Self-Learning Embedded Services for Integration of

Self-Learning Embedded Services for Integration of Complex, Flexible Production Systems Dragan Stokic (1), Ph.D., Sebastian Scholze (1), Jose Barata (2) (1) Institute for Applied System Technology, ATB-Bremen, Germany (2) UNINOVA - Instituto de Desenvolvimento de Novas Tecnologias, Lisbon, Portugal [email protected]

Abstract- A new approach, based on self-learning embedded services, allowing for high adaptation and integration of complex, flexible production systems is presented. The context sensitive solution for adaptation of discrete, flexible assembly/manufacturing lines is proposed. The proposed solution includes context extractor, adapter and self-learning modules allowing for integration of the control and so-called secondary processes depending on the extracted context. Both context extraction and adapter are continuously learning and improving their performance. Service Oriented Architecture principles are used to implement these modules. Generic solution and specific applications in three various manufacturing environments are presented.

I.

INTRODUCTION

Modern highly flexible manufacturing processes require both effective control of production systems and efficient (socalled) secondary processes1 in order to assure their highest availability and efficiency. Currently, secondary processes such as maintenance, energy saving etc. and production control activities are often separated, leading to low efficiency and high costs for both users and manufacturers of production equipment, especially for those acting at the global market. As indicated in [2], merging of e.g. the world of maintenance and the world of control in production systems may lead to enormous benefits regarding efficiency of manufacturing/assembly processes and maintenance activities, as well as regarding flexibility, equipment availability etc. There is a clear need to establish a real-time management of maintenance, energy savings etc. in order to open possibilities for new services and new business models. In order to merge the two worlds in an effective way, the monitoring and control of production system must exhibit high (self-learning) adaptability to fluctuation in a flexible manufacturing process (e.g. product mix), changes of process and equipment parameters. Changes in process and equipment parameters have to be effectively identified and control and secondary processes activities have to be adapted/synchronized accordingly. Such an integration of control & secondary processes is related to considerable organizational changes, specifically within an Extended Enterprise (EE), addressing a needed tight collaboration 1 Under Secondary (support) Processes are understood business processes that produce products or support primary processes that are invisible to the external customer but essential to the effective management of the business [1].

between equipment manufacturers and users. On the other hand, in order to support this integration, implementation of complex models and algorithms for such self-learning production systems require powerful ICT infrastructures. A service oriented approach opens entire new perspectives in the embedded networking for self-learning & self-adapting production systems. It is likely that approaches based on Service Oriented Architecture (SOA) principles, using distributed networked embedded services in device space (sensors, controllers etc.), are the most appropriate for implementation of self-learning production systems in general, and especially those aiming to merge control & secondary processes. The objective of the work presented in this paper is to develop reliable service-based self-learning production systems aiming at integration of control and secondary production processes within complex assembly/manufacturing lines. The key assumption of the work is that context awareness approach will allow for adaptation and integration of control and secondary processes in highly flexible manufacturing systems. The paper explores how context awareness can be used to assure self-adaptation of production systems, with integrated control & secondary processes, applying embedded services, i.e. SOA principles. The paper is organized in the following way. Section II provides a brief overview of the state-of-the art. In Section III the overall concept is elaborated. In Section IV the three applications of the proposed concept are briefly described, while in the conclusion (Section V) the future plans are indicated. II.

SURVEY OF THE STATE-OF-THE ART

The brief overview of the state-of-the art is focused upon several key RTD problems and technologies relevant for the approach proposed. In the area of self-learning production systems, the research has demonstrated that the application of machine learning techniques, dynamic self-adaptation and operator’s feedback in the loop promises to increase the intelligence of the overall system. In this approach, machine learns whenever it changes its structure, program, or data based on inputs or in response to external information in such a manner that its overall performance is expected to improve [3-5]. In production systems in particular, these methods have been proven to be

especially useful for monitoring/diagnosis [6-8] and control [9]. However, the applications of self-adaptation of production systems and learning in industrial practice are still in initial phase. In this paper a novel approach in production systems is presented, which is context aware, adaptive to contextual changes at run time and learns from adaptation and operator’s action. Of special interest for the work presented in this paper is Service Oriented Architecture (SOA) in manufacturing i.e. the relation between self-learning production systems and SOA. SOA gained popularity with the advancement of web service technologies in industrial networks. For example, DPWS (Devices Profile for Web Services) allows devices with embedding computing capabilities, sensors, actuators etc. to interact within the industrial network and enabling SOA integration to higher levels [10]. OPC Unified Architecture is the recent OLE (Object Linking and Embedding) for process control (OPC) specification from the OPC Foundation and differs significantly from its predecessors. The Foundation's goal was to provide a path forward from the original OPC communications model (namely COM/DCOM) to a crossplatform SOA for process control, while enhancing security and providing an information model [11]. Scalable SOA holds promise for seamless integration, interoperation and flexibility in manufacturing environment. But, there is a lack of adoption of overall SOA based self-adaptive production systems in discrete manufacturing environment. In this paper, SOA based self-learning solutions are proposed as add-on features in existing shop floor infrastructure, Context awareness is widely applied in modern ICT solutions. Different approaches to context modeling are developed, where ontology based context modeling is mostly investigated. Based on the formal description of context information, context can be processed with contextual reasoning mechanisms. Since contextual information has some inherent features (it can be considered incomplete, temporal, and interrelated) context reasoning can exploit reasoning mechanisms to deduce high level, inferred context from low-level raw contextual information. Furthermore, contextual reasoning can be used to verify and possibly solve inconsistent context knowledge due to imperfect input. For example, Luther et al. [12] show the needs for ontology support and reasoning in mobile applications; their case study is conducted with the Protégé knowledge workbench [13] for ontology modeling and OWL editing, and the RACER inference engine [14] for proof checking, ontology validation and classification. A more flexible use of ontological reasoning is presented in [15]. Their framework utilizes context-awareness for service classification. Application of context awareness for SOA based selflearning production systems has not yet been sufficiently investigated. The approach proposed in this paper applies the ontology based context modeling, and re-use of experience of other projects [16] for context extraction in highly flexible and dynamic self-learning production systems [17]. In the area of integration of control of production processes and secondary processes (e.g. maintenance) a

number of research projects have been carried out (e.g. EU project: InLife [18]). In spite of that, both technical and organizational aspects for e.g. control & maintenance integration are still subject of intensive investigations. The lean approach may serve as baseline for organizational aspects related to this integration [19]. The solution proposed in this paper addresses adaptation of various process/control parameters to achieve such integration considering both technical and organizational aspects. III.

PROPOSED APPROACH FOR SELF-LEARNING EMBEDDED SERVICES

A. Proposed Concept The classical manufacturing concepts, often assume a ‘division’ of control and secondary process (e.g. maintenance) of production systems (complex assembly/manufacturing lines). The objective of the work presented is to integrate these ‘activities’ by provision of a self-learning adapter which ‘adapts’ and synchronizes both control and secondary processes (e.g. for maintenance - see Fig. 1). The challenge is to define an adapter able to handle wide scope of ‘disturbances’/changes coming from either enterprise level, or process and equipment parameters changes, requiring harmonized adaptations of both control and “secondary” activities. The proposed approach is to (on-line) identify current dynamically changing context in which the production system operates and to ‘use’ this identified context to adapt both control and secondary processes. Therefore, the proposed approach includes a self-learning context extractor (as a generalized ‘observer’ providing current context) and an adapter (as ‘active’ part) – see Fig. 1.

Fig. 1 Approach to integrate control & secondary processes SOA principles are the most appropriate for implementation of such self-learning production systems aiming to merge control & secondary processes. Context awareness, providing information about the process & equipment and the circumstances under which the services operate and allowing them to react accordingly, is a

promising approach to assure, needed dynamic self-learning adaptation to changes in the context, including changes in processes & equipment parameters. The context can be seen as a bonding element between control and secondary processes of production systems and different services/networks. This approach has not been explored up to now. The basic assumption is that holistic, simultaneous and harmonized usage of context awareness, based on (on-line) extraction of a current context as indicated in Fig. 1, for self-adaptation of production systems (integrating control & secondary processes), is the effective way to assure considerable advantages regarding efficiency and availability of production systems, when applying self-learning systems. B. General architecture A generic solution for context based self-adaptation of production systems is proposed which can be applied in various manufacturing systems. The features and functionality for the overall architecture are derived to meet industry specific needs, but allowing the applications of generic results to the wider scope of discrete manufacturing industries. Context awareness allows for dynamic self-adaptations and learning in production systems. The system adapts to run time critical contextual changes and learns from it. Learning is also enhanced by operator’s feedback and experiences. The overall proposed reference architecture is illustrated in Fig. 2. The architecture is designed following SOA principles as an add-on to the standard control2 following the conceptual approach as presented in Fig. 1. The components of the proposed system include [20]: • Context Extractor, Adapter and Self-learning services - see the text to follow • Expert Collaboration Platform (user validation) module: the identified solutions are required to be validated by user, where the user can manually/automatically accepts/rejects any new solution. The user validation UI sends the feedback to the adapter and learning module. • Evaluator: Performance of adaptation and context extraction are measured by the evaluator either manually via operator’s feedback or automatically via mapping against objective functions at run time. Evaluation results are sent to the Learning module. • Data access layer: Generic component, responsible for accessing the device layer (data from shop floor infrastructure). Information from ERP level, devices or plant data servers are brought to data access layer directly or via middleware depending on plant specific equipment and communication protocols. • Data Processing: These services are responsible for the bidirectional processing of information and perform e.g. pre-processing of monitored raw data acquired via the data access layer, before the context is identified. Main 2 Please note that for the sake of simplicity not all elements of the architecture are presented in Fig. 2.

functionality is to transform the raw data in a format which serves as basis for context identification. The Model Repository contains ontology based plant specific models for equipment, production processes and products. The models are shared by different software components at run time. The Context Repository allows update and storage of extracted/processed contextual information for later retrieval. Information flow among the modules is event driven in some cases and time based in other cases. • Service Infrastructure: underpinning framework ensuring information is securely gathered from trusted context data sources and that the control updates are securely communicated to control systems with appropriate levels of authentication. The communication authentication components ensure seamless and secure connectivity with existing manufacturing and information system communication protocols and security mechanisms.

Fig. 2: Proposed reference architecture [20] 1. Context extractor The Context Extractor is the set of embedded services responsible for identifying changes in the context of the environment. The current identified context is used to extract available context knowledge. The results of the Context Extractor are used in the Adapter which is responsible for updating the system behavior. The purpose of the proposed ontology is to define a fundamental data model for context extraction. The ontology is mainly used to model the device context. As shown in figure 3, the Self-Learning ontology will be defined in such a way that it can be expanded: it is a layered ontology, the generic part forms the core ontology and the application specific part forms the domain specific ontology. The basic principles for context modeling were followed: Support description of main context: In practices, all context information cannot be modeled. The context model should consider those most relevant concepts and information needed to meet the requirement of context sensitive adoption. Model the context that is (easy) acquirable: The context factors considered should be identifiable and acquirable,

whether provided through computer automatically, or by user input explicitly.

monitoring

Figure 3: Device Context Model - ontology Trade-off between investment of context modeling/extracting and effects of context sensitive adoption: Intuitively, if we could model as much context factors in as much details, the accuracy of context will be higher. However, this does not come for free: more time and efforts are needed for context modeling and more computing resources are needed to handle the context, which will bring deficiency to the adoption process. The Context Extractor uses all monitored “raw data” provided via the data access layer to derive the machine’s current contextual situation. Using the ontology/context model the monitored data are evaluated and the context extracted. Based on the identified context, situations can be compared to previous ones and stored. A continuous process, coordinating with the monitoring and followed by the adaption process to give current contextual meaning to the provided knowledge is built based on reasoning around the extraction of a contextual situation. The main elements of the context extractor are: • Adapter Interface: Via this interface the Context Extractor and the Adapter exchange data and call specific functionality on both sides. • Context Identification: It is responsible for the identification of the current context, based on monitored raw data, the adopted ontology and historic context information stored in the context repository. • Rule Engine: Responsible for providing appropriate rules for the identification of context – context reasoning. • User Interfaces: User interfaces for maintaining and administering the rules and the context repository. • Application Specific Modules: This set of services includes application specific components and user interfaces. 2. Adapter The Adapter is the set of services responsible for updating system behavior (locally and/or globally) in response to a

change of context in the environment. The adaptation is based upon available context knowledge to identify the best set of rules to employ in each context. The Adapter is informed by the Context Extractor about a variation on the system (change of context) and adapts the system to handle the new situation. The Adapter is guided by a set of rules that describe how the system should behave in each particular context. These rules can be updated through learning based on lifecycle history data, context and user validation. Although triggered by the Context Extractor whenever a change of context is detected, the Adapter is able to periodically enquire the Learning module to check for a new set of rules that promise to improve system performance in accordance with production objectives. This information can be then presented to the user in order to validate or not the new proposal. In the positive scenario, this proposal will be saved and become the new de facto set of rules, which will delineate system behavior for that particular context. Each application case comprises a specific set of rules and output, but the Adapter structure is generic. • Rule Engine: an engine framework that is able to process application specific rules. • Context Action Selector: These services are responsible for the definition of the adaptation depending on a particular context. These services interpret the results obtained by running the rules in the Rule Engine. They may also be constrained by the user input concerning the validation or not of the proposed solution(s). • Context Action Optimizer: Set of services to update of the current set of rules, either due to a Learning module suggestion (validated by the user) or by direct input. • Application Specific Module: This set includes services that need to be developed for each application individually. Application specific rules and information templates are also defined here. 3. Learning services The proposed concept for self-learning represents the capability of a complete system to learn along its life-cycle based on the experience retrieved from past and current relations and share of information between its different elements in a distributed manner. Learning services allow the system to learn relying on Data Mining (DM) and operator’s feedback to update execution of adaptation and context extraction at run time. The results (adaptation rules) are suggested to the Collaboration platform, and must be backed up by the user (either upon request or automatically). Adaptation rules and context identification procedures are updated accordingly. DM has been considered to be appropriate approach for the concerned applications in complex production systems. The approach relies upon the extraction of previously unknown and potentially useful patterns from data sets. DM is often referred in the context of Knowledge Discovery in Databases (KDD), which consists in selecting data samples from a large

database, treating it and analyzing it for pattern extracting. The sampling process is used because, most of the times, the total amount of data is too large to be fully analyzed. There are several types of methods for DM to discover the patterns in the data. The methods applied in this research are: • Classification – learning of a function aiming to map the data inside a set of classes in the best way possible. This function is based in the attributes and quantitative information of the data items to map. The method is implemented through the use of Decision Trees, K-Nearest Neighbor. • Regression - focuses on the relationship between a dependent variable and one or more independent variables. By those means the dependent variable can be studied for changes whenever any of the independent variables vary. Algorithms that are used to calculate regression are adaptive spline and projection pursuit regressions. • Clustering – seeks to find a finite number of categories or clusters to classify the data. The categories can be mutually exclusive and exhaustive or consist of a richer representation, such as hierarchical or overlapping categories. The algorithms that implement this method are DBSCAN, kMeans and Support Vector Clustering. • Association Rule Learning - Discovers relations between variables in databases and describes them as rules. Algorithms for generating rules are the Apriori Algorithm, Eclat Algorithm and One-attribute-rule. IV.

APPLICATIONS

The proposed concept has been developed following SOA principles and applied in three different scenarios. The three application scenarios belong to different industrial sectors (although all three address discrete manufacturing and machine vendors’ views). Their main characteristics are presented in the following table. For example, the first scenario addresses application of the approach to increase energy efficiency of machines [21]. In a common approach to increase the machine utilization degree a timeout is set for different machine subsystems, which can be shut off during idle time. There is no need to know, when an operation will be executed. The subsystems are automatically turned on after the incoming of a new operation. Due to this reactive behavior, the time needed to execute the planned operations is extended according to the wake-up delay of the subsystems. Therefore, the programmed deadline is exceeded. As an alternative to this common approach for reducing energy consumption, the proposed context sensitive self-learning approach is applied. This solution exhibits advantages compared to the timeout procedure. Although production planning & control may not be available at a process control level, the continuous observation of production context makes it possible to predict future machine utilization. In this way, there is no need to set a timeout. Subsystems of a machine can be shut off directly after the advent of an idle time. If the characteristic wake-up

delay is known, it is possible to turn on the subsystems in advance, avoiding the exceeding of the deadline.

Table 1: Application scenarios overview Main Business

Application Scenario

Objectives/Technical issues addressed

control, automation and drive systems

Control systems of machine tools

Improve current service platform by integration of control and secondary processes (e.g., power saving, maintenance), specifically by improved transparency of complex machines, improvement of tools and methods for the analytical optimisation of secondary processes (e.g. maintenance plan / scheduler, NC-program with “Power save commands”).

machines and automation systems for shoe industry

Control systems of machines/autom ation systems under development

Enhance machines with self-learning features by allowing machines to inspect statistically the condition of products and equipment, report and analyse proactively the gathered statistic values, enabling the machines to decide and adapt the parameters and keep them always inside the “optimized” working range. 3 challenging scenarios defined aiming to identify the unique selflearning solution

highly customized FMS systems

FMS experimental cell

Optimisation of usage of FMS by maximum utilization rate of production machines, minimization of the lead time of production orders. Optimisation of reactive scheduling model for selflearning scheduling and dispatching in FMS.

The objective of the second above listed scenario is the adjustment of parameters from different parameter sets in order to integrate control and maintenance. The objective is to achieve high adaptation of the machines to the changing conditions. Therefore, the proposed solution is applied to react to the changing situations/contexts associated with variations in different parameter sets. The parameter variations are in terms of pressure and temperature, speed frequencies of drives and pumps, proper material mix ratio and filling of materials into shoe forms. For example, one of the scenarios addresses synchronicity of the different valve circuits when dosing of several components. As the valve synchronization is designed by a mechanical system, and/or due to valve abrasion, with different valve opening times, caused by e.g. different force requirements or different air supply, it may come to flaws in the product. Currently the synchronization can only be adjusted by a technician during downtime of the machines. By implementation of the proposed self-learning solution an automatic adjustment of the valve switching to different conditions is achieved. The context monitoring serves as a basis for identifying valve adjustment parameters. The advantage of having such an intelligent monitoring for the valve synchronizations is that it provides a basis for preventive maintenance. Device centric infrastructure is based on machines, the corresponding PLC’ and operational PCs. Ethernet, Field bus, CAN Open, Profibus based communication protocols are used to

communicate within the equipment level. Main information sources from device layer are sensors for pressure/temperature, pumps/drives for speed frequencies and controllers for parameter sets. Enterprise level parameters are accessible in plant database which are mostly associated with varied customization of shoe types and sizes related to the different production lot numbers. Operators adjust the boundary values or parameter sets through the operational PC of machines. The context extractor services extract and process relevant parameters from the database to input to the Adapter set of services. The adapted solutions, in terms of adjusted parameters, changes to speed frequencies of pumps/drives, changes in the process cycle time etc., are sent to the Expert collaboration for user’s feedback. Based on adaptation results and operator’s feedback, the learning services learn how changing cycle times and ambient conditions are influencing the production process (e.g. above explained valve synchronization) and also update the rules for context identification and adaptation.

ACKNOWLEDGMENT This work is partly supported by the Self-Learning (Reliable self-learning production system based on context aware services) project of European Union’s 7th Framework Program, under the grant agreement no. NMP-2008-228857. This document does not represent the opinion of the European Community, and the European Community is not responsible for any use that might be made of its content. REFERENCES [1] [2] [3] [4] [5] [6]

V.

CONCLUSIONS

In this paper a novel approach in production systems is presented, which is context aware, adaptive to contextual changes at run time and learn from adaptation and operator’s action. The solution proposed addresses adaptation of various process/control parameters to achieve integration of control and secondary processes. Two aspects will be specifically elaborated in future research. High complexity of data acquisition and real-time data analysis algorithms and tools for self-learning production systems imposes high demands on quality and security of SW services which are implementing such self-learning algorithms. Therefore, the main obstacles to fully utilize the opportunities offered by service-based self-learning systems are problems related to reliability, availability, interoperability and security & trust of diverse SW services, especially real-time services in device space. This will be the subject of further investigations. As indicated above, an effective integration of control and secondary processes requires solving both organizational and technical problems. The tight integration of control and secondary processes may lead to critical organizational problems taking into account current ‘separation’ between these activities by many equipment users. Effective integration requires consistent application of EE principles (e.g. trust between equipment manufacturers and users), since e.g. equipment manufacturers often need to be strongly involved in the maintenance activities which are, according to the proposed approach, tightly integrated with control of production. Lean production principles (such as minimization of ‘waste’ activities, where ‘wastes’ are all activities not bringing direct added value to product) in combination with the technical concept will be used as a leading approach.

[7]

[8]

[9] [10]

[11] [12] [13] [14] [15]

[16] [17] [18] [19]

[20] [21]

A.G. Rummler, and P.A. Brache. Improving Performance: How to manage the white space on the organizational chart. Jossey-Bass, San Francisco.1995. MANUFUTURE Strategic Research Agenda, Report of the High-Level Group, September 2006 E. Alpaydin, Introduction to machine learning. The MIT Press.2004. D. Michie, D. Spiegelhalter, C. Taylor, and J. Campbell, J. Machine learning, neural and statistical classification. Ellis Horwood Series In Artificial Intelligence, page 289.1995. J Carbonell, R. Michalski,, and T. Mitchell, An overview of machine learning. Machine learning: An artificial intelligence approach, 1:3– 23.1983. S. C. Liu and S. Y. Liu, ‘An Efficient Expert System for Machine Fault Diagnosis,’ The International Journal of Advanced Manufacturing Technology, vol. 21, pp. 691-698, 2003. N. Palluat, D. Racoceanu D, and N. Zerhouni ‘A neuro-fuzzy monitoring system application to flexible production systems,’ Comput. Ind., vol. 57, pp. 528-538, 2006. Y. Jianbo, X. Lifeng, and Z, Xiaojun, ‘Intelligent monitoring and diagnosis of manufacturing processes using an integrated approach of KBANN and GA,’ Comput. Ind., vol. 59, pp. 489-501, 2008. L. Monostori, ‘AI and machine learning techniques for managing complexity, changes and uncertainties in manufacturing,’ Engineering applications of artificial intelligence, vol. 16, pp. 277-291, 2003. L. Ribeiro, J. Barata, A.W. Colombo, G. Jammes, ‘A Generic Communication Interface for DPWS-based Web Services,’ in IEEE International Conference on Industrial Informatics INDIN, Daejeon, Korea, IEEE, 2008. W. Mahnke and S-H. Leitner, “OPC Unified Architecture - The future standard for communication and information modeling in automation” 3/2009 ABB Review 3/2009, pp. 56-61 M. Luther, et al. Situational reasoning - a practical OWL use case. Chengdu, Jiuzhaigou, China: IEEE, 2005. F. N. Noy et al., Creating Semantic Web contents with Protege-2000. IEEE Intelligent Systems, 2001. 16(2): p. 60, 2000. Haarslev, V. and R. Moller. RACER system description, Siena, Italy: Springer-Verlag, 2001. F. Jari, L. Ora, and S. Tapio, RDF-Based Model for Context-Aware Reasoning in Rich Service Environment, in Proceedings of the Third IEEE International Conference on Pervasive Computing and Communications Workshops. IEEE Computer Society, 2005. S. Ziplies, S. Scholze, D. Stokic and K. Krone, Service-based Knowledge Monitoring of Collaborative Environments for Usercontext Sensitive Enhancement, ICE-2009, 2009. M. Sorli, and D. Stokic, Innovating in Product/Process Development, Springer-Verlag, Heidelberg, London, New York, 2009. InLife, - 'Integrated Ambient Intelligence and Knowledge Based Services for Optimal Life-Cycle Impact of Complex Manufacturing and Assembly Lines', Public report, 2007. Stokíc, D., Kirchhoff, U., Cordes, J., “Application of Lean Production Principles in Implementing CIM Technologies in European Industrial Enterprises”, Proceedings of the CIM-EUROPE Conference, Amsterdam, May 1993. Self-Learning, EU project NMP-2008-228857 “Reliable self-learning production system based on context aware services”, Public report, 2010. .L. Bittencourt, R. Bonefeld, S. Scholze, D. Stokic, M.K. Uddin, and J.L. Martinez Lastra, “Energy efficiency improvement through context sensitive self-learning of machine availability,” in IEEE Int. Conf. on Industrial Informatics INDIN, Lisbon, Portugal, IEEE, 2011.