Mikron demonstrator has been demonstrated to provide sensible predictions for all test case scenarios. The system level, (which takes into account data from the integrated station, the operator, the pen layout being used and a manual checking station at the end) has also been demonstrated to be functional. After running the station / system implementation of the FRAME software it was clear that the theoretical definition of the boundary between station and system chosen as the basis for the technical specifications is not perfectly aligned with the Mikron operator's working strategy. Rather than modelling the entire machine as a station, it was concluded that future projects should further break large stations such as that into substations, with the subdivisions being instigated at any point where an automated or manual check is made. This fits with the operator's working strategy of repeatedly trying a single arm over and over again, without using any other part of the machine until they are happy with the results. Such a strategy would not only fit better with the Mikron procedures, but also would yield more portable knowledge between stations with similar mechanical features. It is of note that such a strategy would require significantly more processing power and the CSM would require changes to handle the new partitions. However, this concrete understanding of the system / station boundary is an important result of the FRAME project.
The behavioural model is of limited use to the Mikron demonstrator. Whilst a fully working, coherent model of the Mikron demonstrator has been built successfully, the synchronisation of the Mikron machine via the central shaft, means that the discrete event simulation is unable to recommend advice in terms of cell utilisation. However, despite these challenges, it is possible to run through self learning scenarios and demonstrate a level of learning from the Mikron machine. In addition a new scenario, designed to illustrate system / station interaction has been run where an operator is found to be the root cause of failures elsewhere in the system. During the FRAME deployment to the Mikron demonstrator, it was identified that changes to the FRAME enabled station should be passed up to the system level. These changes have now been incorporated into the FRAME mechanic.
Potential impact:
From a strategic standpoint, the FRAME project contributes to the thematic area of adaptive production systems, in particular to NMP-2008-3.2.2 Self-learning production systems, where the expected impact of self-learning production systems is expressed as: 'Factories with self-learning capability are expected to contribute towards increased competitive advantage of European manufacturers by 10 to 30 % in the medium to long term'. FRAME is an essential step towards achieving this target.
Results achieved and validated exercises performed on the FRAME demonstrators using specific test case scenarios indicate support towards the following strategic impacts:
i. Reduction in development and commissioning time for factory assets:
- Support for rapid adaptation and integration of systems, enabling fast system deployment and thus shortened commissioning times by 30 %.
- Reduction in ramp-up time by up to 50 % through a self-aware system that is able to self-optimise.
ii. Reduction in downtime:
- Production environment that pro-actively predicts potential conflict situations based on the continuous analysis of the real-time behaviour at station and system level. By comparing these production scenarios with previously accumulated knowledge, mitigation strategies can be applied to avoid the occurrence of conflict situations.
- Reduction in initial ramp-up times during system commissioning and support to reduce downtime during reconfiguration.
iii. Improvement of product quality:
- Real time tracking of both the assembly processes and the product outputs allow the system to adapt to optimise production, thus increasing product quality.
- Storage of specific process sequence information regarding individual products, thus allowing back-tracing of faulty products to a specific process and system learning to prevent it from happening in future products.
iv. Increase in station availability:
- Pro-actively address production scenarios and simulate optimisation strategies that can be applied to avoid system failure, thus increasing availability.
- Use of continuous real time data analysis for predictive maintenance strategies to support the maintenance-on-demand approach.
v. Improved efficiency of complex production systems:
- Self-awareness offering from the station level to system level thus promoting collaboration between all processes at different levels and increasing their efficiency.
- Provide self-learning system which enables faster configuration and reconfiguration and cross-learning between different stations and production lines for a steeper learning curve.
In summary, the following targets are set for FRAME enabled devices:
- 30 % reduction in downtime;
- 50 % reduction of ramp-up delay factors;
- 30 % reduction of rework;
- 30 % reduction of time for system reconfiguration;
- 30 % reduction of ramp up times;
- 40 % increased level of production customisation through increased system flexibility.
These targets reflect the philosophy of FRAME to maximise knowledge reuse and accelerate both ramp-up and repair.
The time to event transformation (T2ET) has been identified of particular interest to automation specialists, who estimated it could reduce the cycle time of feeding components by between 5 and 10 %. In addition, the context-aware HMI has been assessed as being capable of reducing recovery time from disruptive events by 10 to 50 % and commissioning time by 5-10 %. This analysis reflects well on the adherence of FRAME enabled devised to the set targets.
During the test case investigations, industrial members of the consortium benefitted directly from their interaction with FRAME. Not only were the intended improvements in ramp-up and downtime reported, but additional improvements, such as product quality, were achieved, through the simple introduction of quality as a KPI instead of purely temporal measures. To this end, the aerospace demonstrator reported the following improvements in quality:
- more than 30 % failure decrease;
- 31 % failure decrease for non compliance;
- 5eduction of 23 % in rejected outputs;
- increase of 64 % in ideal outputs;
- 50 % improvement in quality of acceptable products (according to in-house metrics);
- 12 % relative reduction of cycle time.
Not only have the benefits been seen at the demonstrator level, but technology from the FRAME project has already been implemented by one of the industrial partners into their everyday processes, with an estimated cost saving of EUR 6 million in the first year.
Exploitable results
A number of tools and features encompassed within the FRAME Architecture, have been both developed and investigated during the project. Six clear exploitable results have been identified for further dissemination within the industrial world:
1. multivariate system analysis on PLC cycle, time to event transformation;
2. proactive HMI;
3. self-awareness using CSMs;
4. Self-learning;
5. Self-adaptation;
6. fast ramp-up and adaptive manufacturing environment.
A description of the noted exploitable results is detailed below:
1. Advanced data processing - Interactive high volume data visualisation and multivariate system analysis. Sophisticated mathematical features and contextual selection support. User Plug Ins for PLC based time to event transformation, virtual material tracking, cycle exploration and online reaction.
2. Heitec - Pro-active HMI for capturing context-specific operator actions at the point of occurrence. These includes the HMI Framework, the three-dimensional (3D) machine model, the innovative pro-active intent-activity-system which can be used standalone and as functions customized for FRAME.
3. IPA Fraunhofer - The CSM captures relevant information about a production line, its processes, products, mechanical entities, process and quality parameters. This knowledge is available at the line for the life-time of the line, starting with the initial ramp-up activities. It can be browsed by human engineers and operators and can be used to track all important changes and adjustments.
Additionally, this knowledge is available for software reasoning, as well, and therefore forms a 'self awareness' capability of the production line. The self-awareness aims to provide all important (base) information for further software components, which subsequently perform maintenance, adjustments and associated analysis.
4. University of Nottingham - Self-learning (comprising the self-learning core and the experience recognition system) provides assembly systems with the ability to identify the quality of adjustments made during the ramp up process. This knowledge forms a 'knowledge base' which can be used to inform decision making (see self adaptation).
At the centre of this process is the 'knowledge object algorithm', a novel technique designed to address the unique challenges presented by the ramp up problem. This technique allows the system to learn even when the physical structure of the machine is being changed and allows the seamless integration of inferred and expert knowledge.
5. University of Nottingham - The self adaptation is a mechanism which gathers information from various sources, (CSM, knowledge base) and uses it to recommend adjustments that will improve the quality and efficiency of the device being monitored. Such a system is invaluable for handling knowledge acquired automatically.
6. University of Nottingham / IPA Fraunhofer - Integrated environment to capture and analyse ramp-up experience of complex assembly stations and systems liking data capture with latest machine learning techniques.
Ramp up business case
In the case of assembly systems, different phases of ramp-up can be described when considering the entire production life.
With reference to the phases:
1. A period of ramp-up is conducted at the machine constructor's site, from the commencement of the build phase to the fulfilment of the factory acceptance testing (FAT) by the customer. This ramp-up impacts on the lead time to market of the assembly system. The system is then delivered and installed at the customer site, where a second period of ramp-up is completed for machine set-up at the production site.
2. This phase leads to the site acceptance testing (SAT) and commencement of actual production.
3. Finally, during the production phase, systems may require reconfigurating to fulfil new user requirements or to overcome failures. A new period of ramp-up is therefore required to cover system update or failure recovery.
Assembly system manufacturers, aim to decrease their lead time to market when the system is at the manufacturer's site, and reduce the ramp up phases once the system is at the customer site. For both the Bosch and Mikron business cases, the aim is to decrease time to market and improve production ramp-up at the customer site. At present, FRAME is focussed on improving the ramp-up period at the manufacturer's site. However, there is potential to extend the results to the customer site as well as applying the FRAME system to different production systems.
With reference to the customer site, it is appropriate to consider the application of FRAME for predictive maintenance. This will require close collaboration and exchange of sufficient data between both the manufacturer and customer during the 'functional life' of the production system whilst taking into consideration data ownership. From a commercial perspective, one viable solution would be to not equip production lines with the FRAME systems, but to supply the production capability as a service. Ramp-up and maintenance would therefore be the sole responsibility of the manufacturer. In some sectors this may not be feasible due to the subcontracting of some production to specialist third parties.
Based on the validation results from the deployment of FRAME to the industrial test cases, the end users have outlined that FRAME also supports decision making in the case of reduced data volume capture and is able to represent and filter the data, thus simplifying decisions: 'The FRAME system captures the logic behind decision making and, at a certain point, starts to suggest solutions'. In particular, the self awareness component displays the status of the machines and the evolution of their status in a clear and objective manner.
Given the positive performance of the FRAME system when deployed to the industrial demonstrators four groups of potential customers for the FRAME technology have been identified. The potential customers of the FRAME technology are predicted to have a similar profile to the end- users in the FRAME project. Given the fundamental value and customisable features of FRAME, potential customers are likely to be manufacturing companies dealing with varying phases in the production value chain. A profile of potential customers is detailed below.
- Component manufacturers - Provide components for building systems, automation components producer
- System integrators - Receive systems from component manufactures and assemble, in charge of the ramp-up period, receive requirements from customer(s) and transform the requirements into solutions.
- End-users - System integrator customers - Operate the machines and form the final part of the chain, interested in the learning curve as profit from the final learning phase.
- Consultants - Assist the whole manufacturing chain, starting from the system integrators, apply the tools for developing consultancy business.
Project coordinator: Professor Svetan Ratchev, the University of Nottingham
Address: Head of Manufacturing, Faculty of Engineering, Coates Building, University Park, Nottingham, NG7 2RD, United Kingdom (UK)
Email: Svetan.Ratchev@nottingham.ac.uk