Diagnosis and Supervisory Control - DISCO

Head: Yannick PENCOLE
Phd Student Representative:Ibis VELASQUEZ 
Financial assistant: Justine Praneuf

News:  DISCO is co-organizing with VERTICS the 14th French speaking symposium on Reactive System Modeling (MSR'23), 22-24 November 2023.

Website: https://msr2023.sciencesconf.org/

Team intranet

The objective of the DISCO team is to develop a broad-spectrum methodological research in the field of diagnostics. This team, whose scientific skills are at the frontiers of Automation and Artificial Intelligence, aims to study abductive reasoning, applied to a wide variety of system classes (static systems, dynamic systems: discrete, continuous or hybrid). The fundamental principle of a diagnostic process is to compare the  observation of a real system (uncertain or not, noisy measurements, alarms, messages, tests) with the available knowledge of this system (models, uncertain or not) in order to establish a health state. This research is motivated by the fact that diagnosis is crucial to improve, among other things, the safety, resilience and maintainability of systems. To conduct this research, DISCO's activities can be divided into three types.

The DISCO team contributes to the full chain of a diagnostic process:

1.  Acquisition models for diagnosis by machine-learning methods.
   
   DISCO develops static classification methods with extraction of indicators (normal, abnormal, healthy, healthy, sick, faulty, etc.) on populations of individuals exploiting fuzzy logic [hal-01998674], neural networks [hal-01998620]. A dynamic classification method, on the other hand, offers the possibility of learning dynamic states of behaviour and detecting changes between these states by also detecting new states [hal-02135580]. DISCO also contributes to the machine-learning of time models (chronicles) by using time data mining techniques on system logs to discover time-discriminatory operating classes [hal-01817529], [hal-01611635].
    
2.  Analysis of diagnostic models by formal methods.
   
   Whether the models are obtained by machine-learning or by expertise, the development of an efficient and effective diagnostic method requires analyses of the system model beforehand. The performance of a diagnostic method is based not only on the quality of the models but also on the measurement capacity and the structure of the system. DISCO develops methods for verifying the diagnosability of discrete [hal-01574475], continuous [hal-01198408] and hybrid systems [hal-02004402] by identifying critical pairs (certificates) whose non-existence ensures that it will always be possible in finite time to provide a single diagnosis. Some work on sensitivity analysis in continuous systems also proposes to define experimental conditions that ensure faults will be detected [hal-01914534]. Structural analyses of such systems are also proposed in order to define relevant diagnostic tests in advance, to select discriminating sensors or to decentralize diagnostic algorithms [hal-01882324].
    
3. Development of diagnostic algorithms
   
   The synthesis of diagnosers (diagnostic algorithms) is based on the existence of a model, a flow of observations and a level of objective depending on the knowledge derived from the model. The basic objective is the estimation of system states. One of the difficulties of this estimation is the management of the different sources of uncertainty (set-membership computing, interval/ellipsoid/zonotope, interval Kalman filter) in continuous systems [hal-01739540], [hal-01884592], [hal-01884636] or in hybrid systems (reachability) [hal-01650701] and also the management of statistical uncertainty sources (interval Kalman filters [hal-01561951]) . The second level of objective is the location of faults in the system (where is the fault?). DISCO develops localization techniques by setting up CSP-type problems (constraint satisfaction problems) in logical systems [hal-01929533], or by exploiting the theory of the residual of (max,+) algebra on timed discrete event systems [hal-01954270].  The third level of objective is fault identification. This diagnostic problem is addressed on all types of systems: model-checking diagnosis with Petri nets [hal-01827362], non-linear observers [hal-01763870] or exploitation of analytical redundancy expression on continuous/hybrid systems [hal-01229077] , management of bounded and mixed uncertainties [hal-01739540], use of distributed diagnostic methods [hal-01229097]. DISCO also contributes to the synthesis of plans and orders taking into account faults (fault-tolerant control)   [hal-01884700] or used to refine a diagnosis (active diagnosis) [hal-02056090]. The fourth and last level of objective is fault/failure explanation where we look in the system for unanticipated behaviours (patterns, propagation of events) that are the cause of the appearance of the fault (explainable system behaviours, algorithm explainability).
    
4. Diagnostic/prognostic integration for predictive maintenance
   
   DISCO is studying the integration of diagnostic methods with ageing prognosis methods in order to improve the predictive maintenance of systems. The objective is to provide a degradation diagnosis that feeds an aging model of system equipment to determine the residual useful life of the system. DISCO addresses this theme by modelling the prerequisites of a good ageing model  [hal-01993849] and by formally defining a diagnostic and prognostic framework (Hybrid Particle Petri Nets) [hal-01229083]. DISCO also develops diagnostic/prognostic coupling techniques based set-membership computing [hal-01929470].