MAS and SOA: A Case Study Exploring Principles and Technologies to Support Self-Properties in Assembly Systems

Transcription

1 2008 Second IEEE International Conference on Self-Adaptive and Self-Organizing Systems Workshops MAS and SOA: A Case Study Exploring Principles and Technologies to Support Self-Properties in Assembly Systems Luis Ribeiro, Jose Barata Electrical Engineering Department New University of Lisbon Monte de Caparica, Portugal {ldr, jab}@uninova.pt Armando Colombo Schneider Electric Seligenstadt, Germany awcolombo@ieee.org Abstract Holonic Assembly systems require intelligent and distributed IT support. Traditionally this has been achieved through the use of multiagent systems (MAS). Recently a significant research effort has been invested in applying Service Oriented Architectures (SOA) in automation domains. MAS and SOA are often perceived has competing concepts and technologies. In this application oriented paper an informal comparative analysis between MAS and SOA, derived from the authors experience, is presented with the goal of unveiling their strengths and weaknesses in supporting self-properties. A case study demonstrating selfmonitoring/diagnosis/healing in the assembly domain illustrates a first approach to merging the best of both worlds in a lightweight infrastructure that can be ported to embedded devices, following the trend that advocates the use of pervasive computing materialized in tiny devices for automation. 1. Introduction The advances in networked information technologies (IT) verified in recent years are inducing significant socioeconomic changes. Intelligent software entities are increasingly acting on our behalf automating processes, protecting sensitive data, filtering useful information, etc. The pervasive nature of network enabled devices, its availability and reduced cost, are progressively transforming the internet, which mainly targeted people, into a hybrid mesh of autonomous devices that seamlessly interact with each other and with humans. Only recently have these ideas hit the industrial domain traditionally conservative to technological changes. It is expected that in the future the survival of an enterprise depends on how well established is its IT infrastructure and how quickly can it accommodate changing requirements. In production domains, from which assembly is a particular case, this scenario is more dramatic as it involves machinery and shop floor changes. Emergent control approaches advocating the use of distributed and modular intelligence to maximize enterprises agility and flexibility include: bionic manufacturing systems (BMS) [1], holonic manufacturing systems (HMS) [2], reconfigurable manufacturing systems (RMS) [3], evolvable assembly systems (EAS) [4] and evolvable production systems (EPS) [5]. More than distributed platforms or technologies, Service Oriented Architectures (SOA) and Multiagent Systems (MAS) provide fundamentally new general purpose modeling metaphors. Rather than understanding MAS and SOA as competing paradigms it is the authors belief and experience that profiting from the best of both worlds presents a step further in modeling, designing and implementing self-properties to maximize robustness and performance of distributed assembly systems. The present work presents a test case in the assembly domain where each module in the shop floor is abstracted as an intelligent entity that is modeled accordingly to SOA and MAS principles and concepts. Furthermore these entities cooperate and coordinate to overcome disturbances in the shop floor relying in individual and collective selfmonitoring/diagnosis/healing. The subsequent details are organized as follows: section 2 briefly frames the problematic focusing in the assembly domain; section 3 overviews the concept of SOA and briefly surveys existing research, section 4 provides a similar analysis for MAS; section 5 sets forward a comparative study between MAS and SOA; section 6 details the infrastructure developed towards merging both /08 $ IEEE DOI /SASOW

2 concepts; and in section 7 the main conclusions are discussed. 2. Problem Framing To understand the infrastructural challenges of modern production systems it is worth describing the emergent approaches that promise to enhance shop floor s agility and flexibility. The BMS [1] are inspired in the functioning of natural organs. Organs are composed by cells and through harmonious interaction with other organs support different organisms. This approach denotes the idea of a hierarchical system where information travels bottom up and top down along the chain. The basic modeling entity is the modelon which can in turn be composed of other modelons forming a hierarchy. The system heavily relies in self-organization, passing from layer to layer DNA-type information, to gather submodelons that will be involved in the execution of a given task. A similar approach relies in the concept holon, a combination of the Greek word holos (whole) and the suffix on (part). The HMS emerged in the intelligent manufacturing systems (IMS) community as a response to the continuously changing requirements and demands posed to modern enterprises. In HMS the holons must accomplish a certain task using: coordination, communication, cooperation and negotiation while being immerse in a holarchy (a society of holons where each component behaves simultaneously as a part and a whole). An extensive review in HMS can be found in [2]. Modern industrial control paradigms recognize modularity as a key aspect to attain flexibility and agility. Research in RMS attempts to find a balance between full customization and mass production. As discussed in [6] the fundamental characteristics of RMS include: modularity, integrability, flexibility, scalability, convertibility, diagnosability, etc. Similarly to HMS and RMS, EPS and EAS rely in the modularization of hardware. Nevertheless EPS considers a finer granularity level. In this sense, the concept of module is within grippers, robots, sensor, actuators, etc; rather than cells and stations. Additionally EPS focus on production change-overs relying in biologically inspired concepts such as adaptability and evolvability. A certain set of characteristics are immediately identified as common: Modularity distributed building blocks are considered. Autonomy the modules are created independently of each other and provide selfcontained functionalities. Interaction the modules must interact according to certain rules or agreements. Structure all the modules play a certain role in the system. The definition of a structure supports the progressive encapsulation of complexity. Dynamics Most of the approaches support the idea of evolution, either by introducing changes in the environment or learning. Heterogeneity the systems targeted by these paradigms are compositions of distinct units. Current control approaches are on the edge of becoming obsolete. Emergent approaches set new challenges and requirements. General purpose modeling paradigms such as MAS and SOA currently align as the probable vehicles for embedding the conceptual frameworks hereby briefed on a ground of their own, as they by default support the features earlier detailed and additionally assure overall interoperability and integration in heterogeneous environments. The discussion in the next sections will attempt to unveil the best of MAS and SOA proposing the unification of strengths and the mitigation of weaknesses towards its application in assembly environments. 3. Service Oriented Architectures (SOA) SOA s basic building block is the service abstraction. The definition of SOA is far from being agreed as a search in the literature easily confirms. Contact points between the numerous definitions frequently include the following topics: Autonomy: there are no direct dependencies between the services. Interoperability: is specified at interface level omitting unnecessary details. Platform Independence: the services are described using interoperable XML-based formats. Encapsulation: services provide self-contained functionalities that are exposed by user defined interfaces Availability/Discovery: the services can be published in public registries and made available for general use. As an emerging modeling paradigm for distributed systems SOA is often confused with a wide range of networked information technologies. In this context, Web 193

3 Services are the preferred mechanism for SOA implementation. The Web Services Working group of the World Wide Web Consortium (W3C) defines Web Service as: a software system designed to support interoperable machine-to-machine interaction over a network. It has an interface described in a machine-processable format (specifically WSDL). Other systems interact with the Web Service in a manner prescribed by its description using SOAP messages, typically conveyed using HTTP with an XML serialization in conjunction with other Web-related standards. Although a significant share of the research in SOA focus on modeling and supporting inter enterprise relationships, there is a favorable convergence of factors that are rendering it attractive in the establishment of networks of devices namely: the availability of affordable and high performance embedded devices, the expansion and low cost of Ethernet based networks and its acceptance in the industrial domain, the ubiquitous nature of the Internet, the existence of lightweight, platform agnostic communication infrastructures, etc. This has triggered several European projects in the field including industry s heavy weights. Among them one may mention as examples: SIRENA ( award winning project that targeted the development of a Service Infrastructure for Real time Embedded Networked Applications; SODA ( creation of a service oriented ecosystem based on the Devices Profile for Web Services (DPWS) framework developed under the SIRENA project; SOCRADES ( t.html) development of DPWS-based SOA for automation systems; InLife ( including a test case that explores service oriented diagnosis on distributed manufacturing systems [7]. The Devices Profile for Web Services ( whose initial publication dates from May 2004, defines the minimal Web Service s implementation requirements for: secure message exchange, dynamic discovery, description and subscribing and eventing. DPWS specifically targets peripheral-class and consumer electronics-class hardware guaranteeing compliance without constraining richer implementations. In [8] a roadmap for SOA is presented and identifies the following research areas: Service Foundations, Service Composition, Service Management, Service Design and development. The main open challenges, also described in the roadmap include: dynamically reconfigurable run-time architectures, services discovery, autonomic composition of services and orchestration, self-* (selfconfiguring/healing/diagnosing ) services, design principles for engineering service applications. While some of these challenges can be eased by emergent standards such as BPEL4WS and WSCI others, service composition in dynamically reconfigurable run-time architectures, are harder to tackle specially in heterogeneous systems, which are typically SOA s target environments. 4. Multiagent Systems (MAS) Most definitions for agents are of functional nature and relate to their authors background and the systems under study. Nevertheless, it is possible to isolate a common set of characteristics widely accepted: Autonomy agents act individually fulfilling their individual goals. Sociability agents interact among each other establishing an intelligent society. Rationality an agent can reason about the data it receives in order to find the best solution to achieve its goal. Reactivity an agent can react upon changes in the environment. Proactivity a proactive agent has some control on its reactions basing them on its own agenda and objectives. Adaptability an agent is capable of learning and changing its behaviour when a better solution is discovered. A Multiagent System (MAS) is, in this context, a composition of several agents, each having incomplete information on a particular problem, communicating and cooperating, in a decentralized and asynchronous manner, in order to solve it. MAS results are broader than the sum of individual contributions. As stated earlier, emergent industrial paradigms are: modular, decentralized, changeable and complex. A multiagent system is by nature a decentralized and modular (and thus, easily changeable) environment, able to solve complex problems. Pilot experiments have occurred in industrial setups [9] there are, however, open challenges. Most development platforms target local area networks and implement centralized agent management systems that constitute both a bottleneck and centralized node of failure. Furthermore most agent development environments are still heavy from a computational point of view with significant memory footprints and lack operational support for real-time environments. 194

4 5. Merging MAS and SOA At a glance, both paradigms support the idea of distributed autonomous entities and provide an effective modeling metaphor for complexity encapsulation. Nevertheless SOA emphasizes contract-based descriptions of the hosted services and does not provide a reference programming model. MAS, on the other hand, support well established methods to describe the behaviour of an agent. Automation environments are typically heterogeneous and the lack of a structured development model/template may render system designing, implementation and debugging harder. Furthermore agents are regulated by internal or environmental rules that dictate social behavior and support, to some extent, flexibility and adaptability to changes in the environment. This is of major importance when considering systems that undergo dynamic runtime changes which is the case of the production paradigms earlier referred. SOA are typically supported by widely used web technologies and assure interoperability with a wide range of systems easily spawning over the internet. Most well-known MAS platforms are optimized for LAN use and are restricted to compliance with well defined but less used interoperability standards. Additionally, recent frameworks like DPWS provide high performance Web Service support for devices, with limited resources, without constraining services implementation. Most MAS platforms are computationally expensive. The discussion around the fusion of MAS and SOA is not fundamentally new. However the research focus has been in enabling agents in existing systems to request, provide or manage web services [10, 11]. In [12] a more detailed comparison between MAS, SOA and their potential of application for automation environments is presented. Self-properties in assembly systems are only now starting to be regarded as beneficial. In fact, until recently, self-organization and emergence have been forbidden concepts in the assembly domain under the argument that one would not like industrial manipulators to come up with surprises in a production line. The Evolvable Assembly and Evolvable Production paradigms have emerged to clearly explore this gap, advocating the use of biologically inspired self-organizing multiagent systems for production systems. In the next section some preliminary work on the implementation of agent-like processing on top of DPWS will be detailed. The proposed architecture heavily relies in self-monitoring/diagnosis and healing to react to failures through individual and collective behaviour. 6. A Case Study in an Assembly Installation 6.1. Reference Architecture The reference architecture considered in the following implementation falls under the scope of EPS[7] and considers intelligent modules with a fine granularity level. In this context the minimal functional entity is set at device level (gripper, robot, conveyor, etc). Furthermore each entity is abstracted and classified according to its role in the process. The following terminology is considered: Coalition Leader Service (CLS) entity in charge of aggregating and orchestrating other services; Manufacturing Device Service (MDS) abstracts each individual manufacturing component: tools, robots, conveyors, etc; Service to Machine Interface (SMI) The SMI harmonizes legacy equipment with the DPWS platform, the wrapper approach is considered. When developing a monitoring/diagnostic system there are some general requirements that should be met namely: early fault detection and diagnosis, proper fault discrimination, robustness, identification of multiple faults, fault explanation, adaptability, etc. The present architecture explores intelligence at device level (self-monitoring and diagnosis) to cope with these requirements. Interaction faults and fault propagation are typical problems that often fall beyond individual capabilities [7]. To circumvent these scenarios some entities must necessarily monitor interactions. In the present architecture that task is reserved for CLS devices. At MDS level the diagnosis uses qualitative model based reasoning. Each device senses the environment and matches the observations against the previsions established, using device s goals and behavioural model. Upon detecting and diagnosing a fault the MDS device either attempts to recover from the fault or, when that is not possible, forwards the diagnostic to the device that can best cope with the fault. This behaviour leads to the diagnosis architecture that is depicted in Figure. 1. At CLS level diagnosis is based on the model of the process. For each running task, the CLS matches the effective execution against the expected behaviour. When an MDS device reports a fault from which it could not recover from, that information is passed to the CLS. The CLS will then first check if there is a specific recovery action for the fault being reported and will attempt to execute it; otherwise it will use the default actions for a given operation. 195

5 Fig. 1 Monitoring and diagnosing faults that individual entities cannot detect or recover from In any case, recovery actions at CLS level include the collaboration of several devices under the coordination of the CLS. These mechanisms will become apparent in section 6.1. The installation where the architecture was implemented is the NOVAFLEX cell which comprises: two industrial robotic manipulators (Bosch SCARA SR800 and ABB IRB 2000), a conveyor network and an automatic warehouse. The demonstrator described in this section is to shows the feasibility and benefits of taking advantage of MAS and SOA features to implement Self-properties in distributed assembly systems. To model NOVAFLEX a DPWS stack [13] was used to provide the necessary infrastructural support delivering the inherent technological advantages. Behind each service runs an agent-like code that supports selfmonitoring/diagnosis and healing through the use of a generic service interface [14]. This interface supports the implementation of generic mechanisms defining the agents behaviour when receiving messages, of different types, and led to the development of generic code for message handling and processing. In NOVAFLEX s assembly scenario the systematic use of this interface allowed the componentization of the environment (40 services) and their interactions. The footprint of each service is 150 kilobytes. Furthermore the generic interface enables orchestration in a simple and yet effective manner. The agents behave as generic processing machines. Given a process description, each agent is able to, using the standard discovery mechanisms, identify other agents capable of executing a certain subtask and coordinate them. The notion of skill, as a functionality that an agent offers to the remaining system, is fundamental to enable light-weight orchestration. Process descriptions are provided in the form of skills that can be simple or the complex (a sequence of other skills): Each skill s, complex or simple has the form s(arga, argb, argc, rslt1,, rsltn) where argn denotes the arguments of the skill and rslt1 the results of the execution of the skill. Complex skills have the form csn s1,, st, where csn is a complex skill defined as a sequence of skills s1,, st. A complex skill is only executable if all the s in s1,, st exist and their owners have not issued any alarms. The orchestration mechanism is crucial in enabling selfproperties in the system as internally each entity reasons on what skills to use in order to overcome an occurrence. Figure 2 depicts the execution of a pick and place operation. The participant entities in the process are: a device orchestrating the operation, an external device that requested the operation, a robot and gripper. For the sake of simplicity, from this point on, these entities will be referred as agents. Figure 2. Dialogue maintained by the agents during a pick and place operation. The gripper agent is self-monitoring and detects a fault. Failing to recover from the fault, the gripper agent reports it, to the agent that is orchestrating the operation (station 2 CLS). The CLS agent suspends the pick and place operation and checking its database indentifies the process 196

6 to handle that failure and starts executing it. In the present case a gripper replacement is considered. This action includes a third party, the tool warehouse agent, which was not present during the normal pick and place operation. Further the new operation (the gripper replacement) has to be executed If the replacement is successful the initial pick and place operation is re-executed. Despite the simplicity of the example it clearly shows the ability of the system to suspend ongoing operations and using skill based re-orchestration recover from intricate fault scenarios. In the present scenario, should the gripper replacement fail to solve the problem, the system will attempt another action until all conceivable recovery rules have been excluded. In this scenario, where the dimension of the fault is completely off the action scope of the orchestrating agent, the CLS will forward the fault to an higher abstraction level were another self-healing action will be attempted until the failure is reported as unrecoverable by all existing self-healing mechanisms at all levels. 7. Conclusions Despite the advances in industrial control paradigms their effective implementation remains elusive. In order to remain high-end competitors, future enterprises will necessarily have to embrace innovative production paradigms to support agility and flexibility. However emergent approaches have demanding IT infrastructural requirements. General purpose paradigms, namely MAS and SOA, and related technologies seem to provide the adequate support. Nonetheless, and despite the impressive achievements verified in these domains, the support for self-properties is inexistent. The present paper demonstrates the importance of such features in assembly systems showing how a high level of self-functionality, including collaborative fault recovery, can be achieved using a light-weight SOA platform and applying agent modeling principles such as the use of structured communication mechanisms, generic processing of skills and automated interaction. The outcome was a reduced footprint framework that avoids reprogramming supporting scalability and plugability as denoted by the 40 interacting agents when the cell is operating. Although this work shows promising results the authors believe that further research needs to be carried out, given the complexity of the systems being considered, to fully understand and formalize the dynamics of such systems and perfect the technological IT platforms supporting future s assembly systems. References [1] K. Ueda, "A concept for bionic manufacturing systems based on DNA-type information," in PROLAMAT Tokyo: IFIP, [2] R. Babiceanu and F. Chen, "Development and applications of holonic manufacturing systems: a survey," Journal of Intelligent Manufacturing,, vol. 17, pp , [3] Y. Koren, U. Heisel, F. Jovane, T. Moriwaki, G. Pritchow, A. G. Ulsoy, and H. Van Brussel, "Reconfigurable Manufacturing Systems," CIRP Annals - Manufacturing Technology, vol. 48, pp , [4] M. Onori, J. Barata, and R. Frei, "Evolvable Assembly Systems Basic Principles," in Conference on Information Technology for BALANCED AUTOMATION SYSTEMS in Manufacturing and Services Ontario, Canada: Springer, [5] J. Barata, R. Frei, and M. Onori, "Evolvable Production Systems Context and Implications," in International Symposium on Industrial Informatics Vigo: IEEE, [6] E. Hoda, "Flexible and reconfigurable manufacturing systems paradigms," International Journal of Flexible Manufacturing Systems (Special Issue: Reconfigurable Manufacturing Systems), vol. 17, pp , October, [7] J. Barata, L. Ribeiro, and M. Onori, "Diagnosis on Evolvable Production Systems," in International Symposium on Industrial Electronics Vigo: IEEE, [8] M. P. Papazoglou, P. Traverso, S. Dustdar, F. Leymann, and B. J. Krämer, "Service-Oriented Computing: A Research Roadmap," in Dagstuhl Seminar Proceedings (SOC), [9] V. Marik, P. Vrba, K. H. Hall, and F. P. Maturana, "Rockwell automation agents for manufacturing," in Fourth international joint conference on Autonomous agents and multiagent systems, The Netherlands, 2005 pp [10] D. Greenwood and M. Calisti, "Engineering Web Service - Agent Integration," in 2004 IEEE International Conference on Systems, Man and Cybernetics, [11] M. Lyell, L. Rosen, M. Casigni-Simkins, and D. Norris, "On software agents and Web services: Usage and design concepts and issues," Proceedings of WSABE2003, [12] L. Ribeiro, J. Barata, and P. Mendes, "MAS and SOA: Complementary Automation Paradigms," in Innovation in Manufacturing Networks. vol. 266/2008, A. Azevedo, Ed.: Springer Boston, 2008, pp [13] F. Jammes and H. Smit, "Service-oriented architectures for devices - the SIRENA view," in International Conference on Industrial Informatics, Perth, Western Australia, 2005, pp [14] L. Ribeiro, J. Barata, A. W. Colombo, and F. Jammes, "A Generic Communication Interface for DPWS-based Web Services," in IEEE International Conference in Industrial Informatics INDIN Daejeon, Korea: IEEE,