Research Program — Synergistic transformations in model based and data based diagnosis

Sensors are multiplying on machines, networks and living things. From the data produced and the symptoms observed as well as from verifications and tests, diagnosis aims at estimating the internal state of a “system” and to identify the nature and the cause of a failure, an anomaly or an illness. There are two main paradigms in the field of automatic diagnosis: model-based diagnosis and data-driven diagnosis.

Model-based diagnosis (MBD) is based on representing the system structure and behavior with a mathematical model that can be used to predict what sensors should observe. Inconsistencies not only allow to detect, but also to identify the fault responsible for a malfunction (Travé-Massuyès 2014). When one is interested in systems with continuous or hybrid dynamics, one carries out transformations of the model so as to exhibit fault indicators computable from the data returned by the sensors. Structural analysis or elimination theory can be used for this in the case of analytical nonlinear models. We conducted work according to both approaches (Perez et al., 2017)(Chanthery et al. 2016)(Jauberthie et al. 2016), (Verdière et al. 2015). When one is interested in discrete-event systems, most approaches use explicit behavioral models and obtain diagnoses by corroborating faulty behavioral patterns with the observations, which again involves model transformations (Bouziat et al. 2018)(Pencolé et al. 2017)(Cordier et al. 2006). Data-Based Diagnosis (DBD), for its part, proceeds with an exploration of sensor data to infer, by learning, a classification model (Ding 2014). We have built a long-standing experience in this field (Barbosa et al. 2016)(Hedjazi et al. 2015)(Carrete & Aguilar-Martin 1991). Beyond the fact that some classification methods are supervised, i.e. they require historical data labeled by the desired abnormalities/faults, and others are unsupervised, we will focus here on the fact that some do not allow a nonlinear separation of the data while others allow it. Among these are the kernel methods that make it possible to find nonlinear decision functions thanks to specific functional transformations (Steinwart & Christmann 2008).

The idea of this project is to analyze the transformations performed by MBD and by DBD diagnosis methods. The first objective is to highlight and understand the correspondences that may exist between them and how they could synergize each other. The second objective is to be able to abstract up data classifiers and map them to symbolic or analytical models suitable for diagnosis reasoning, gaining better explainability and acceptability. To our knowledge, there is no work that has analyzed MBD and DBD approaches, trying to match them, possibly deriving mutual benefits, and integrating them closely. In this sense, this project is innovative and original.

Integration within ANITI program

The proposed project is perfectly consistent with ANITI's research core "Hybrid Track I: From Subsymbolic to Symbolic and Back". It is related to ANITI IP4 “AI Assistants” because diagnosis, including monitoring complex systems, detect anomalies to prognosis and health management, is a task for which an integrated human/AI system is particularly relevant and leads to better performance. It is also related to ANITI IP3 "AI towards autonomous critical systems" because diagnosis allows to estimate the internal state of a system and is hence an essential function for adaptable autonomous systems.

The Application Areas defined by ANITI that are targeted by this project are Transport/Mobility and Industry 4.0 (Continental, ACTIA, AIRBUS, ATOS, CS Communication & Systems have already expressed their interest).

Education Topics: Logic theory of model based diagnosis, Diagnosis and diagnosability of discrete event systems, Fault indicators generation via structural analysis, Machine learning for diagnosis and supervision, Prognosis and Health Management, Bridging control and artificial intelligence theories for diagnosis, Ethics and integrity in science.

AI dissemination activities to the general public: Popularization talks, Meetings in high schools, Female high school student’s awareness of AI, AI-based escape game.