A global modular framework for automotive diagnosis

Advanced Engineering Informatics 26 (2012) 131–144

Contents lists available at SciVerse ScienceDirect

Advanced Engineering Informatics journal homepage: www.elsevier.com/locate/aei

A global modular framework for automotive diagnosis A. Azarian ⇑, A. Siadat LCFC-ENSAM, 4 Rue Augustin Fresnel, 57078 Metz, France

a r t i c l e

i n f o

Article history: Received 27 June 2010 Accepted 3 October 2011 Available online 2 November 2011 Keywords: Diagnosis Heuristic diagnosis Model-based diagnosis Diagnostic algorithm Causal dependency graph Knowledge management

a b s t r a c t The automotive after-sales dealers lack solutions for accurate, comprehensive and efficient fault localization. However, such services in the after-sales networks are crucial to the brand value of automotive manufacturers and for client satisfaction. In this paper, a new approach for the off-board diagnosis is presented, with significant improvements compared to the current technologies. A more robust approach that allows, per the additions of functional modules, to enhance traditional computer aided diagnostic systems towards knowledge based systems that emphasize the whole life cycle of the vehicle. Once the design of a new vehicle has begun, information like the dependencies between the components could be re-used for the models dedicated to the diagnosis task. The massive use of electronics dramatically increases the amount of data to manage for the testing of ECU (Electronic Control Unit) functionalities. The complexity of the sub-systems leads to breakdowns that need qualitative symptom description for fault localization. Finally, a feedback engine completes the expensive models for the diagnosis and returns critical dysfunctions to the design department. In this paper, we present our research on a global modular framework for the diagnosis. It encompasses the needs and requirements of automotive manufacturers. The results are presented with data obtained from low, middle and luxury class vehicles. They demonstrate the performance in real field conditions of our different modules. They are based on the interpretation of observations, the fault localization and isolation and the evaluation of feedbacks for model auto-completion. These experiments show the potential of our proposed approach for the automotive off-board diagnosis task. Crown Copyright Ó 2011 Published by Elsevier Ltd. All rights reserved.

1. Introduction It has been repeatedly reported that computer aided diagnostic tool’s accuracy and efficiency depend on the quality of the models used [1]. Current techniques are based on expert systems or in combination with CBR (Case Base Reasoning) engines [2]. Model based diagnostics are very accurate but time-consuming and labor-demanding, and therefore too expensive to be comprehensively applied in workshops [3,4]. The economic pressure leads manufacturers to a high degree of innovation with a massive use of electronics to distinguish their products with the consequences that new technologies such as X-by-Wire, Electric and hybrid engines, Car to car communication, wireless sensors [5] are used. Moreover the clients have the possibility to customize their cars with optional equipments. As a result the dependence between components and the management of vehicle variants makes it difficult for traditional expert systems or CBR engines to diagnose broken components in a reasonable amount of time.

⇑ Corresponding author. Tel.: +49 (0) 1701124043. E-mail addresses: [email protected], (A. Azarian), [email protected] (A. Siadat).

[email protected]

This section is divided into three parts beginning with a summary of existing research, strategies and tools, followed by a presentation of the problem statement and objectives of our contribution. Finally, a brief description of our industrial platform used for the experiments is given in the form of an overview. 1.1. Existing research Many research initiatives are investigating the fields of diagnosis allowing a progression in efficiency, in particular: model based diagnosis which improves the precision of fault localization [6], qualitative reasoning which handles incomplete models [7], and distributed diagnostic agents which process parallel information [8]. 1.1.1. Knowledge based systems Several knowledge based systems have been proposed for the diagnosis task [9–11]. Their general strategies rely on (qualitative or quantitative weighted) inference rules or on a combination of expert rules with a case base. One common problem encountered with those strategies is that the diagnostic knowledge is encoded in a rule [7] which is vehicle dependent. If we consider the symptoms: ‘‘the headlights do not work’’, it could be linked to the engine

1474-0346/$ - see front matter Crown Copyright Ó 2011 Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.aei.2011.10.001

132

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

battery (for a low class car with only one battery) or with the comfort equipment battery (for a high class car with 2 power supplies). The second test for the confirmation or rejection of the hypothesis of a faulty battery will be to start the engine in the first case or any entertainment equipment in the second case. The construction of the diagnostic tree (hypothesis and tests) varies with the car configuration. The MBD (model based diagnosis) technology consists in comparing the actual behavior of a system, as it is observed, with the predicted behavior of the system given by a corresponding model (see Fig. 1). A discrepancy between both states (observed and predicted) is a clear indication that a failure is present in the system. But the reasoning with an MBD engine would allow to identify which test would have to be performed to confirm a hypothesis. With our previous example, if in both cases the power supplies and their connections are modeled, the diagnostic engine can immediately determine the actuators (or equipment) to test (the motor engine or any entertainment component) and provide an explanation of why this test is performed (based on the construction knowledge, or in the case of our example, based on the wiring diagram of the power supply). Case based reasoning engines have also encountered a great success in the industry. They are adequate to reason on global diagnostics in automotive cases [12], such as for instance a faulty ECU engine management (which initiates around 60 fault codes in ECUs of a modern car [13]). But for hypothesis refining, it is difficult to find a global similarity metric or to ignore certain attributes in the case description after a global test. Despite the quick response time of CBR and the widespread physical domain covered, it is not adequate with the requirement of the automotive industry which needs the models of new vehicle to be operational in the workshops before commercialization starts, meaning the possibility to perform a diagnosis in the workshops without any ROE (return of experience about the encountered failures of components). Moreover, CBR does not provide any explanation about the delivered results, and does not handle the configuration problems of modern vehicles. Murphey et al. [8] proposed a distributed diagnosis agent structure with an engine reasoning from global diagnostics. The state of the system composed by n diagnostic agents can consequently be described by a signature vector. The discrimination comes from the comparison of the current signature with the one from known problems. The interesting aspect of this approach is that the distance between the different states (or coordinates of the signature) can be computed considering the constraints of the states as ‘‘undefined’’ or ‘‘not at disposal’’. This aspect is very valuable in automotive problems, where some electronics components often cannot return a signal, or due to a missing value in the database

The headlights do not work

the signal returned cannot be interpreted as a known state. This is a weakness of CBR engines. A statistical analysis of the sales of BMW [14] reveals that less than 1% of cars are identical, that implies a real challenge in the editing and updating of the vehicle model in diagnosis tools. This is particularly true for MBD techniques which need high postdevelopment cost of models for each different vehicle variant. The next section outlines the evaluation comparison of the strengths and weaknesses of current technologies. 1.1.2. Computer aided MBD systems To facilitate the assessments and comparison of diagnosis technologies, the NASA has developed the Advanced Diagnostics and Prognostics testbed called ADAPT [14]. It is a platform which allows to benchmark different tools and technologies under equable conditions in order to measure their performance. The diagnostics competition was defined at the 20th International Workshop on Principles of Diagnosis. Details about the general set-up are given in [15,16]. In the last decade, model-based technology in diagnosis has matured to the point where it was transferred from academic research into real applications. It provides an alternative to more traditional techniques based on experience, such as rule based reasoning systems or case-based reasoning systems. Early model-based diagnosis tools include MDS [17], RAZ’R [18], LYDIA [19], DSI Express [20] or RODON [21,28] or the module DIAGSYS in SIMFIA [22]. In [23], diagnosis is performed using RAZ’R [24] a consistencybased diagnosis engine. For the fault localization the software compares the observations with predictions based on the model behavior. Then a list of potential faulty candidates is generated and tested. This diagnosis relies on finding the assignments of modes (correct or faulty) to a set of component which is consistent with the system description and the observations. Consistency-based diagnoses can be generated by some sort of best-first search according to criteria such as maximal probability, minimal cardinality of the set of faulty components, and other orders on faults models. The tool provides at least an 88% compliance of core diagnosis results in this study [23]. The module DIAGSYS is also exploiting a consistency based approach with the help of causal dependencies. The module relies on a constraint propagator and carries out a diagnosis with a minimal cardinality in order to provide the least possible number of suspects. RODON’s model based diagnostic (MBD) engine is based on the notion of conflicts (or contradiction between observed and simulated behavior) known as the GDE (General Diagnostic Engine) already described in 1987 [25]. de Kleer and Williams [26] reports the synthetic track of the diagnostic competition with RODON employed for ADAPT (tier

ECU XAND: Fault FC3_G80

Repair action User

System

Mechanic Fault codes

Qualitative symptom description Observations ECU Vref

Diagnosis

D r ai n

FB

S our ce

C o mp

S ht dw n

Predicted behaviour

Re s e t

I-s e n s e

Model

Diagnostic reasoning engine Fig. 1. Illustration of MBD principle applied in automotive industry.

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

1 and tier 2, details given in [27]). In both cases, the isolation accuracy is close to 1 and the computing time under 25 s with reasonable peaks of memory usage. But these standardized metrics from [16] used for the evaluation do not take into account the implementation effort of the models. This effort is essential for automotive manufacturers who are looking for an economical objective commissioning, such as ratio of good answers divided by implementation cost of models. Moreover, the trend is to diminish all the authoring cost in automotive diagnostic tools [12].

1.1.3. Specificity and problem statement of the automotive diagnosis task As a conclusion, the model-based solution is a highly challenging technology in several aspects, all the more so if the process to diagnose contains high uncertainties, if the available observations are reliable and accurate, and if the available data for the completion of the models have a sufficient degree of quality. In order to benefit from the high accuracy of the MBD technology, we decided to develop a global framework for the diagnosis that emphasizes the most important aspect of automotive diagnosis requirement as stated by car manufacturers which are the precision and the performance. As depicted in Fig. 1, the first exchange of information about the system state that occurs is the qualitative symptom description, between the user and the technician in the workshop. Taking into account this description, in order to maximize the information input for the initialization of the diagnosis session, may reduce the research area of the diagnostic engine. Following this line would require the diagnostic tool to be able to handle natural language and at least map these sentences on a pre-compiled database of observable symptoms. The second information exchange happens between the car and the diagnosis station. In this step, the car communication allows the acquisition of the ECU’s fault codes which are accurate observations and therefore crucial information for the diagnosis task. The problem is the interpretation by the diagnostic engine of these fault codes. They vary for each vehicle and according to the market report of automotive electronic in [26] the trend to use more and more ECU is on the rise. Therefore we focused on the development of a module which allows to scan automatically the description files of the ECUs called ODX (Open Diagnostic eXchange), and import them as completely functionalized data into the database. Both of these problems address the specific point of the availability, accuracy and reliability of the observations as depicted in Fig. 2. Based on our experimental platform called SIDIS Enterprise developed by Siemens AG, which relies on an inference engine with all the drawbacks described in Section 1.1.1, we designed a module in order to improve the diagnostic algorithm to benefit from the accuracy offered by the MDB technology. At least, as outlined by a report about the European automotive aftermarket in [3], the ratio of authorized repairer against independent repairer tends to diminish. Nowadays the workshops are partially integrated in the workflow of the after-sales department of car manufacturers. Therefore, to benefit from the scale of this valuable source of information, we developed a feedback evaluation module. It inspects the protocols of the performed diagnosis sessions and auto-completes the models for the diagnosis in order to be cost-effective for the database edition process. The objective is to improve the current diagnostic engine of SIDIS Enterprise into a combined reasoning strategy with expert and a model based engine re-enforced by ROE knowledge. The resulting framework for the diagnosis will benefit from all source of available knowledge in a car manufacturer in order to cover the whole life cycle of a vehicle (design: construction knowledge or model; rules: diagnostic experts; ROE: feedback evaluation).

133

1.2. Contribution Over the last years, the major field of innovation in the automotive industry has been in the field of electronics, which has had the following impacts: (1) The different subsystems are more and more interconnected so that they can share information. The different subsystems must be entered in the diagnosis knowledge database by experts, which increases the maintenance cost of diagnosis tools for car manufacturers. (2) More subsystems mean more (suspicious) components and consequently a higher risk of multiple faults and a loss of time to perform test on functioning components. Our previous work in [12] had allowed on the one hand the taking into account of the qualitative symptom description by a search engine in natural language. On the other hand, we developed a module to administrate automatically the ECU description files, which reduces the maintenance cost of diagnosis tools. The model based diagnostic strategies as described in [26] may handle the listed difficulties with great accuracy, but a qualitative analysis of the cost of the models will reveal the cost to be significantly higher than that of an expert system. Moreover, the model based technology only takes construction knowledge into account, not expert or experience knowledge. The first contribution of this paper is an improved diagnosis algorithm that bridges traditional expert knowledge with models and optimizes the models through feedbacks in order to provide an incremental innovation which allows the migration toward the MBD technology. A radical innovation in products for complete automotive after-sales networks may induce too much change in the structures of car manufacturers. The success of a combined approach for fault detection and isolation was already used in [29]. The objective of our contribution is to modify the diagnosis engine depicted in Fig. 1 into a meta-engine combining the rule based diagnostic with model based diagnostic. Moreover, it will auto-complete or optimize the models for the diagnosis by a feedback evaluation module as depicted in Fig. 3. It is shown that with the qualitative causal models as in DIAGSYS through multiple experiments the resulting diagnosis has the following strengths: - Accuracy: the causal relationships allow a better detection than traditional expert systems. - Response time: the reduced detection time has a direct impact on the answer time. The scale of an automotive after-sales network of a manufacturer allows around 10.000 diagnosis sessions a day in a country like Germany. Therefore each saved minute has a strong economic impact on the return on investment. - Compatibility: the algorithm could be programmed as generic as possible and switch between the different sources of knowledge. If for example the models have not been edited, it should run as a traditional expert system in order to allow the manufacturers to change their process and information source for the data edition (this constraint is strongly related to the technology adoption lifecycle [30]). - Ratio good answers divided by the maintenance cost: this metric is the most important factor for manufacturer because it reflects the return of investment of the acquisition of a complete automotive diagnosis solution. Therefore, our first objective will be to develop a new diagnostic engine that combines model based knowledge and expert knowledge [31,32]. Besides, as mentioned in Section 1.1.3, the

134

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

Symptoms Identification of symptoms

Client

Vehicle communication

Technician

Control interior lights

Natural language

Data abstraction and import ODX File 1

Job status OK

Client’s vehicle with faulty behavior

Diagnosis station

` AS Author

ODX File 2

ECU rough communication data Fig. 2. SIDIS Enterprise diagnostic station processing the evaluation of observation [27].

Expert knowledge

On Fig. 4, the knowledge structure related to the diagnosis task is represented. These data are organized in trees where child nodes inherit the properties and attributes of parent nodes. The trees relative to the diagnosis are:

Diagnosis Engine Model based knowledge

Return of experience

Fig. 3. A frame of unification for automotive diagnosis.

automotive manufacturer has a large network of after-sales service points (around 10.000 diagnosis session a day for a country like Germany). The second contribution of this paper is to take into account the valuable source of information of the feedbacks from performed diagnosis sessions. The long term vision of this new engine is to return failure information about a specific component directly to the design department of the manufacturer and integrate the whole value chain from R&D to after-sales in a same workflow. The originality of this solution consists in a diagnosis algorithm that allows to combine the different sources of knowledge and to migrate from a traditional expert system towards a more accurate model based diagnosis. 1.3. System overview SIDIS Enterprise is based on a client/server architecture which is composed of two parts: - The authoring system: which allows to edit simultaneously (on different clients) all the information in the database (diagnostic data, ECU data, service repair documentation, etc. . .) - The workshop system: which is used in the service points and enable the technician to acquire the fault codes, perform measurements, show repair documentation.

- The perceived symptom tree: (top left of Fig. 4) contains system and senses oriented failure descriptions as for example: ‘‘the engines vibrates during the ignition’’. A symptom node is characterized by a weight coefficient and by outgoing suspicion links which points toward diagnosis node. - The fault code tree: (bottom left of Fig. 4) contains system oriented symptoms as for example: ‘‘short cut to ground of electromagnetic valve #7’’. These informations are read from the ECU for the initialization of the diagnosis process. The fault code nodes contain the same characteristics as the perceived symptom. - The diagnosis object tree: (top right of Fig. 4) is a model of the car in which a node represents a component or subsystems. The hierarchy links between the diagnosis nodes are equivalent to a ‘‘is composed of’’ or ‘‘is part of’’ relationship. The diagnosis nodes are associated to tests which allow checking if the component is functioning or not. The other important attributes are the life time of the corresponding component, the level of the node in the diagnosis object tree and the number of incoming suspicion rules. - The test objects: (not represented on Fig. 4) are associated to diagnosis node. They allow to incriminate or discriminate components during the diagnosis process. They can for example involve car communication through the ECU or ask the technician to perform a tension measurement. They are characterized by: the test cost for an OK answer, the test cost for a NotOK answer and the coverage of the subcomponents. - The suspicious links: (bottom middle on Fig. 4) are basic expert rules coming from fault codes or perceived symptoms (source) and pointing at diagnostic objects (target). They are characterized by their rank if several links point out a same node.

135

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

Perceived Symptom tree

Diagnosis Object tree

Perceived symptom node object

Diagnosis node object

Symptom #ID:

152

DO #ID:

2052

Symptom Text:

Root symptom

DO Text:

Root Diagnosis

Symptom Weight:

GS

Associated Test by ID :

Suspicion Rules by ID :

[0]

Test cost:

Engine

Carriage

Carriage

Life time:

Engine

12000

Incoming suspicion rules by Ids : Node level : Cooling system

Ignition

Temperature

Breaking System

Losing water

Ignition

ABS

3012 [15,22]

[22,37,45] 0

Fuel supply

Ignition cycle

Injection cycle

Causal links Suspicion Links Too high

Too low

Fault Code tree

Suspicion Links

Causal links

Fault Code node object FC #ID

1054

FC Text:

Comfort Equipment

Root Fault Code

FC Weight:

GS

Suspicion Rules by ID :

[0]

Engine

Suspicion link object

Causal link object

Suspicion link #ID:

1045

Causal link #ID:

Source node #ID:

10500

Source node #ID:

9654

Target Node #ID:

12400

Target Node #ID:

13421

Rank:

2

Weight:

2985

GC

Electronic V2.1

Electromagnetic valve 1

Electromagnetic valve 7

Interuption

Short cut

Fig. 4. Diagnosis related knowledge structure and their properties in SIDIS Enterprise.

In the current context, authors who edit the knowledge structures cannot complete the whole information. Actually parameters like perceived symptom of fault code weight, life-time, test cost, test coverage are not used by car manufacturer. The next section investigates different possibilities to improve the diagnostic engine and presents our approach.

ðk þ 1ÞRSLn PMn ¼ GS Pk i¼1 RSLi where:

 GS: the weight of the symptom or fault code.  PM: the partial moment or partial suspicion degree induced by a rule.  K: the number of rules from the symptoms or fault code.  N: index of the considered suspicion rule.  RSL: rank of the considered suspicion rules.

2. Proposed diagnostic algorithm As shown in Fig. 3 the diagnosis algorithm is divided into three different parts: the heuristic knowledge reasoning, the model based and the feedback evaluation. These steps are described in the following three sub-sections. Section 3 presents experimental results demonstrating the performance of this approach compared to the current technology.

2.1. Expert knowledge In the current system, each time a symptom is observed on the treated case, the suspicion links (see Fig. 4) attached to that object will induce a partial suspicion moment to the targeted diagnostic object. This partial suspicion moment or PM depends on the number of suspicion links from the symptom or fault code called k for each symptom Si, given in (1):

ð1Þ

The rank of a suspicion rule is its order or priority against others if several rules starts from a same symptom (or fault code). Once the total suspicion moments denoted SM (sum of all the PM for a given diagnosis object), have been evaluated, the objects are ranked by a heuristic metric h given in (2)

h¼

SM x Tcx

ð2Þ

where:  SM: the total suspicious moment (or the sum of all partial moments) of a node.  Tc: the average test cost (between Ok and notOk answers) of the node.

136

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

This metric h represents the suspicion degree of the components. All the components are ranked according to this value and the seven components with the highest suspicion degree are listed into a test agenda. The technician can select one of the objects of the test agenda to incriminate or discriminate it by performing a test. Once the fault has been identified the technician proceeds with the repair and a feedback protocol of the diagnosis session is sent to the manufacturer. However, considering the amount of electronic subsystems, a simple expert system does not return adequate response in a reasonable time. Specially for critical errors like failure codes of the engine management ECUs which activates around 60 failures codes in different subparts as in the ABS or ESP (Anti-lock Braking System, Electronic Stability Program) and the semi-proportional suspicion degree attribution does not discriminate enough the different objects. The main advantage of a rule based system is that engineers can easily enter the data; standardized documents like FMEA (Failure Mode and Effects Analysis) are the main source of diagnostic knowledge in the automotive industry. But for larger systems declined in many variants the overall view of the model is lost and authors who edit the databases do not commit a weight coefficient to a rule with, as a critical consequence, the loss of discrimination power. As a result the candidates have more or less the same rank in the test agenda (as the test cost are usually set to their default value) leaving the decision of further investigation for the technician in the workshops.

2.2. A new diagnostic strategy 2.2.1. Previous work There are a lot of different diagnostic methods. Several are presented by Piechowiak [33] but hereunder only the most relevant to our case will be discussed. The first one is the Case-Base Reasoning method, detailed by Leake [34]. Another solution consists in the implementation of model based diagnostic with fault models. This approach has the advantage of being very accurate. However, the models are very expensive. The MBD principle issued from the work on the hitting set algorithm from [35] can help to improve the discrimination power of inference rules which correspond to the possible conflicts. The weight of each symptom will be proportional to the number of independent linked diagnostic objects (by a suspicion rule). A subset of this strategy is the implementation of causal networks which are qualitative models representing the physical dependencies between components. This method had the strong advantage of not radically changing the information model. A third method is the a–b search. The principle is to reduce the search space and focus only on the portion which contains relevant information while irrelevant nodes are deleted. The algorithm looks for the best result at the nth depth level and makes the move to go to the closest possible node. Transposed to the diagnosis, a metric, based on factors such as suspicion degree, failure rate, vehicle maintenance history, etc. needs to be implemented in order to rank the results at the nth depth. However, this solution needs high resources during the exploration of the graph. Another solution consists in data-driven methods like the use of conditional probabilities as illustrated by Piechowiak [33] or Schwall and Gerdes [36] with the help of Bayesian networks (BN) which could result in accurate fault localization if applied to the automotive domain. These probabilities could be evaluated by the analysis of the feedbacks of the diagnosis sessions (performed in the manufacturer’s after-sale networks). But these methods are not adequate for the automotive industry, because it is difficult to derive the data from FMEA documents for a new vehicle. An extension of this strategy in order to establish a diagnosis, will consist in the implementation of physical characteristics for each (mechanical or electrical) element. Information such as the following examples could be useful for suspecting more or less an object:

- The test cost which allow to rank the test object in an timeoptimized manner. - The test coverage because a test cannot cover all possible failure modes of a component. - The life-time because elements are built to last a certain amount of kilometers or working hours. - The weight of causal dependencies or the relative importance of a rule. - Unobserved causal dependencies or suspicion links. In all cases, the integration of new knowledge requires higher maintenance costs for the models and authors will require more sources of information, but the causality or influence relationships are generally defined in FMEA documents and may be a promising issue for the problem statement. 2.2.2. Selection of an approach In order to focus our work on a solution that requires the smallest amount of effort on the implementation side compared to aforementioned methods, we will retain the criteria already discussed in Section 1.2 (accuracy, response time, compatibility, ratio: correct answers by maintenance cost) and also the operational applicability of the approach. This means that improvements must be easily interpretable for users, in particular the authors who establish the models for the diagnosis. The candidate generation must also be interpretable for the technician, e.g. the activation of a causal link which leads to the suspicion of a particular component. A qualitative comparison of the methods is summarized in Table 1. Considering the first two methods, their main drawbacks are their lack of applicability due to the changes requested in the model edition or case description. But the integration of causal networks as well as the physical characteristics may be a promising solution when combined together. Furthermore the automatic learning of the physical attributes of the components and the causalities could be realized through BNs. Therefore, our approach will rely on a diagnostic module that integrates horizontally the heuristic diagnosis, the model based diagnosis with the implementation of causal networks and physical component’s attribute. As mentioned, the causal relationships represent physical dependencies between components therefore they will be implemented in the diagnostic object tree (see Section 1.1.3) as oriented weighted links which will be vehicle dependent. The physical characteristics of the components will be stored as attributes of the objects in the diagnostic tree. The last module will be composed of the evaluation of the ROE through BNs in order to complete a modular framework as depicted in Fig. 3. This module is implemented on a feedback evaluation server which receives the protocols of the diagnosis session from the after-sales service points. The data will be evaluated and discovered knowledge will be transmitted to the car manufacturer who can decide or not to update the knowledge base of the diagnosis tools. The next section details the formalization of the horizontal combinations of the diagnosis reasoning engines and the evaluation of the feedback. 2.2.3. Principle and formalization The proposed development consists in implementing two fundamental equations for the candidate generation. The first one deals with the evaluation of the causal moment whereas the second deals with the life time parameter, both are presented in the next sub-sections. The causal links have the same properties as suspicion links except that they point out a diagnosis node to another diagnosis node (see Fig. 4). The last sub-section details the specific evaluation of the ROE. Section 2.2.4 details the combination of the different approaches.

137

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144 Table 1 Comparative table of the suggested solutions.

CBR MBD with fault models Qualitative MBD (causal network) ab Search BN Physical characteristics

Accuracy

Response time

Compatibility

Ratio correct answers/maintenance cost

Operational applicability

High Very high High High Average Low

Very high Low Average Average High High

Low Very low High Average High Average

Very high Low High Average Average Low

Low Very Low High Low Average Average

(a) The causal factor Because the system should emphasize the elements which are the most suspected, i.e. whose causes play an important role, the idea (as described in Section 1.2) is to use the information theory to quantify the importance of causal ancestors of an already suspected diagnosis node. In particular, the information entropy based on [37] and generalized by [38]. Adapted to our diagnosis reasoning the causal suspicious moment denoted SMa is defined by (3):

 P  a   a  1 LevelðnodeÞ 1 ei GC i  SM a ¼  log2 o þ log2 o þ Pi  1a DepthðtreeÞ 1  a k GC k ð3Þ where:  SMa: the causal suspicion moment of a node.  Level(node): the depth in the diagnostic object tree of the considered node.  Depth(tree): the depth of the diagnostic object tree.  a the order of the entropy.  GCi: the weight of the causal link indexed by i.  ei: a boolean value equal to 1 if the causal successors of the node is already suspected and 0 else.  o: an offset equal to 1. The first part of Eq. (3) corresponds to the precision and contributes to the information gain. The higher the position of a node, the less information it contains and the lower its entropy in the tree. This factor benefits to precise fault localization compared to candidates, which have a lower level (or objects that are nearer to the root of the tree) which are essentially the LRU (Least Replaceable Units). The denominator of the ratio contributes to a normalization resulting in the ratio to be comprised in the interval [0, 1]. The second part of Eq. (3) deals with the impact of the previous causal ancestors (like backward chaining), a cause with three active heavy weighted previous causes has definitely not the same importance as one with one light weighted previous cause. Here, GCi represents the weight of the previous causes and ei equals 1 if the previous cause is active, 0 otherwise. If no previous cause is active, then the expression equals 0 (thanks to the offset). (b) The lifetime factor As discussed the lifetime parameter of one component plays an important role in the diagnosis. The usual curve of the lifetime of an equipment is known as the bathtub curve where three different periods can be recognized: early period (premature defects), maturity period, wear lifespan. At the end of the maturity period, the probability of a breakdown increases with the number of working hours of the considered equipment (see Fig. 5). The older the equipment is, the higher the probability of a failure. This type of behavior is characteristic of systems that underlie wear or other progressive deterioration (e.g. oxidation) causing the failure rate to grow.

The last part of the curve is significant for electronic equipments in the automotive industry, the modeling of this part of the curve is done with a threshold function as described in (4) called Heaviside function.

PMLT ¼ HeavisideðLTðvehicleÞ  LTðnodeÞÞ

ð4Þ

where:  PMLT: the partial suspicion moment of a diagnostic object node induced by the life time behavior.  LT(vehicle): the km reading of the current vehicle being diagnosed.  LT(node): the life time in km of the corresponding diagnosis node.  Heaviside(x): The Heaviside function (or unit step function) is equal to 0 for x < 0 and equal to 1 for x P 0. The Eq. (4) triggers a partial suspicion moment if the current km reading of the vehicle is higher than the given lifetime of the node currently being inspected. Otherwise the partial suspicion of the lifetime is equal to zero as depicted in Fig. 5. The Heaviside approximation acts as a threshold and fires a suspicion only if the prior life-time of a component is exceeded augmenting the discrimination power between candidates. This approximation is therefore more suited due to the fact that it has an upper bound and better imitates the wear out behavior of the lifetime periods of electronic equipments. (C) Evaluation of the return of experience Despite the existence of hundreds of new machine learning algorithms (a study of a review can be found in [39]) from a machine learning academic researcher’s perspective, the current described contribution is only preliminary and does only address the question ‘‘how to search through the feedback protocols (hypothesis space) to converge to an optimal hypothesis f consistent with the diagnosis session and also performing well over other unseen observations in order to improve the diagnosis?’’. In order to answer this question, we focused on the learning of the following parameters (see Section 2.2.1):  Optimization of the weights coefficients of suspicion rules, symptoms and of causal links.  Dependencies, suspicion rules between components.  The life-time of the components.  The test cost.  The test coverage. In our research, we also considered learning the failure rate of equipments in order to approximate the bathtub curve, but benchmarks on our datasets of feedback protocols only allow to learn and optimize the above mentioned parameters with a high degree of certainty. For the learning of parameters, the Bayesian network [40] considers that all possible suspicion rules are present in the model (or

138

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144 Failure rate

Bathtub curve

PMLT

Early « Infant Mortality » failures

Constant random failures

Wear out failures

Time (km) Lifetime

Fig. 5. Model of the bathtub curve behavior for the candidate generation.

in the diagnostic object tree, symptom and failure code tree; see Section 1.3) and also all causal links with a weight of 0. The protocol sent to the central server will then evaluate the probability of acuteness of the link or of the causal relationship. For example, for a given symptom or fault code a suspicion rule is taken. Based on all feedback protocols and the results of the associated test to the incriminated diagnostic object by a rule, a new rank can be calculated based on (5):

Pk R ¼ Pk

i¼1 P NotOKi P þ ki¼1 P OKi

ð5Þ

i¼1 P NotOKi

where:  R: new rank of the considered suspicion rule.  K: total number of feedbacks containing the considered suspicion rule.  PnotOK: number (1 or 0) for, NotOK’ answers for the tests associated to the component targeted by the rule.  POK: number (1 or 0) for, OK’ answers for the tests associated to the component targeted by the rule. In order to illustrate the Bayesian network, the strategy will be applied on an abstract example based on the knowledge structure of SIDIS Enterprise. In Fig. 6 a diagnostic object tree is depicted

S2

S1

with 3 symptoms, which represent the beginning of a guided fault diagnosis session. The nodes n2 and n5 answer ‘NotOK’, but nodes n3 and n4 answers ‘OK’ despite the fact that they are candidates. The guided fault finding procedure of this example begins with the node n4, then n3, n2 and n5, which are ranked first in each corresponding test agenda. In this example, it is supposed that only default values are affected to the different weight parameters and no causal links have been modeled. Considering that the previous example delivers a feedback protocol which is analyzed, in particular the Symptom S2 and the rules n°1 pointing to node n2 and rule n°2 pointing to node n3 with both a default rank of 1. The details of the calculation of the new rank for both rules are given by (6):

Pk

i¼1 P NotOKi ¼ Pk ½k¼1 i¼1 P NotOKi þ i¼1 P OKi

R1 ¼ Pk R2 ¼

PNotOK fn2 g 1 ¼1 PNotOK fn2 g þ POK fn2 g 1 þ 0

PNotOK fn3 g 0 ¼ ¼0 PNotOK fn3 g þ POK fn3 g 1

It must be outlined that if the model in the example is designed only to diagnose this one case, then the second rule has a rank equal to 0, because it is considered as useless. This property can be very useful in case of models, which are too complicated to maintain. The statistical analysis can provide to clear the model of all useless rules which do not help the diagnosis engine.

S3

n1

n2

n5

n6

n3

n7

ð6Þ

n4

n8

Fig. 6. Example for Bayesian networks.

n9

n10

n11

139

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

The other parameters: weight parameters, dependencies, suspicion links and test coverage learning are based on the same principle. The weights of the symptoms and fault codes can be adjusted with the same formula, except that the classifier ‘NotOK’ and ‘OK’ are replaced by ‘acute’ and ‘in-acute’. But in this case the algorithm has to identify the ‘acuteness’ of a link which needs to interpret the whole diagnosis path from the initialization of the session by the activation of fault codes and symptoms toward the fault localization (as mentioned in the previous example). This ‘acuteness’ classifier of a link can be directly evaluated in the workshop system at the end of the diagnosis session, because the knowledge structures are already filtered to match the current vehicle. The implementation effort of such a method is very low because only digital object identifiers and object definitions are needed to compute the posterior probability If the feedback evaluation server has to evaluate the acuteness of a link for each protocol, it must re-simulate the entire diagnosis path based on the considered vehicle. This simulation would require 200 Mb of memory for each diagnosis protocol. Nevertheless, Section 3.3 outlines the encountered problem of this unsupervised learning method and proposes some threshold values based on our experience. The second aspect of the feedback evaluation is to calibrate the lifetime of the components and the test costs which allow to re-enforce the discrimination between candidates and to optimize the ranking in the test agenda. A solution for the calibration of these values consists in the use of decision trees. Decision tree maps observations about an item to conclusions about the item’s target value. Basically, the goal of a decision tree is to create a model that predicts the value of a variable based on several input variables. If for a diagnosis session the tested components and the km reading of the vehicle is sent to a central server very simple mathematical operators on the occurrence permit to build the decision tree as in Fig. 7. These collected frequencies permit to construct the hypothesis table (see Table 2) with the implicit hypothesis that if the km reading of the vehicle is higher than the supposed life time, then the test should reply ‘NotOK’. One way to control the quality of the hypothesis is to compute an indicator called Giny impurity given in (7). The value of this indicator reaches zero when all cases in a node reach a single label.

IG ¼ 1

m X

fi2

ð7Þ

i¼1

where:  m: the number of values a variable can take (here m = 2 for true or false).  fi: the fraction of item labeled with the value i in the dataset. In the dataset given in example it is possible to compute the Giny impurity for a hypothesis that the component had a life time

True

False

Lifetime 6 16.000 km Lifetime > 16.000 km

4 1

1 3

of 16,000 km. The number of occurrences where the hypothesis is true (means the km reading is superior or equal to 16,000 km and the test of the component answers ‘notOK’ AND the number of occurrence where the km reading is inferior to 16,000 km with a component’s test answer of ‘OK’) is equal to 3 and 6 when the hypothesis is false. The value is 1–(3/9)2–(6/9)2 equal to 0.44 which means that the data are quasi equitably categorized in both labels. To confirm an hypothesis, the posterior probability for the hypothesis to be true must be at least equal to 90% and the Giny impurity lower than 0.2. The problem of the decision tree is to construct an optimal decision tree with the best hypothesis and the best classifiers at the root of the tree. This problem is known to be NP-Complete (Non-deterministic Polynomial time). But despite this difficulty, a very simple dichotomy procedure can be used here. As a starting point the average value of the km reading of the dataset can be taken and 3 different leaves of the tree can be computed: average value minus standard deviation, average value, average value plus standard deviation. If no leaf is adequate with the previous criteria, then a dichotomy is made between both leaves that deliver the best percentage for the hypothesis equal to true (or if the scores are increasing then a new leaf is created with the hypothesis average value plus 2 times the standard deviation or vice versa). This minimizes the finding of an optimal tree with an algorithm of a complexity of o(nlog(n)) with n the number of times the dichotomy is applied. The same techniques are applied for the test cost whereby 2 values are learned: the first one for a test reply ‘OK’, the second one for a test reply ‘notOK’. This distinction has to be made, because the time varies depending on test results which are logical, because the test logic is to investigate deeper and deeper all possible failure modes. Moreover, the dichotomy procedure doesn’t need to be applied for the learning of the test cost, because the test costs do not vary as widely as the km reading of a vehicle. Nevertheless, to protocol the durability of the tests a timer needs to be started at each execution of the test and paused for example if the test is waiting for a user action (the user can for example be currently working on another vehicle). 2.2.4. Summary (a) Combination of all factors For the alignment of the suspicion induced by the expert knowledge, see (1) in Section 2.1, for the one induced by the causal network (Section 2.2.3.a) the calculation of the heuristic moment denoted SMh was computed according to (8).

ð8Þ

where:

Km reading

Component broken: 4 Component OK : 1

Hypothesis

 P 0 a    1 e GSi   SMh ¼  log2 o þ Pi i  1a k GSk

Component Broken: 5 Component OK : 4

KM≤16.000km

Table 2 Occurrence table.

KM>16.000km

Component broken: 1 Component OK : 3

Fig. 7. Example of a decision tree for the lifetime parameter.

 PMnode: the partial suspicion moment of a diagnostic object node.  a the order of the entropy.  e0 i: equal to 1 if the symptom indexed by i is active, and 0 else.  GSi: is the weight of the symptom indexed by i.  o: an offset equal to 1.

140

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

The two moments (SMh and SMa) obtained must now be put together with the lifetime factor in order to get a final moment called SMf for each object. The horizontal combination of the methods allows a weighted sum of these factors as in (9).

SM f ¼ bSM h þ ð1  bÞSM a þ cLT PM LT

ð9Þ

where: SMh: the heuristic suspicion moment obtained by (8). SMa: the causal suspicion moment obtained by (3). b: a real number comprised in the interval [0; 1]. PMLT: the partial suspicion moment induced by the lifetime of a diagnostic object given in (4).  cLT: the relative influence of the life time suspicion moment.

   

Other formulas can be proposed for the combination, but this one had the advantage that the parameter b could allow to switch or to balance relatively the expert and construction knowledge depending on the type of information available. The same applies for the coefficient cLT which can be equal to 0 if the information about the lifetime is not available. The b parameter allows, in accordance with our long term-objective, to migrate from an expert system towards an MBD engine. (b) Synthesis The proposed approach combines three different sources of diagnosis knowledge: first the expert knowledge via the suspicion rules, then the model based knowledge via the causal links, and finally the model optimization through probabilistic research. In the basic diagnosis algorithm the discrimination relies only on expert rules from fault codes or symptoms pointing towards possible faulty components such as, for example, ‘‘the ABS indicator is always on ? breaking pedal sensor’’. But this expert rule is not enough to build up fault hypothesis; with the learned values of life-time of the components through ROE, the discrimination can be slightly improved. The main source of accuracy for the diagnosis lies in the causal relationships between the components. Then, if the test of the breaking pedal sensors returns an OK, the diagnosis engine can explore the possible involved components like the cockpit panel-ECU or the wheel speed sensors, the engine management ECU, etc. If the models for the diagnosis contain only the simple expert rule, the technician will check the breaking pedal sensors, then he/she will test related components to the ABS system based on his/her knowledge until he/she finds the fault. The feedback of this diagnosis session will contain vital information because the component tested after the breaking pedal sensors may have causal connections to each other. For example, if this case appears frequently and if the cockpit panel-ECU is systematically tested and in the ‘‘not OK’’ state after the test of the breaking pedal sensors, then a causal relationship between both components will be learned. The causal links increase powerfully the discrimination capacities of the candidates depending on their weight. The interpretation behind a link like A ? B is ‘‘A may cause B’’, associated with correct weight coefficients corresponding to probabilities obtained through frequency analysis by the Bayesian network. These links improve the fault localization procedures. Moreover, the function used for the suspicion moment induced by the lifetime is emulating the end of the bathtub curve through a Heaviside function, which represents more likely the behavior of electronic system’s defect rate. Other parameters like failure rate of a component could not be learned, the experiments based on our dataset of feedback protocols show too much divergence in the learned values. In Azarian et al. [6], we conclude that the best values for a, b and cLT are 0.9, 0.4 and 0.2.

The feedback evaluation allows: to learn the causal dependencies between the components of a vehicle and weight their relative importance, to discover expert rules and prioritize them by updating their weight coefficient, and finally to learn the life-time and calibrate the test cost of the components (see Fig. 8). The changes for an implementation of these modules can be done with little efforts if the targeted system has enough opened interfaces as provided by SIDIS Enterprise. This approach leads to a bridge between both worlds of heuristic and model based, with a reduced effort for the model edition through a probabilistic method. The next section describes the performed experiments and reports the major results in terms of performance. 3. Experimental results 3.1. Case description Instead of carrying out the experiments with ADAPT [14], we prefer to select vehicles in order to show the whole potential of our approach on a complex mechatronics system which encompasses more physical domains then the ones in ADAPT’s testbed. For the simulations, we developed an independent prototype with nearly the same information model. The three tested vehicles belong to the low, middle and luxury car classes, which represents around 14.000 objects in the database. The simulated breakdowns contained single and multiple faults, and depending on the case, we fed the feedback engine with real or with highly divergent diagnosis sessions to monitor the stability. The performance indicator used was the orbit of a diagnosis session, which is the number of executed intermediate test for the fault isolation. The test cost assigned to the diagnostic object reflected real practical conditions including the car communication, the differences between OK and not OK answers to the test. In the first series of experiment (Section 3.2) the measures of the orbit of the diagnosis sessions were made without any feedback data. This means that the models contained only the expert rules. In the second series of experiment (Section 3.3), the data was updated with the results of the feedback evaluation such as discovered expert rules, causal relationships, life-time of components and updated weight coefficients. The comparison of the results obtained in both cases showed the usefulness of the data acquired by the return of experience. 3.2. Simulation without feedback data In Fig. 9 the orbit of the diagnosis session are represented for the different vehicles and for two or three cases (the curves of the diagnoses of vehicle B and C with SIDIS Enterprise are overlapped). In almost all cases, our approach presents better results (the exception being the last case vehicle A). The third case contains a multiple fault case where subcomponents of the ABS and ESP did not work and a wrong intermediate test answer had a dramatic impact on the orbit. In these vehicle models, no causal relationship had been entered, therefore it is only a heuristic diagnosis that is performed based on the modified engine of SIDIS Enterprise detailed in (8). The coefficient b lowered the suspicion moment in that case which, divided by the test cost, modified the second and third test agenda and included two unnecessary tests. On average, the orbit was diminished by 13.1%. It is also noteworthy that the evaluation of the suspicion rule by the logarithmic function alone permits to improve the diagnosis. This factor relies on a logarithmic function to assign a suspicion moment to the diagnostic object, which compared to the linear function of SIDIS Enterprise provides a better distinction between the candidates. The higher the number of active suspicion links, the higher its partial suspicion moment.

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

141

Fig. 8. Data extraction, evaluation and filtering for the diagnosis.

3.3. Simulation with feedback data Despite the possible automatic learning method and its implementation in the authoring system, the results must be analyzed carefully. The discovered rules correspond to hardcoded diagnosis knowledge or correlation and therefore simulate a physical reality.

In a first series of experiments we decided to automatically complete our database without any supervision. Due to an approximate value of 30 feedbacks per vehicles around 60 links (suspicions and causal) appear in the database. As a result, we observed on average an increase of 30% of the orbit due to new rules which provided too much dispersion in the fault localization process. In order to con-

142

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

Fig. 9. Orbit of the guided fault findings.

trol the learning process and the database manipulation, we configured limitations. Actually an acceptance limit could also be configured, for example in terms of number of protocols. A rejection limit should also be defined (due to the naïve hypothesis of the Bayesian network); for example, a causal link that has a weight of 0.0001 can correspond to a case that occurs once every 10,000 feedbacks! A reasonable acceptability limit for the weight is 20%, because it is around this limit that links have an impact on the test agendas. But depending on the size and the policy applied in a car manufacturer after-sales network (e.g. the protocol can be sent optionally or the protocol must be sent to the feedback server), that limit should be configurable. Generally, an acceptability limit around 10% less than the lowest weight is enough to keep the relevant links for the diagnosis. The results obtained are depicted in Fig. 10. For the comparison we also provided the orbit of the diagnosis sessions performed with SIDIS Enterprise with the models enhanced by the feedback data. These vehicle models contain causal relationships between the components, new expert rules, the life-time of the components and updated weight coefficients. In two cases (for the vehicle B and C), the learned values from the feedback evaluation cause the algorithm of SIDIS Enterprise to perform one more test. This is due to the test ranking metric detailed in (6) which uses the ratio suspicion by test cost for the ranking. In both cases the protocols of diagnosis sessions contain various breakdowns, from the headlight to the door-locking, whereas the faulty component is a temperature sensor of the

cooling system and the sensor of the break pedal. The dispersion of feedbacks causes the Bayesian network to diminish the weights of the relevant rules and causality links. The consequence is that this divergence is reflected in the test agenda with unnecessary tests. Despite this local phenomenon, the feedbacks decrease the average orbit of SIDIS Enterprise’s algorithm by 16% proving its contribution on a large number of cases. Furthermore, the average performance gain for the orbit of the diagnosis sessions is equal to 26%. 3.4. Synthesis and discussion The results in terms of performance are promising with an average decrease of 20% for orbit (without and with feedback data). The combined approach has a higher discrimination power of the diagnostic knowledge thanks to the entropy formula implying normalized logarithmic distinctions. The parameter b is of no importance when the models do not contain causal links, because the candidate generation still remains the same with different values b, except that the partial suspicion moment is scaled. It is the heuristic metric of the test ranking that can play a differentiation role due to the division by the test costs (if they are updated, for example). The parameter a influences only the precision factor, when the models do not contain causal links. For that reason, most of the suspicion links are pointing towards diagnostic objects with the same level; the influence of the precision factor is therefore negligible. Once

Fig. 10. Orbit of the guided fault findings with feedbacks.

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

the causal links are added to the model, the parameter b plays a central role because it balances the certainty between suspicion and causal links. If the balance exceeds the value 0.4 the causal links are nearly ignored and the orbits of diagnostic sessions with the hybrid approach are very similar to the ones obtained by SIDIS Enterprise’s algorithm. For lower values, the causal suspicious moment comes in the foreground for the candidate generation, with the consequence that the hybrid approach emulates a directed weighted graph exploration. The reason is that the causal relation corresponds to frequencies of correlated suspects, the graph exploration induces source elements (with the maximum of outgoing causal links) to be ranked first in the test agendas and these are not always the cause of the faulty behavior of the vehicle. This is the main drawback of the whole feedback engine: all adapted weights and new links correspond to frequently encountered cases and not to background knowledge. In other words, Baye’s rule reveals possible correlations but these are strictly different from any causal reality. An automatic self learning system as well as any large knowledge based system may contain erroneous data and therefore there is a need to perform a ‘‘data diagnosis’’. This aspect was already identified 20 years ago by [41] but due to the complexity of mechatronics systems and their multiple configurations this point will represent a crucial challenge for the applicability in the industry of knowledge based systems. Despite this, in the last decade, model-based technology in diagnosis has matured so far that it has been transferred from academic research into limited real applications providing useful features like explanation driven diagnostic or simulation. In fact, there is a lack of research on systems which re-use knowledge from previous technology and allow technological transfer. Our framework raises this opportunity as well as an important performance improvement through the combination of the different sources of knowledge. Moreover, for the start-up operation toward a model based culture a guide of best practice has already been published by [42] and the work of [5] presents useful information on the integration of IT-solutions.

4. Conclusions This paper had explored the field of diagnosis applied to the automotive industry. At the beginning we outlined the encountered difficulties faced by car manufacturer to find time-performant and accurate diagnostic tools for their workshops. The performance measurements resulted from our experiences on three vehicles (a low class, a middle class and a luxury class car), a case base of seven breakdowns and data fed with 90 protocols of diagnosis sessions, show a possible application on a large scale of our global framework on a whole range of vehicle of a manufacturer. The decrease of the orbit reaches an average value of two tests per diagnostic session. This represents an economy of 2500 working hours a day if scaled to the after sales network of a manufacturer in a country like Germany. Combined with our previous work (see Section 1.1.3), the module of symptom retrieval reaches an economy of 3.28 tests per diagnostic session and the module for the support of the edition of the ECU database reaches 37.5% time economies. But still some related areas such as automated modeling, automated knowledge extraction from documents, automated model abstraction and automated diagnosis on data should be explored in order to bridge the gap between research world and industry for a successful transfer of MBD technology. The work presented in this paper has allowed to provide a practical oriented framework for the automotive diagnosis, with the combination of different knowledge sources and their access for an efficient fault localization like: expert knowledge (suspicion rules), construction knowledge (causal links), return of experience (feedback engine)

143

with the result to diminish the time of diagnosis sessions in the workshops through the optimization of the models of diagnosis. Acknowledgment This work has been partly performed within Siemens AG Company (Germany). The authors wish to thank this company for the support and opportunities offered for performing this research work as well as the investment of the automotive department for giving access to the product SIDIS Enterprise, to training material (presentations, specifications) and for the discussions and sharing of the knowledge about the after-sales process in major automotive industries. We are also indebted to the members of the European Network of Excellence (MONET) and the Qualitative Modeling Group (MQM) for their support and their documentation. References [1] J. Baumeister, F. Puppe, D. Seipel, Refactoring methods for knowledge bases, Lecture Notes in Artificial Intelligence 3257 (2004) 157–171. [2] T.L. Cornelius, Intelligent Knowledge-Based Systems, Springer, 2004, ISBN: 9781402077463. [3] T. Chieux, C. Guillaneuf, The European Aftermarket–Trends in the Market and Pros-pects for Authorised and Independent Repairers, ICDP (International Car Distribution Programme) Research Report 05/05, Solihull, United Kingdom, 2005. [4] G. Leen, D. Hefferman, Expanding automotive electronic systems, Computer Journal 35 (1) (2002) 88–93. [5] W. Shen, Q. Haoa, H. Maka, J. Neelamkavila, H. Xiea, J. Dickinsona, R. Thomasa, A. Pardasania, H. Xuea, Systems integration and collaboration in architecture, engineering, construction, and facilities management: a review, Advanced Engineering Informatics 24 (2) (2010) 196–207, doi:10.1016/j.aei.2009.09.001. [6] A. Azarian, A. Siadat, M. Martin, A new strategy for automotive off-board diagnosis based on a meta-heuristic engine, Journal of Engineering Applications of Artificial Intelligence 24 (5) (2011) 733–747. [7] P. Struss, Model-based problem solving, in: F. van Harmelen, V. Lifschitz, B. Porter (Eds.), Handbook of Knowledge Representation, Elsevier, 2008, pp. 395– 465, ISBN: 978-0-444-52211-5. [8] Y.L. Murphey, J. Crossman, Z.H. Chen, J. Cardillo, Automotive fault diagnosis – Part II: a distributed agent diagnostic system, IEEE Transactions on Vehicular Technology 52 (4) (2003) 1076–1098. [9] A.R. Tyler, Expert Systems Research trends, Nova publishers, 2007, p. 217, ISBN: 1600216889. [10] H. Ueno, Y. Yamamoto, Fuduka H. Kiyoki, Y. Wangler, B. Jaakkola, Hannu Kangassalo, Hannu, Stock Image Information Modelling and Knowledge Bases XVI, 2011, Stratford Books ISBN: 1586034979. [11] T. Denton, Advanced Automotive Fault Diagnosis, second ed., 2006, pp. 40–89, ISBN: 0750669918. [12] A. Azarian, A Doctoral Thesis: A New Modular Framework for Automatic Diagnosis of Fault, Symptoms and Causes Applied to the Automotive Industry, University of Karlsruhe, ENSAM, 2009. [13] C. Ertz, Ohne DIS und VAS keine Chance bei der Fehlersuche, in SIK Journal: Das Magazin für den Siemens Industriepark Karlsruhe, 2008, pp. 6–7. [14] BMW Group, Pressemeldung: BMW Group steigert 2007 Absatz um 9,2 Prozent. 2008 Available on: . [15] Norm Picker, Shawn Puma, Scott Poll, Ann Patterson-Hine, Joe Camisa. Advanced Diagnostics and Prognostics Testbed (ADAPT) System Description, Operations, and Safety Manual, NASA Ames Research Center, ADAPT Projet, 2006. [16] T. Kurtoglu, S. Narasimhan, D.S.P. Garcia, L. Kuhn, J. de Kleer, A. van Gemund, A. Feldman, Towards a Framework for Evaluating and Comparing Diagnosis Algorithms, in: Proceedings of the 20th International Workshop on Principles of Diagnosis (DX-09), Stockholm, Sweden, 2009. [17] T. Kurtoglu, S. Narasimhan, D.S.P. Garcia, L. Kuhn, J. de Kleer, A. van Gemund, A. Feldman, First International Diagnosis Competition – DXC’09, in: Proceedings of the 20th International Workshop on Principles of Diagnosis (DX-09). Stockholm, Sweden, 2009. [18] J. Mauss, M. Volker, T. Mugur, Towards model-based engineering: Failure analysis with MDS, in: Proceedings of the ECAI-2000 Workshop on Knowledge-Based Systems for Model-Based Engineering, Berlin, Germany, 2000. [19] M. Sachenbacher, P. Struss, M.C. Claes, A prototype for model-based on board diagnosis of automotive systems, AI Communications 13 (2) (2000) 83–97. [20] A. Feldman, P. Jurryt, G. Arjan van, All roads lead to fault diagnosis: modelbased reasoning with Lydia, in: Proceedings of the Eighteenth BelgiumNetherlands Conference on Artificial Intelligence, Namur, Belgium, 2006. [21] E. Gould, Modeling it both ways-hybrid diagnostic modeling and its application to hierarchical system designs, in: Proceedings of the Annual Systems Readiness Technology Conference, San Antonio, Texas, USA, September, 20–23, 2004.

144

A. Azarian, A. Siadat / Advanced Engineering Informatics 26 (2012) 131–144

[22] K. Lunde, R. Lunde, M. Burkhard, Model-based failure analysis with RODON, in: Proceedings of the 17th European Conference on Artificial Intelligence, Riva del Garda, Italy, August 29–September 1, 2006. [23] Press release CD 37, MAG: Magazine d’information de l’outil de Diagnostic Citroën, December 2006. Available on: . [24] P. Struss, B. Ertl, Post-mortem diagnosis of bottling plants based on recorded data, in: Proceedings of the 7th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes (Safeprocess’09), Barcelona, 2009. [25] OCC’M Software GmbH, 2010. Available from: (last consultation: 18/06/2010). [26] J. de Kleer, B.C. Williams, Diagnosing multiple faults, Journal of Artificial Intelligence 32 (1987) 97–130. [27] SIDIS Enterprise User Manual, Siemens AG, 2004. [28] P. Bunus, O. Isaksson, F. Beate, M. Burkhard, RODON – A model-based diagnosis approach for the DX diagnostic competition, in: Proceedings of the 20th International Workshop on Principles of Diagnosis DX-09, Stockholm, Sweden, 2009. [29] O. Isaksson, Model-based Diagnosis of a Satellite, Electrical Power System with RODON. Master Thesis. Department of Electrical Engineering, Linköping University, 2009. [30] R. Everett, Diffusion of Innovation, Free Press, 2003, ISBN: 0743222091. [31] P. Struss, B. Ertl, Diagnosis of bottling plants – first success and challenges, in: proceedings of the 20th International Workshop on Principles of Diagnosis, Stockholm, 2009. [32] A. Evsukoffa, S. Gentilb, Recurrent neuro-fuzzy system for fault detection and isolation in nuclear reactors, Advanced Engineering Informatics 19 (1) (2005) 55–66, doi:10.1016/j.aei.2005.01.009.

[33] S. Piechowiak, Intelligence artificielle et diagnostic, Techniques de l’Ingénieur, traité Informatique industrielle, No. S7-217, 2004, can be find on . [34] D. Leake, Case-Base Reasoning: Experiences, Lessons and future Directions, Chapter 1, second ed., AAAI Press/MIT Press, 2000, ISBN: 0-262-62110-X. [35] X. Zhao, D. Ouyang, Improved algorithms for deriving all minimal conflict sets in model-based diagnosis, in: Proceedings of the Third International Conference on Intelligent Computing Qingdao, China, 2007. [36] M. Schwall, J. Gerdes, Multi-modal diagnostics for vehicle fault detection, American Society of Mechanical Engineers, Dynamic Systems and Control Division 70 (2002) 801–810. [37] C. Shannon, A mathematical theory of communication, The Bell System Technical Journal 27 (1948) 379–423. [38] A. Rényi, On measures of entropy and information, in: Proceedings of the 4th Berkeley Symposium on Mathematical statistics and probability, Berkeley University, California, USA, 1960, pp. 547–561. [39] E. Alpaydin, Introduction to Machine Learning, MIT Press, 2004, ISBN: 9780262012119. [40] O. François, P. Leray, Etude comparative d’algorithmes d’apprentissage de structure dans les réseaux bayésiens, Journal électronique d’intelligence artificielle 3 (2004) 1–19. [41] P. Struss, Extensions to ATMS-based diagnosis. In: Artificial Intelligence in Engineering: Diagnosis and Learning, 1988, pp. 3–27, ISBN: 9781853120121. [42] S. Smith Prabhu, L. Callanan, Z. Chambers, M. Herniter, Development of Model Based Design Curriculum, ASEE Annual Conference and Exposition, Conference Proceedings, 2007.