Embedded holonic fault diagnosis of complex transportation systems

Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Contents lists available at SciVerse ScienceDirect

Engineering Applications of Artificial Intelligence journal homepage: www.elsevier.com/locate/engappai

Embedded holonic fault diagnosis of complex transportation systems$ Antoine Le Mortellec a,b, Joffrey Clarhaut a,b,n, Yves Sallez a,b, Thierry Berger a,b, Damien Trentesaux a,b a b

University of Lille Nord de France, F-59000 Lille, France UVHC, TEMPO Lab, ‘‘Production, Services, Information’’ team, F-59313 Valenciennes, France

a r t i c l e i n f o

abstract

Article history: Received 29 November 2011 Received in revised form 16 July 2012 Accepted 3 September 2012 Available online 16 October 2012

The use of electronic equipment and embedded computing technologies in modern complex transportation systems continues to grow in a highly competitive market, in which product maintainability and availability is vital. These technological advances also make fault diagnosis and maintenance interventions much more challenging, since these operations require a deep understanding of the entire system. This paper proposes a holonic cooperative fault diagnosis approach, along with a generic architecture, to increase the embedded diagnosis capabilities of complex transportation systems. This concept is applied to the fault diagnosis of door systems of a railway transportation system. & 2012 Elsevier Ltd. All rights reserved.

Keywords: Embedded diagnosis Holonic architecture Cooperative fault diagnosis Model-based diagnosis Corrective maintenance Railway transportation system

1. Introduction To deal with the complexity of modern transportation systems (e.g., commercial aircrafts, trains, ships), an efficient maintenance strategy is essential for maintaining and improving the availability and reliability of assets, while minimizing maintenance and total life-cycle ownership costs (Discenzo et al., 2002). Normatively, maintenance is classified as preventive or corrective maintenance (DIN EN 13306, 2010). While preventive maintenance focuses on reducing the probability of failures by replacing components before they fail, corrective maintenance has the objective of returning an item back to service after a failure occurrence. Before any corrective actions can be taken, one of the most time-consuming step of corrective maintenance is the fault diagnosis procedure, which consists of identifying the faulty components to be repaired (Feldman, 2010; Khol and Bauer, 2010). Unlike diagnosing complex industrial systems and static machines, several considerations must be taken into account when diagnosing complex transportation systems. In this paper,

$ This paper was selected from the proceedings of the 4th International Conference on Industrial Engineering and Systems Management IESM 2011. n Corresponding author. Tel.: þ33327511321. E-mail address: [email protected] (J. Clarhaut).

0952-1976/$ - see front matter & 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.engappai.2012.09.008

we assume that complex transportation systems are characterized by the following properties:  System complexity—a complex transportation system is assumed to be decomposable into a set of interacting subsystems, composed of a control part and a controlled part. These sub-systems are designed by several suppliers, using computational and physical components, and heterogeneous technologies (e.g., electrical, mechanical, hydraulic, pneumatic, hardware and software parts) (Dievart et al., 2010).  System variability—the sub-systems may differ from one system to another (e.g., change of suppliers, design changes, product evolutions) (Azarian et al., 2011).  System environmental context—each sub-system is assumed to have its own context, which can be either physical (e.g., climate impact, vibrations, electromagnetic disturbance) or informational (e.g., functioning mode, component states) (Monnin et al., 2011).  System maintainability—a complex transportation system is usually linked to a stationary maintenance center and needs to communicate with it (Jianjun et al., 2007). In addition, maintenance operations cannot be executed immediately in the system (Umiliacchi et al., 2011). To allow the diagnosis of this kind of complex transportation systems, a diagnosis system must be defined. This diagnosis system must not interfere with the normal operation of any

228

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

sub-system. This no-intrusiveness constraint is mandatory. In this paper, a diagnosis system is assumed to fulfill the following requirements:

in a real train. Finally, Section 7 offers conclusions and perspectives for future research.

 Accuracy—the diagnosis system must be adapted for isolating uniquely the faulty components among various interconnected sub-systems.  Ease of explanations—the diagnosis system must allow a user to understand how the diagnosis procedure came to the results.  Adaptability—the diagnosis system must be adapted when changes in components and sub-systems design occur.  Reactivity—diagnosis results must be delivered in a timely manner to the maintenance center in order to improve the maintenance management.  Confidence—the diagnosis system must avoid producing false alarms.

2. Condition monitoring for diagnosing transportation systems

Through the continued advances in infotronics and in communication technologies (Clarhaut et al., 2011), intelligent diagnosis systems based in particular on artificial intelligence (AI) and multi-agents (MAS) allow the diagnosis procedure to be automated on-line, observing continuously the system (Campos, 2009; Ng and Srinivasan, 2010). In this context, the main contribution of this paper is to propose a fault diagnosis approach that supports all the previously introduced assumptions. This approach is applied within a joint research-industry project, called SURFER (SURveillance active FERroviaire, translated as ‘‘active train monitoring’’), led by Bombardier-Transport. The aim of the SURFER project is to provide a more advanced solution for the on-line diagnosis of incipient failures and faults that can occur during the train service, besides the existing ORBITA system developed by Bombardier-Transport in 2006 (Orbita-BT, 2006). This paper is organized as follows. Section 2 presents a literature review about condition monitoring and diagnosis standards, along with the main diagnosis methods. This section highlights the limits of currently-used diagnosis approaches, emphasizing the need of a robust embedded diagnosis. Section 3 proposes an embedded decentralized cooperative fault diagnosis approach, based on a generic holonic model. Section 4 applies the proposed embedded diagnosis model for advanced fault diagnosis of train door systems, within the context of the SURFER project. Section 5 presents the experimental platform used for the implementation of the holonic diagnosis architecture. Section 6 exhibits the first results obtained in the implementation

A condition monitoring system involves raw data acquisition, processing, analysis and interpretation of faults to provide useful maintenance information (Campos, 2009). In this section, we refer to condition monitoring as a means of applying on-line fault diagnosis procedure to a complex transportation system, focusing on diagnosing abrupt faults rather than diagnosing incipient faults or performance degradation. Then, the literature is surveyed on condition monitoring, diagnosis standards, and diagnosis architectures. Next, the main diagnosis methods are analyzed. According to the requirements of a diagnosis system, first a diagnosis architecture is chosen, followed by the selection of a relevant diagnosis method. 2.1. Condition monitoring and diagnosis standards The reference standards for condition monitoring in industry and transportation have been registered under the ISO 13374. This standard defines a generic model of a condition monitoring architecture, using six-layer processing blocks (ISO 13374-1, 2003). These successive layers progress from raw data acquisition to useful maintenance advisories, as the data evolve into information. The layers defined by this standard are: (1) data acquisition, (2) data manipulation, (3) state detection, (4) health assessment, (5) prognostic assessment and (6) advisory generation. In the prescribed model, the first three layers are assumed to be technology-dependent on the monitored system. From there, the diagnosis layer #4 (health assessment) may handle incipient or abrupt faults, while the prognosis layer #5 is specific for incipient faults analysis (IAEA, 2007). The last layer aims at delivering recommendations on maintenance actions or operational changes based on information delivered by lower layers. 2.2. Condition monitoring and diagnosis architectures Focusing on the diagnosis layer of the prescribed model (i.e, layer #4), Fig. 1 illustrates the two fundamental partitioning approaches: off-board diagnosis and on-board diagnosis (Alanen et al., 2006; Bengtsson, 2003).

Fig. 1. Diagnosis partitioning: (a) off-board diagnosis and (b) on-board diagnosis.

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

2.2.1. Off-board diagnosis In the off-board diagnosis approach, layers 1–3 exist in the on-board system, while layers 4–6 are performed on-ground. This approach is presently widely used in transportation (Umiliacchi et al., 2011). Essentially, the diagnosis procedure centrally stores and processes the raw data and gives alarms in the remote center. This is called as Remote Centralized Diagnosis (RCD) (Fig. 1(a)). RCD is a common standard solution that is easy to support with a traditional company information system, taking full advantage of the available remote computational capabilities. When considering the requirements for the diagnosis system, some drawbacks of the RCD approach arise:  Massive amounts of raw data are sent to the maintenance center, which implies transmission constraints (e.g., data loss, transmission delays and limited bandwidth). In addition, the maintenance workers often have difficulty finding the exact origin of a failure in this mass of raw data, involving time-consuming data-mining work (Khol and Bauer, 2010). These raw data only represent symptoms and do not take into account the equipment’s evolution (e.g., wear, component replacement). As a result, the delays from a problem’s occurrence to its diagnosis solution cannot be ignored, and maintenance interventions cannot be scheduled efficiently.  The ‘‘fugitive’’ faults are not handled correctly, which implies a significant diagnosis effort has to be engaged without identifying the cause of the failure and without any real improvement.  False alarms are common because the sensors monitoring the equipment do not take the context into account. For example, if a sensor connected to a piece of equipment senses abnormal vibrations, problems with internal components may be suspected. However, if another piece of equipment in the same complex system senses the same vibrations at the same time, the vibrations are probably due to an environmental factor. In the ORBITA railway system (Orbita-BT, 2006), data about sub-system states (e.g., air conditioning, brakes, doors) are collected in real-time from the vehicle and are sent in a remote maintenance center, where they are analyzed for troubleshooting purposes. A huge amount of data is sent to a central database via General Packet Radio Service (GPRS) or wireless links in the railway stations, involving a very significant delay between the fault occurrences and the appropriate diagnosis.

2.2.2. On-board diagnosis In the on-board diagnosis approach, layers 1–4 of the ISO 13374 are embedded inside the transportation system (Benedettini et al., 2009), while layers 5 to 6 are performed off-board (Fig. 1(b)). This provides more reactivity, allowing raw data measurements to be analyzed during the system’s operation, reducing communication costs considerably. Only diagnosis information is transmitted (Byington et al., 2003). According to the main approaches in the literature for diagnosing distributed systems (Roychoudhury et al., 2009; Ferrari, 2009), on-board diagnosis architectures are divided into three major categories: centralized, decentralized, and decentralized & cooperative diagnosis. In Embedded Centralized Diagnosis (ECD), a central diagnosis entity, called diagnoser, is responsible for computing the raw data measurements acquired from the sub-systems, and delivering a global diagnosis. In the aeronautical field, Lefebvre et al. (2007) computed centrally on-board the time when the alarm occurs by various sub-systems. However, ECD can be unrealistic when dealing with large complex systems because the system is too complex to be diagnosed as a whole (Pencole´ and Cordier, 2005). Also, ECD architecture has several limitations linked to on-board

229

computational requirements (e.g., large amount of data to be analyzed) and the modularity of the architecture (Kurien et al., 2002). Alternatively, embedded diagnosis can also be deployed in a decentralized way. This architecture, called Embedded Decentralized Diagnosis (EDD), promotes closer computation, in which a diagnoser is assigned to each sub-system. With this architecture, each diagnoser is responsible for delivering a local diagnosis of its monitored sub-system, which is then reported to the remote maintenance center. Usually, some fusion mechanisms (e.g., coordinator, supervisor) are implemented in order to merge these local diagnoses (Nasri et al., 2012). However, each diagnoser operates independently, without communication among diagnosers (Qiu and Kumar, 2004). This architecture can have significant drawbacks, especially when a fault with the same origin is detected and reported by several diagnosers. In order to deal with the lack of robustness of decentralized approaches, local diagnosers could evolve toward proactive entities, by exchanging information during the diagnosis task execution. This architecture is called as Embedded Decentralized & Cooperative Diagnosis (EDCD). Provan (2002) proposed an example of EDCD architecture for diagnosing distributed systems. These authors employ generic diagnosers, along with a message-passing algorithm, for delivering a system-level diagnosis. In the aerospace industry, Koutsoukos et al. (2010) and Dievart et al. (2010) presented other examples of EDCD implementations. The cooperation among diagnosers can prevent from erroneous fault detection and misdiagnosis, by taking into account current environmental and operational information, as well as status of neighboring sub-systems (e.g., stress on a specific sub-system, condition of other similar sub-systems, mission profile). As a result, only diagnosis information, without any false alarms, is transmitted to the remote maintenance center. As a summary, Fig. 2 proposes a typology organized according to three axes: (1) embeddability, (2) diagnoser mapping, and (3) organization. Focusing primarily on embedded diagnosis, the last two axes are only considered for on-board diagnosis architectures. As stated previously, the lack of robustness and the introduction of significant delays in data analysis, preventing corrective maintenance reactivity, make the RCD architecture inefficient for transportation systems. Among on-board diagnosis architectures, EDCD is the most appropriate architecture for preventing false alarms and delivering pertinent diagnoses. Thus, the EDCD architecture is chosen. The next sub-section describes different methods for fault diagnosis. Section 2.4 presents an analysis of the potential benefits for the EDCD architecture that was chosen. 2.3. Relevant methods for fault diagnosis Fault diagnosis is an important research domain, mainly studied by the Artificial Intelligence (AI) and Fault Detection & Isolation (FDI) communities (Cordier et al., 2000; Simeu-Abazi et al., 2010). Fig. 3 presents two major categories based on the classification schemes proposed by Venkatasubramanian et al. (2003a) and Katipamula and Brambley (2005). The first category (Model-based methods) uses quantitative or qualitative models. Quantitative models are based on explicit mathematical relationships describing the physical system to be diagnosed; qualitative models exploit more qualitative relationships describing the system. Unlike the first category in which a priori knowledge of the system is assumed, the second category (Process-Historybased methods) employs a huge amount of historical data to perform diagnosis. The significant diagnosis methods belonging to these categories are introduced below.

230

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Fig. 2. Proposed typology of diagnosis architectures for complex transportation systems.

Fig. 3. Classification of diagnosis methods.

2.3.1. First Principle Reasoning (FPR) At the heart of any First Principle Reasoning method is the comparison between the observed behavior of the system and its predicted behavior, inferred by an explicit model of the system (De Kleer and Williams, 1987). The system’s faulty behavior leads to discrepancies when the observed behavior differs from the expected behavior. Then, a list of possible fault candidates that might explain these discrepancies is computed. The FPR’s accuracy is closely linked to the quality and the validity of the system model used. Azarian et al. (2011) present an example of an FPR method and its performance for an automotive application.

2.3.2. Fault trees (FT) Fault trees are cause/effect models used for analyzing system safety and reliability. A fault tree graphically represents the fault interaction within a system. Basic events (e.g., component or human faults) at the bottom of the tree are linked via logic symbols, called gates, to one or more top events, called dreaded events. Diagnosing such events tests each of the possible causes,

going through the fault tree gates until a root cause is identified. Lefebvre et al. (2007) used fault trees for on-line fault diagnosis in aeronautics. These authors did experiments with dynamic fault trees for modeling temporal dependencies between alarms and failure rules. However, the accuracy of fault-tree diagnosis is limited to analyzing common-cause failure, and it is more a safety/risk assessment method rather than a pure fault diagnosis method (Ribeiro and Barata, 2011).

2.3.3. Rule-based reasoning (RBR) Rule-based reasoning methods are used to solve diagnosis problems by using human expert knowledge. These methods implement a knowledge base that represents the required expertise to diagnose the system, usually described as a set of rules (e.g., if-then-else rule statements (Zaki et al., 2007)). This knowledge is then used by the inference engine to assess abnormal behaviors, providing possible explanations to maintenance operators. RBR methods need enough experience about the system to constitute the knowledge base, which can be difficult to maintain

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

and expand. They are limited to already known faults (Rao et al., 1998). Li and Shen (2007) present examples of expert systems for diagnosing avionic devices. 2.3.4. Case-based reasoning (CBR) In case-based reasoning methods, a knowledge base containing a set of experiences is used to solve new diagnosis problems. An important feature of the CBR methods is their learning capabilities (Berenji et al., 2005). Unlike rule-based systems, the knowledge base is enriched automatically through experience. In order to solve new diagnosis problems, the system’s observed behavior is matched with similar cases that have already been solved or remain unsolved. A solved problem allows the diagnosis system to reuse the solution for future similar cases. Otherwise, the system proposes possible explanations by retrieving past cases that are most relevant for the observed situation. In transportation systems, Vong et al. (2011) worked on a CBR method for diagnosing automotive engines. 2.3.5. Artificial neural networks (ANN) Inspired by the structure of biological neural networks, Artificial Neural Networks are layered, interconnected processing elements, used for modeling complex relationships between inputs and outputs, capable of non-linear modeling and pattern recognition. In fault diagnosis, ANNs are attractive for their learning capabilities and useful for classification tasks. Several ANN structures can be used for fault diagnosis, such as multilayer perceptions and radial basis function networks (Yang et al., 2002). When used for fault classification, ANNs are timeconsuming for training, and maintenance operators risk not understanding the diagnosis reasoning. 2.3.6. Support vector machines (SVM) Support Vector Machines are learning machines that are able to recognize patterns. They are mainly used for fault classification tasks. After a learning phase, the classification is carried out following decision planes for separating data into two or more classes. In fault diagnosis, SVMs are widely used for recognizing fault patterns using data features. Ganyun et al. (2005) present examples of SVMs methods for fault diagnosis. Although SVMs provide accurate results, these methods remain computationally costly for classification of various faults, and they require additional methods for extracting features from raw data measurements. 2.3.7. Bayesian networks (BN) Bayesian networks are based on a directed acyclic graph, in which each node of the network represents an event, and the arcs between nodes rely on conditional probabilities. An important feature of BNs is their ability to model uncertainties, as well as their learning capabilities. In fault diagnosis, BN classifiers are primarily established using expert knowledge for encoding probabilistic information between the nodes, which represent either observable or non-observable events, in order to compute the probabilities of the faults. (Ferreiro et al., 2012) present several examples of BN applications in fault diagnosis. 2.4. Comparative analysis of diagnosis methods for the EDCD architecture Inspired by the feedback of Ribeiro and Barata (2011) and Venkatasubramanian et al. (2003a,b), Table 1 summarizes the advantages and limitations of the previously cited methods when applied with the EDCD architecture. Six criterias were retained for the comparative analysis, corresponding to the rows of the table.

231

Table 1 Strengths and limitations of diagnosis methods for the EDCD architecture. Diagnosis requirements

Accuracy Ease of explanations Adaptability Reactivity Confidence Embeddability 1 2 3 4 5 6 7

Diagnosis methods FPR

FT

RBR

CBR

ANN

SVM

BN

þþ1 þþ1,7 þ 2,7 þ 1,2 þþ7 þ7

þ3 þþ3 03 þ3 þþ3 þ3

þ 2,3 þ3 þ 2,3 þþ3 þ3 þþ3

þþ1,3 01,3 þ 1,3 þþ1,3 þþ3 þþ3

þ 3,6 03,6 þ 3,6 þþ3,6 þþ3 þþ3

þþ4 þ5 þ4 þ5 þ5 þ 4,5

þ3 þ3 þ3 þ3 þþ3 þþ3

Azarian et al. (2011). Zaki et al. (2007). Ribeiro and Barata (2011). Widodo and Yang (2007). Vong et al. (2011). Venkatasubramanian et al. (2003b). Chittaro and Ranon (1999).

The first five are issued from the requirements presented in the introduction. To deploy EDCD architecture, an additional embeddability criterion was examined (e.g., algorithm complexity, response time, memory size). FPR methods present the advantage of accuracy, delivering useful explanations with reasonable computing time (Azarian et al., 2011, Chittaro and Ranon, 1999). However, the model is dependent on the system and can be time-consuming to develop (Zaki et al., 2007). FT is based on a reliability analysis, which is highly dependent on the system properties. Although FT provides explanation and confidence, the accuracy is closely linked to the reliability assessment. In addition, FT lack in adaptability and computation (Ribeiro and Barata, 2011). RBR methods are very fast for isolating faults since fault symptoms are matched directly with the expert knowledge, but they lack the diagnosis of unforeseen faults (Zaki et al., 2007). In addition, the adaptability can be difficult when revising the rules in case of system change (Ribeiro and Barata, 2011). The CBR methods provide accurate results in a timely manner, but need enough feedback about the system. In addition, CBR is system-dependent and delivers poor explanations about the results obtained (Azarian et al., 2011). ANNs provide confident rapid isolation time, but they have several limitations for the ease of explanations and adaptability (Ribeiro and Barata, 2011; Venkatasubramanian et al., 2003b). SVMs are highly accurate methods for fault classification with fast response time, but require prior knowledge on fault features and tends to be memory intensive for multi-class classification (Widodo and Yang, 2007). BNs provide confident results, assuming that preliminary stochastic knowledge is available to initialize the model (Ferreiro et al., 2012; Ribeiro and Barata, 2011). The current phase of the SURFER project is focused on the commissioning and commercializing new trains. As a result, the process-history-based methods cannot be applied because no feedback is available. As shown in Table 1, among the model-based candidates, the FPR and RBR methods are the only ones that provide the most numerous advantages and are compatible with embeddability criterion. Due to this train commissioning phase, it is important to validate the model used for the diagnosis. This involves understanding accurately how the diagnosis procedure finds the faulty components. Consequently, accuracy and ease of explanations become major criteria, and finally leading us to choose the FPR method. This does not prevent considering other interesting embeddable methods (e.g., RBR, CBR or SVM) in future phases

232

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Fig. 4. Holonic architectural model for embedded diagnosis.

of SURFER, for example, during the commercial train operations, where feedback is available.

method f i (Eq. (1)). Di ¼ ðfeip =0 op r NDi g,fLig =0 o g rcardfeip ggÞ ¼ f i ðC i ,Si ,M i ,Dni Þ

3. Proposed approach for the embedded fault diagnosis of complex transportation systems This section describes the proposed generic approach for diagnosing complex transportation systems, by combining the EDCD architecture and the FPR method. The holonic view clarifies and expands the usual systemic framework (Koestler, 1967). Its major advantages are:  a holon is capable of autonomous decision-making in relation to its objectives and cooperation with other holons (this is one way towards an autonomous, cooperative diagnosis).  a complete holon is composed of physical and informational parts, making it possible to model the two important aspects of complex transportation systems.  the holonic view is associated to recursive decomposition (i.e., a holon is both a part of a whole and a whole composed of other holons), thus meeting the systemic principle of decomposition. The capacity of autonomy and cooperation of the holonic view is very pertinent for supporting proactive diagnosers and information exchange during the diagnosis task execution (McFarlane et al., 1995; Jarvis and Jarvis, 2003; Vinatoru, 2008). This holonic view is very appealing for designing the EDCD architecture proposed in the sub-Section 2.2.2.

3.1. Architectural model for embedded diagnosis The generic holonic architectural model proposed for the diagnosis of a complex system is depicted in Fig. 4. Based on our experiments in manufacturing systems (Sallez et al., 2010; Trentesaux, 2009), this model is composed of recursive diagnosis structures. This generic model shows the decomposition of the system into sub-systems and their associated diagnosis methods. As shown in Fig. 4, a system Si is decomposed in sub-systems. Sij is a j sub-system of the system i. For example, S12 is a subsystem of S1 , and S112 is a sub-system of S11 . At one level in the decomposition of the system in sub-systems, one system Si is associated with its context and is diagnosed by a diagnosis

ð1Þ

where: Di is the result of the diagnosis for the system Si , which consists of a set of discrepancies, denoted feip g, together with the set of the lists of the possible faulty components, denoted fLig g, that may explain these discrepancies, NDi being the total number of discrepancies that might occur in the system i. C i denotes the n o context associated to the Si system. iÞ is the set of all sub-systems of Si . Si ¼ ScardðS ijj ¼ 1 M i is the model of Si according to the system’s abstraction level (e.g., structural and/or behavioral model). Dni ¼ fDij , j A ½1; cardðSi Þ, j A Ng denotes all the diagnoses results (i.e., eij ,Lij ) recursively issued from the different sub-systems composing Si . cardðSi Þ is the total number of sub-systems of the system Si . For each triplet ðf i ,C i ,Si Þ of the holonic architecture, a holon Hi is defined. For example, in Fig. 4, the triplet (f 1 ,C 1 ,S1 ) constitutes the holon H1 . Moreover, each holon Hi is characterized by its two main properties: Autonomy: at level i, a holon elaborates autonomously a diagnosis (Di ) by using a model-based fault diagnosis method (f i Þ. This diagnosis is performed based on observation of its Si system and its associated context C i . The accuracy of the diagnosis is closely linked to the quality of the system observations, so the diagnosis has to be executed as close as possible to each sub-system. Cooperation: at level i, in order to elaborate the diagnosis Di , a holon can interact with other holons located on level i, and with the holons on level i 1 and iþ1 of the holonic architecture. This collaboration occurs in a collaboration space (see Fig. 4). A collaboration space represents an area where an information exchange takes place for a diagnosis, taking into accountC i , Si and Dni . 3.2. Diagnosis method for embedded diagnosis A systemic and functional view of the diagnosis method is shown in Fig. 5. Each diagnosed system Si is composed of a control part and a controlled part, which operate in a context C i (block (1) in Fig. 5). The control part executes an algorithm to control the controlled part and, in return, the controlled part adopts an expected behavior. At the lowest level of the holonic structure, the controlled part is typically composed of physical elements (e.g., sensors, switches and actuators) that are linked to mechanical and electrical constraints.

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

233

Fig. 5. Diagnosis method for a holon at level i and exchanges with other holons at different hierarchical levels.

Fig. 6. Petri Nets model of the context analysis process, shown in Fig. 5.

The system Si may be affected by failures either in the control part or in the controlled part. Possible faults for the control part concern physical faults (e.g., hardware faults) or design faults (e.g., software bugs). For the lowest level of the holonic structure, only physical faults are considered (e.g., broken sensors and actuators, component hardware or electrical discontinuities). Two cases can be distinguished when the holon executes the fault diagnosis method f i , due to the holonic architecture used:

 The holon analyzes the observed behavior of the system Si continuously. Based on processed data and the system model



(Mi in Fig. 5), the holon detects any abnormal behavior (i.e., fault symptoms) and generates discrepancies (output of the block (2), denoted ei in Fig. 5). These discrepancies are then analyzed by the diagnosis algorithm (block (3) in Fig. 5), which uses the system model to provide a list of the potential faulty components related to these discrepancies (Di ). The holon receives diagnoses from holons of level i  1 (Di *).

The context analysis process (block (4) in Fig. 5) is one of the innovative features of the proposal. Fig. 6 shows this process, which is modeled using Petri Nets. In cases of ei or Di *, the holon

234

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Fig. 7. The proposed holonic EDCD approach and the FPR method applied to a train system.

analyzes the system context and exchanges with other holons to elaborate the system’s enriched context, which is useful for refining the diagnosis results. If the enriched context explains Di * or ei , the resulting diagnosis at level i (Di ) is inhibited for the related Di * or ei to prevent a misdiagnosis. If the enriched context at the system’s level i does not provide an explanation to Di * or ei , the resulting diagnosis at level i (Di ) is not inhibited. Then, the holon’s diagnosis results (Di ) are sent to the holons at level iþ1. Some features of the well-known subsumption architecture (Brooks, 1986) are reflected in the inhibition principle. Another approach would use a previous filtering principle instead of the inhibition principle, used once a diagnosis has been computed. This filtering principle would be executed before the diagnosis process (block (3) in Fig. 5). This approach will be considered in future research for a comparative performance analysis. The main difference with the inhibition principle is that, in the filtering principle, no diagnosis occurs if the context analysis does not allow it. In the inhibition principle, even if a context analysis inhibits the diagnosis, the diagnosis has nevertheless been made.

4. Applying the holonic approach to a railway transportation system The following sections illustrate the corresponding instantiation in the passenger access function, which is a safety critical function (Umiliacchi et al., 2011).

4.1. Instantiation of the holonic approach for the diagnosis of a train system A train is a complex transportation system (Si ) that has several vehicles, according to the model, denoted Sij , Si0 j0 . Each vehicle is composed of a certain number of distributed sub-systems (e.g., Sijk , Si0 j0 k0 ), including door sub-systems located at either side of the vehicle (Fig. 7). Each part of the train system has its corresponding context (e.g., C i , C ijk , C i0 j0 k0 ). A passenger door sub-system is composed of a door control part (i.e., door control unit) and a controlled electromechanical part (i.e., a door mechanism and one or two moving steps). All the door sub-systems receive control information (e.g., closing order) from the centralized vehicle

control system. Each vehicle control system in the train receives information from the centralized train control system. The centralized train control system, the centralized vehicle control system and each door sub-systems are equipped with a holon (e.g., Hi , Hij , Hi0 jk ) that monitors and diagnoses its systems or sub-systems. Due to the hierarchical decomposition of the train system, the holon in the centralized train control system has a higher functional level than holons in the centralized vehicle control system. Similarly, vehicle holons have a higher functional level than the holons in the door sub-systems. Each diagnosis method at the lowest holon level (e.g., f ijk , f i0 jk ) implements a specific FPR diagnosis method, developed by PROSYST, a company involved in the SURFER project. This innovative method, called ‘‘automatic diagnosis’’, has been successfully applied to solve diagnosis problems in industrial applications (Willaeys and Asse, 2011). This method has the objective to achieve a complete automation of the fault detection and the fault isolation steps, using the structural model of the monitored system. In this method, the systems to be diagnosed belong to the class of discrete events systems, since the signals exchanged between the control part and the controlled part consist of a timed sequence of discrete events. The ‘‘automatic diagnosis’’ is implemented in each holon in each passenger door sub-system (i.e., the architecture’s lowest holon level) and is not implemented for holons at higher levels. The holon analyzes the observed behavior of the system Si0 jk continuously (block (1) in Fig. 7). The outputs of the door control part, denoted Oi0 jk , allow the expected future information to be predicted based on the system model (block (2) in Fig. 7). At the same time, if one or more components of the electromechanical part have failed, discrepancies appear (denoted ei0 jk in Fig. 7) between the predicted set of information and the monitored feedback information, denoted Ii0 jk . These discrepancies are automatically analyzed by the diagnosis algorithm (block (3) in Fig. 7). The next section describes the context analysis algorithm, which is the most innovative feature in the proposal. 4.2. The train system’s context analysis algorithm and the inhibition principle The algorithm described below implements the inhibition principle in a holon on the level ij (i.e., the vehicle system level) in this example. It is assumed that holons on the same hierarchical level (e.g., train door sub-systems) cannot cooperate with

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

each other. Holons cooperate only with holons on a higher level. Once the system has been initialized, a holon monitors its system continuously, whatever the occurring events (i.e., discrepancies or diagnoses).

Table 2 Inhibition of discrepancies by context Cijk at level ijk (Door sub-system). Context Cijk

Algorithm 1. Holon’s behavior on level ij (vehicle system) 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26.

Begin Initialization of system Sij model Initialization of context inhibition table at level ij While (true) Monitor system Sij and wait for event (discrepancies eij OR diagnosis Dijk*) If (eij ) Then Diagnose Sij and generate Dij EndIf If (Dijk*) Then Analyze Dijk* and generate Dij EndIf Acquisition of context Cij For (each (eij OR Dijk*) of the context inhibition table at level ij) For (each Cij of the context inhibition table at level ij) Find the event (eij OR Dijk*) Find the acquired context Cij Obtain the resulted inhibition value EndFor EndFor If (inhibition¼TRUE) Then Dij is inhibited for event ðeij OR Dijk*) Else Dij is sent to the ‘‘father’’ holon at level i EndIf EndWhile End

For the train example, due to the three-level holonic view proposed (i.e., train, vehicle and doors), holons use three tables (Tables 2–4 for door diagnosis). In these tables, several discrepancies or diagnoses and various contexts may be chosen depending on the situation. Each context corresponds to its hierarchical level (e.g., door sub-system: door obstructed for level ijk; vehicle system: the vehicle is tilted at station for level ij). Discrepancies, diagnoses and contexts from these tables are obtained from railway standards and expert feedback. For level ij (vehicle system), a holon has two different ways to process the diagnosis (Fig. 6)

 In case of discrepancies, a holon diagnoses system Sij and generates Dij.

 Every time a holon receives a Dijk*, this holon analyzes this Dijk* and generates Dij. In the Petri Nets model in Fig. 6, a holon acquires its corresponding context Cij. To determine if a holon can explain eij or a Dijk* with Cij and inhibit, or not, Dij, a holon uses the context inhibition table at level ij (Table 3). The resulting inhibition is given, along with a cross analysis between eij or a Dijk* and a Cij. In fact, for a precise event (eij or Dijk*), the holon lists the corresponding discrepancy or Dijk* in the table. Then, for this precise event, the holon lists all the possible contexts, finds the acquired context Cij, and directly obtains the result (i.e., Dij inhibited or not). For example, at level ij (vehicle system), a holon uses Table 3 for the analyzing the fault symptom: Door closed with a delay (DT), the context analysis concludes with ‘‘the vehicle is tilted at station’’. Thus, this symptom can be explained, and the diagnosis is inhibited.

235

Discrepancy

eijk

Door cont. open Door cont. closed Door closed with a delay (DT)

Door obstructed

Heavy passenger loadat the doorstep

TRUE FALSE TRUE

FALSE FALSE TRUE

Table 3 Inhibition of discrepancies or Dijk* by context Cij at level ij (vehicle system). Context Cij Vehicle is tilted at station

eij

Door cont. open FALSE or Door cont. closed FALSE Dijk* Door closed with a delay TRUE (DT)

Heavy passenger load in the entire vehicle FALSE FALSE TRUE

Table 4 Inhibition of discrepancies or Dij by context Ci at level i (train system). Context Ci

ei Door cont. open or Door cont. closed Dij Door closed with a delay (DT)

Platform not detected

Train speed 4Opening threshold

FALSE TRUE FALSE

FALSE TRUE FALSE

To illustrate potential benefits of the proposed approach, some simulated scenarios of the cooperative mechanisms among hierarchically-dependent holons are presented in the following sub-section. These scenarios will also be used to validate the diagnosis algorithms embedded into holons. 4.3. Examples of cooperative mechanisms in a railway system To illustrate the dynamic behavior of the fault diagnosis methods based on the simulations of the algorithm introduced above, three possible fault diagnosis scenarios in train passenger doors are presented. A door sub-system can be affected by various fault symptoms that produces discrepancies (e.g., door is still open after a close order signal, referred to as ‘‘cont. open’’; door is closed with a delay). For these three following diagnosis scenarios, the centralized train control system (i.e., S1) is composed of two vehicles (i.e., S11, S12). Each vehicle is equipped with three door sub-systems (e.g., for S11: S111, S112 and S113). Each system and sub-system is associated with a holon (i.e., H1, H11, H12, H111, H112 and H113) and its corresponding system context (i.e., C1, C11, C12, C111, C112 and C113). 4.3.1. Scenario #1: fault on a single door (true alarm) On this example, where behaviors of holons are presented on Fig. 8, a true alarm scenario is studied. Fig. 8 presents the behaviors of the holons. Due to a faulty sensor, a single door sub-system (S111) has a specific discrepancy: the door is still opened after a close order signal. The holon H111 detects this discrepancy, diagnoses S111 and

236

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Fig. 8. Diagram sequence of scenario #1.

Fig. 9. Diagram sequence of scenario #2.

analyzes the local context (C111) in order to explain this discrepancy. Since context C111 doesn’t give an explanation, H111 generates a message to H11 with the diagnosis D111*. H11 analyzes this diagnosis and its corresponding context (C11) in order to explain diagnosis D111*. Since context C11 does not give an explanation, H11 diagnoses S11, thanks to diagnosis D111*, and generates a diagnosis D11 for H1 with the confirmed single failure mode of S111. This scenario was given to illustrate the behaviors of the holons during the fault diagnosis process. In this scenario, holon H1 sends an internally confirmed diagnosis to the maintenance center, using exchanges among hierarchical holons.

4.3.2. Scenario #2: faults on all train doors (false alarm) In this scenario, a false alarm scenario is studied. Fig. 9 presents the behaviors of the holons. Due to the train being tilted at the station because of the inclination of the track, the vehicle’s three doors (i.e., S111, S112 and S113) have the same discrepancy simultaneously: the doors take a longer time to close themselves. H111, H112 and H113 detect this discrepancy, but their limited subsystem context (C111, C112 and C113) cannot provide an explanation for this discrepancy. H111, H112 and H113 generate messages to H11 with diagnoses D111*, D112* and D113*. H11 analyzes D111*, D112* and D113* and generates a diagnosis D11. Then, H11 analyzes

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

237

Fig. 10. Diagram sequence of scenario #3.

the vehicle’s context (C11) in order to explain these diagnoses. Since C11provides an explanation (i.e., vehicle being tilted at the station) for the sublevels diagnoses, H11 inhibits the resulting diagnosis D11, for the specific diagnoses D111*, D112* and D113*, to prevent new false alarms (see Petri Nets model in Fig. 6). This second scenario was given to illustrate the efficiency of the holonic approach for eliminating false alarms. Using the analysis of its higher-level context, a holon can detect and eliminate false alarms from lower-level holons due to their limited local contexts. In comparison, a traditional decentralized embedded approach would have sent a large amount of misdiagnosis information to the maintenance center, without first verifying the vehicle context.

maintenance center or operators since there are no functional anomalies for a local door sub-system. These scenarios will be used on an experimental platform, called EMAPS (Experimental Multi-Agent Platform for the SURFER project), which is briefly introduced in the next section.

5. Experimentation platform The EMAPS platform will allow the holonic diagnosis approach to be validated before the final embedding in the trains. This platform is intended to:

 develop the various configurations depending on the architec4.3.3. Scenario #3: faults on all doors of a specific vehicle In this scenario, a common cause failure is studied. Fig. 10 presents the behaviors of the holons. Due to an electrical ground connection problem between the centralized vehicle control system and the door sub-systems, the door close order signal is sent by the centralized train control system but is not received by the door sub-system. Consequently, all the vehicle’s doors (S111, S112 and S113) remain open. H111, H112 and H113 cannot detect this discrepancy because, at their local context, the situation is normal. So, they don’t send a diagnosis to H11. H11 monitors S11 when it doesn’t receive diagnoses from lower levels. H11 detects a discrepancy (a missing closing confirmation feedback event from its sub-systems) and analyzes the context C11 in order to explain this discrepancy. Since C11 does not explain this discrepancy, H11 generates and sends diagnosis D11 to H1 with a signal reception problem between S11 and its corresponding sub-systems S111, S112 and S113. This last scenario is intended to illustrate the efficiency of the holonic approach for diagnosing common cause failures. In this scenario, due to their hierarchical level at the door sub-systems, the holons cannot detect and diagnose this kind of failure. The proposed holonic architecture, with a higher-level holon that monitors the global system when it does not receive diagnoses from lower levels, is intended to improve the detection of common cause failures. In comparison, a traditional decentralized embedded approach would not have given a diagnosis to the

 

tures proposed (i.e., mapping of monitoring and diagnosis functions to hardware systems), validate the internal algorithms of the embedded holons, which compose the train diagnosis system, assess the appropriateness and reliability of fault diagnosis processes obtained from the cooperation of several holons.

This platform will be initially tested on the door sub-systems and later on the air-conditioning sub-systems of a passenger train. As shown in Fig. 11, EMAPS will emulate two train vehicles and their corresponding doors. The Multi-Agent System (MAS) paradigm (Wooldridge, 2002) will be used to implement the holons and their algorithms. More precisely, one agent will be associated to diagnose one specific door. Each agent could cooperate with other agents through the dedicated diagnostic network. The train’s signals (e.g., process data, context data) are obtained either from real recordings or from signals generated by a train simulator. The resulting train data are available via an OPC server. The EMAPS has the following hardware and software components (Fig. 11):

 A MAS development environment, JADE, which defines various 

architectural configurations for diagnosing sub-systems (Bellifemine et al., 2007); Four computers, each one integrating the behavior of one of the four embedded diagnostic agents;

238

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Fig. 11. Architecture of the EMAPS platform.

 A train data playback system with an OPC data server, which 

allows signals to be disseminated through an emulated internal train network with a TCP/IP protocol; A supervisory system (in Fig. 11, the GUI agent), which provides the operators with an overview of the emulated train network’s signal data and the inter-agent information exchanges throughout the diagnostic network and also presents final diagnosis results.

The next section describes the first results obtained in real train’s exploitation and first lessons learnt from real experiments.

6. The implementation in a real train, first results and lessons learned Low-level diagnosis functions (e.g., monitoring layers) were implemented and embedded in real trains to monitor train signals continuously. Along with information about the system’s operating context, this monitoring system stores the raw data measurements on-board as a collection of log files. These log files provide useful support for understanding the faults that occurred while the train was in service, especially for the door subsystems. The log files analysis demonstrated that particularly abnormal behaviors can only be explained by using a continuous embedded monitoring and cannot be explained by using the current RCD Bombardier-Transport approach: Orbita (Orbita-BT, 2006). In addition to this partial implementation, a specific FPR method, based on a structural model of the door sub-system, is currently being validated off-board. This FPR method analyzes log files from trains (blocks (2) and (3) shown in Fig. 7). Although the approach is used off-board without context analysis (block (4) shown in Fig. 7) and high-level diagnosis processes, the first results are already very promising.

Table 5 Analysis time and diagnosis success rate for the different used off-board approaches. Off-board approaches

Analysis time

Diagnosis success rate (%)

RCD approach Log file with human analysis Log file with FPR method analysis

1 h to more than 6 h 4–6 h 30 min–1 h

80 95 95

Table 5 presents a comparison of the FPR method with the RCD approach and a human analysis based on a manual diagnosis making from log files. This comparison is based on two criteria:  Analysis time, corresponding to the average time for analyzing train data and determining diagnoses, and  Diagnosis success rate, corresponding to the ratio between the number of successful diagnoses obtained and the total number of reported faults. As Table 5 shows, the proposed approach with the FPR method has the same diagnosis success rate as the human analysis (95%), but it is significantly faster than the two other approaches. In addition, after three months of continuous embedded monitoring in trains in real exploitation conditions and log files analysis, BombardierTransport has calculated that the time needed to identify root causes of reported faults was reduced by an average of 40%. However, understanding and analyzing the root cause of reported faults from log files with the approach in off-board use is not enough. For example, a door lock signal is sent to a train line composed of two interconnected trains: train #1 and train #2. Due to a problem with physical data transmission between train #1 and train #2, this signal was only received by train #1. Only a manual cross-tabulation analysis of the two log files from the interconnected trains showed this transmission problem and the root cause. These results encourage us to pursue our research,

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

implementing the proposed on-board holonic diagnosis approach, including the specific FPR method and the context analysis.

7. Conclusion and future works In this paper, a holonic cooperative approach for diagnosing complex transportation systems was proposed. The main contribution of this proposal is the combination of a holonic cooperative diagnosis architecture with a FPR method for the embedded fault diagnosis. The proposed holonic architecture is based on the decomposition of a complex system into sub-systems and is composed of recursive diagnosis holons. At each level in the decomposition, each holon is characterized by its system to be diagnosed, its context and its diagnosis method used. In order to increase the diagnosis confidence and decrease the number of false alarms, a holon considers its system context and may exchange information with holons located at the same level, at a higher level, or at a lower level of the holonic architecture. The holonic approach was applied to the on-board monitoring of a train system in the scope of the industrial SURFER project. A holonic train architecture, coupled with an FPR method and relevant diagnosis algorithms, is proposed. Three diagnosis scenarios with a specific train function (i.e., passenger access) were provided, highlighting the information exchanged among holons. These scenarios illustrate the expected benefits of the approach compared to the other existing diagnosis architectures. In the near future, in order to validate the approach completely, an experimental platform, called EMAPS, is currently being implemented. This platform reproduces the behavior of a train system and will be applied to monitor various emulated sub-systems (i.e., doors, for this first milestone). The encouraging first results make us confident for the project’s continuation and provide some future research topics:

 The short-term perspectives will concern the implementation





of collaboration strategies among holons of the same hierarchical level to improve the context analysis. (In this paper, no cooperation was assumed among holons on the same level.) Another study will focus on the development of the concurrent filtering principle for the context analysis, instead of the inhibition principle proposed in this paper, and relevant comparisons will be made using the EMAPS platform. The mid-term perspectives will focus on another part of the SURFER project for the preventive maintenance (ConditionBased Maintenance) rather than corrective maintenance. This future research will integrate damage diagnosis and prognosis methods into the proposed holonic architecture. This integration will allow us to define prognosis holons for incipient fault detection and localization. In this extended approach, which combines spatially and semantically distributed diagnosis, fault diagnosis holons and prognosis holons will exchange information with each other to refine the diagnosis accuracy. The long-term perspectives are of two types. The first, more theoretical, will consist of studying other interesting embeddable diagnosis methods (i.e., RBR, CBR and SVM) and comparing with reference method FPR. The second will focus on applying the holonic approach for diagnosing other systems than doors, such as the air-conditioning system, leading to a new generalization of the promising approach presented in this paper.

Acknowledgments The project presented in this paper is led by BombardierTransport in collaboration with PROSYST, HIOLLE Industries Group,

239

Universite´ de Valenciennes et du Hainaut-Cambre´sis (UVHC) and Institut franc- ais des sciences et technologies des transports, de l’ame´nagement et des re´seaux (IFSTTAR). This research is supported financially by French Inter-ministerial Funds (FUI) and the Nord/Pas-deCalais Region, and sponsored by the i-Trans and Advancity competitiveness clusters. The authors gratefully acknowledge the support of these institutions.

References Alanen, J., Haataja, K., Laurila, O., Peltola, J., Aho, I., 2006. Diagnostics of mobile work machines. VTT Res. Notes, 23443. Azarian, A., Siadat, A., Martin, P., 2011. A new strategy for automotive off-board diagnosis based on a meta-heuristic engine. Eng. Appl. Artif. Intell. 24, 733–747. Bellifemine, F.L., Caire, G., Greenwood, D., 2007. Developing Multi-Agent Systems with JADE. Wiley. Wiley series in Agent Technology Benedettini, O., Baines, T.S., Lightfoot, H.W., Greenough, R.M., 2009. State-of-theart in integrated vehicle health management. Proc. IMechE, Part G: J. Aerosp. Eng. 223 (2), 157–170. Bengtsson, M., 2003. Condition Based Maintenance On Rail Vehicles, Possibilities for a more effective maintenance strategy. Technical Report, Eskilstuna, Sweden. Berenji, H., Wang, Y., Saxena, A., 2005. Dynamic Case-Based Reasoning in Fault Diagnosis and Prognosis. IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Reno, Nevada, USA. Brooks, A., 1986. A robust control system for a mobile robot. IEEE J. Robotics Autom. 2 (1), 14–23. Byington, C., Kalgren, P., Johns, R., Beers, R., 2003. Embedded diagnostic/prognostic reasoning and information continuity for improved avionics maintenance. IEEE Systems Readiness Technology Conference, Autotescon, 320–329. Campos, J., 2009. Development in the application of ICT in condition monitoring and maintenance. Comput. Ind., 60. Chittaro, L., Ranon, R., 1999. Diagnosis of multiple faults with flow-based functional models: the functional diagnosis with efforts and flows approach. Reliab. Eng. Syst. Saf. 64, 137–150. Clarhaut, J., Hayat, S., Tahiri, H., 2011. Integration of communication systems for a railroad smart wagon, In Proceedings of IEEE Logistiqua Conference, ISBN: 978-1-4577-0322-5, 19–24. Cordier, M.O., Dague, P., Dumas, M., Le´vy, F., Montmain, J., 2000. A comparative analysis of AI and control theory approaches to model-based diagnosis. In Proceedings of the Fourteenth European Conference on Artificial Intelligence (ECAI 2000), Berlin, Germany, 136–140. De Kleer, J., Williams, B.C., 1987. Diagnosing multiple faults. Artif. Intell. 32 (1), 97–130. Dievart, M., Charbonneau, P., Desforges, X., 2010. An embedded distributed tool for transportation systems health assessment. Conference on Embedded Real Time Software and Systems, ERTSS. DIN EN 13306, 2010. Maintenance terminology.European Committee for Standardization. Discenzo, F.M., Nickerson, W., Mitchell, C.E., Keller, K.J., 2002. Open Systems Architecture Enables Health Management for Next Generation System Monitoring and Maintenance. Proceedings conference of the Machinery Failure Prevention Technology, MFPT. Ferreiro, S., Arnaiz, A., Sierra, B., Irigoien, I., 2012. Application of Bayesian networks in prognostics for a new Integrated Vehicle Health Management concept. Expert Syst. Appl.: Int. J. Arch. 39 (7), 6402–6418. Feldman, A., 2010. Approximation Algorithms for Model-Based Diagnosis.Ph.D. Thesis, Delft University of Technology, ISBN 978-90-9025023-6. Ferrari, R., 2009. Distributed Fault Detection and Isolation of Large-scale Nonlinear Systems: an Adaptive Approximation Approach. Ph.D. Thesis, University of Trieste. Ganyun, L.V., Haozhong, C., Haibao, Z., Lixin, D., 2005. Fault diagnosis of power transformer based on multi-layer SVM classifier. Electr. Power Syst. Res. 74, 1–7. ISO 13374-1, 2003. Condition monitoring and diagnostics of machines, Part 1: General Guidelines, Data Processing, Communication and Presentation. International Organization for Standardization. IAEA, 2007. Implementation Strategies and Tools for Condition Based Maintenance at Nuclear Power Plants. ISBN 92-0-103907-7. Jarvis, D.H., Jarvis, J.H., 2003. Holonic diagnosis for an automotive assembly line. In: Deen, S.M. (Ed.), Agent-Based Manufacturing: Advances in the Holonic Approach. Springer, pp. 193–206. Jianjun, C., Peilin, Z., Guoquan, R., Jianping, F., 2007. Decentralized and overall condition monitoring system for large-scale mobile and complex equipment. J. Syst. Eng. Electron. 18 (4), 758–763. Katipamula, S., Brambley, M.R., 2005. Methods for fault detection, diagnostics, and prognostics for building systems—a review. Part I. Int. J. HVAC&R Res. 11, 1. Khol, J., Bauer, A., 2010. Role-Based Diagnosis for Distributed Vehicle Functions. In 21st International Workshop on the Principles of Diagnosis (DX), Portland USA.

240

A. Le Mortellec et al. / Engineering Applications of Artificial Intelligence 26 (2013) 227–240

Koestler, A., 1967. The Ghost in the Machine. Hutchinson publishing Group, London. Koutsoukos, X., Biswas, G., Mylaraswamy, D.A., Hadden, G.D., Mack, D., Hamilton, D., 2010. Benchmarking the Vehicle Integrated Prognostic Reasoner. Annual Conference of the Prognostics and Health Management Society. Kurien, J., Koutsoukos, X., Zaho, F., 2002. Distributed Diagnosis of Networked, Embedded Systems. Proceedings of the 13st International Workshop on Principles of Diagnosis (DX-2002). Semmering, Austria, 179–188. Lefebvre, A., Simeu-Abazi, Z., Derain, J.P., Glade, M., 2007. Diagnostic of the avionic equipment based on dynamic fault tree. IFAC-CEA, Monterrey, Mexico. Li, J., Shen, S., 2007. Research on the algorithm of avionic device fault diagnosis based on fuzzy expert system. Chin. J. Aeronaut. 20 (2007), 223–229. McFarlane, D., Marett, B., Elsley, G., Jarvis, D., 1995. Application of holonic methodologies to problem diagnosis in a steel rod mill. IEEE Int. Conf. Syst., Man and Cybern., Int. Syst. 2lst Century, Vancouver 1, 940–945. Monnin, M., Voisin, A., Le´ger, J.B., Iung, B., 2011. Fleet-wide health management architecture. Annual Conference of the Prognostics and Health Management Society, Montreal, Canada. Nasri,O., Shraim, H., Dague, P., Heron, O., Cartron, M., 2012. Modeling and deployment of model-based decentralized embedded diagnosis inside vehicles: application to smart distance keeping function. Int. J. of Veh. Technol. vol. 2012, Article ID 103530, 14 pages. Ng, Y.S., Srinivasan, R., 2010. Multi-agent based collaborative fault detection and identification in chemical processes. Eng. Appl. Artif. Intell. 23 (6), 934–949. Orbita-BT, 2006. Data monitoring cuts rail running costs. Prof. Eng. J. 19 (20), 50–53. Pencole´, Y., Cordier, M.O., 2005. A formal framework for the decentralised diagnosis of large scale discrete event systems and its application to telecommunication networks. Artif. Intell. 164, 1–2. Provan, G., 2002. A model-based diagnosis framework for distributed system. In Proceedings of 13th International Workshop on Principles of Diagnosis, 16–24. Qiu, W., Kumar, R., 2004. Decentralized Failure Diagnosis of Discrete Event Systems, in Proc. of 7th Int. Workshop on Discrete Event Systems, Reims, France, Sept. 22–24. Rao, M., Yang, H., Yang, H., 1998. Integrated distributed intelligent system architecture for incidents monitoring and diagnosis. Comput. Ind. 37, 143–151. Ribeiro, L., Barata, J., 2011. Re-thinking diagnosis for future automation systems: an analysis of current diagnostic practices and their applicability in emerging IT based production paradigms. Comput. Ind. 62 (7), 639–659.

Roychoudhury, I., Biswas, G., Koutsoukos, X., 2009. Designing distributed diagnosers for complex continuous systems. IEEE Trans. Autom. Sci. Eng. 6 (2), 277–290. Sallez, Y., Berger, T., Deneux, D., Trentesaux, D., 2010. The life cycle of active and intelligent products: the augmentation concept. Int. J. Comput. Integrated Manuf. 23 (10), 905–924. Simeu-Abazi, Z., Di Mascolo, M., Notek., M., 2010. Fault diagnosis for discrete event systems: modelling and verification. Reliab. Syst. Saf. 95 (4), 369–378. Trentesaux, D., 2009. Distributed control of production systems. Eng. Appl. Artif. Intell. 22 (7), 971–978. Umiliacchi, P., Lane, D., Romano, F., 2011. Predictive maintenance of railway subsystems using an Ontology based modelling approach. The 9th World Conference on Railway Research, WCRR 2011, Lille, France. Venkatasubramanian, V., Rengaswamy, R., Kavuri, S.N., 2003a. A review of process fault detection and diagnosis part II: qualitative models and search strategies. Comput. Chem. Eng. 27, 313–326. Venkatasubramanian, V., Rengaswamy, R., Yin, K., Kavuri, S.N., 2003b. A review of process fault detection and diagnosis part III: Process history based methods. Comput. Chem. Eng. 27, 327–346. Vinatoru, M., 2008. Fault Detection in Hydropower Plant using Holonic Structure Concept, 12th WSEAS International Conference on Systems, Heraklion, Greece, 234–239. Vong, C., Wong, P., Ip, W., 2011. Case-based expert system using wavelet packet transform and kernel-based feature manipulation for engine ignition system diagnosis. Eng. Appl. Artif. Intell. 20, 1281–1294. Widodo, A., Yang, B.S., 2007. Support vector machine in machine condition monitoring and fault diagnosis. Mech. Syst. Signal Proc. 21 (6), 2560–2574. Wooldridge, M., 2002. An introduction to MultiAgents systems. John Wiley & Sons. Willaeys, D., Asse, A., 2011. Device and method for a system analysis and diagnosis. United States Patent Application Publication. Pub. no. US008037005B2, Oct. 2011. Prior Pub. no. US 2008/0086288 A1, Apr. 2008. Yang, H., Mathew, J., Ma, L., 2002. Intelligent diagnosis of rotating machinery faults—a review. in Proceedings of the 3rd Asia-Pacific Conference on System Integrity and Maintenance, Queensland, Australia, 385–392. Zaki, O., Brown, K., Fletcher, J.E., Lane, D., 2007. Detecting faults in heterogeneous and dynamic systems using DSP and an agent-based architecture. Eng. App. Artif. Intell. 20 (8), 1112–1124.