 |
Qualitative Diagnosis and Supervisory Control |
 |
The group research lines are the following:
• Diagnosis and decision-making — Model based approaches
• Monitoring and Supervision based on learning and pattern recognition approaches
• Optimisation and supervisory control of non linear systems
The DISCO group develops diagnosis and supervision tools for complex dynamic systems, possibly including human operators. We do stress the qualitative nature of the data that is relevant to supervision in resonance with the continuous aspects of the systems. Hence, hybrid systems and the interface between continuous signals and their interpretation in terms of discrete events are at the core of our activities. The qualitative nature of knowledge and the uncertainties tainting the data induce us to call on qualitative and symbolic formalisms originated from Artificial Intelligence as well as on machine learning and pattern recognition methods.
Diagnosis and decision-making — Model based approaches
Diagnosis can be defined in a broad sense as a set of three tasks: fault detection, isolation and identification. Diagnosis is of prime necessity in many application domains, for instance for monitoring of industrial plants or in the space domain for increasing the autonomy of spacecrafts. Our contributions are both for off-line and on-line diagnosis. Most of our case studies consist of dynamic systems whose models may be spread out in continuous, discreet or hybrid state spaces. The group’s background work addresses modelling and qualitative reasoning, comparing the approaches from the Control and the Artificial Intelligence communities, analysing diagnosability. In other respects, the following projects are representative of recent research topics developed in the DISCO group:
• Automatic generation of optimal diagnosis trees for off-line diagnosis: this problem is solved by a heuristic search method (AGENDA prototype) complemented, during the tree traverse, by a qualitative model based diagnosis method accounting for non anticipated faults. This work finds a preferential application in the automotive domain in the framework of the Common Laboratory AUTODIAG, in partnership with two neighbouring entities which are the ACTIA company and the research laboratory IRIT.
Diagnostic, suivi et reconfiguration à base de modèles hybrides incertains. Application à un sous-système du satellite d'observation de la terre Spot 5
• On-line diagnosis and reconfiguration of uncertain hybrid systems: the proposed hybrid model (KOALA formalism) integrates concurrent automata whose states represent the component operating modes and bounded uncertainty continuous models. The approach relies on a logic representation — called configuration — of the different state space regions, which makes possible the extension of the existing logical theory of diagnosis. The devised diagnosis engine interleaves progressive search in the hybrid space and consistency-checks, making use of truth maintenance and propositional satisfaction techniques. This work was illustrated on the Attitude and Orbit Control system of a SPOT satellite case study in collaboration with CNES and ASTRIUM.
Points of control
• Distributed supervision of complex discrete event systems approached from the two following perspectives: model distribution in the form of a set of Petri net models designed according to the physical organisation of the system; distributed monitoring based on event models expressing temporal constraints under uncertainty related to communication delays.
Monitoring and Supervision based on learning and pattern recognition approaches
This research line aims at providing multisensory fusion tools providing the operators of industrial plants or other artefact with easy-to-interpret information in normal operating conditions as well as in faulty situations. The problem consists of interpreting a set of continuous signals in terms of “operating modes” that have a meaning to the operators and/or to generate directly a performance/criticality index corresponding to the current state of the system. The tools must be consistent with the mental representation of the process that the operators draw up, which calls for interactive tools.
To build these tools, we call on signal and image processing methods which are able to extract relevant events, and from the other hand on classification and pattern recognition methods to characterize the functional states of processes, possibly on-line.
We position ourselves as users of classical statistical classification methods, decision trees, neural networks, fuzzy algorithms like “c-means”. However, our efforts are focused on the development and enhancement of the classification method LAMDA that was devised in the group and includes useful features for combining quantitative and qualitative aspects.
LAMDA was recently applied to petrochemical processes in the framework of the European project CHEM coordinated by IFP
Functional states recognized by LAMDA
In another project, LAMDA is used, possibly combined to fault detectors implemented by non linear observers, in the domain of waste water treatment.
Optimisation and supervisory control of non linear systems
The need for a supervisory control loop, supervising the low-level control loop exists when the system, or its model is far from standard cases, linear or of low dimensional complexity. The DISCO group conducts recognized activities in research domains that are linked to process control, like Takagi-Sugeno fuzzy control, or non linear inverse control which are turned to good account in supervisory control schemas. A great deal of our activities in this research line considers problems in the aeronautic domain, from traffic problems on the ground or in flight to aircraft control problems. These works can be illustrated by the applications in the framework of two collaborative projects, one with CENA on aircrafts relative navigation, and the other with EADS-Airbus Industry on automating the control of aircraft ground taxiing operations.
|
|
|
