Research

System Identification

System identification is the procedure which consists in inferring a mathematical model from a set of input-ouput measurements collected from a dynamical system. We consider mainly blackbox approaches. The first step is to setup a parameterized model structure which is likely to be convenient for describing the system to be identified. Then, comes the step of fitting the collected data to this structure based on optimization or various learning techniques. A third step may be the validation of the model obtained then, that is the demonstration of its ability to predict up to a certain acceptable accuracy, the system behaviour based on its past inputs and outputs.

Hybrid Systems

Hybrid systems are complex dynamical systems which exhibit interacting continuous and discrete-event behaviors. Roughly speaking, hybrid systems can be regarded as switching among a number of dynamical operating modes. The switchings may be state-driven, time-driven, event-driven or totally random. Typical examples of hybrid systems are bouncing ball, walking robots, biological cell growth, chemical processes. The identification of such systems from input-output data is a challenging problem. Our research concentrates mainly on this task and we consider both state space and input-output ARX models.

Machine Learning

We are concerned with data clustering using statistical learning methods, euclidean subspace segmentation techniques, support vectors machines or more generally kernel methods. Given a mixture of data of different natures to a certain extent, the objective of our work is to group or partition them according to some similarity function.

Fault Detection and Isolation

Linear multivariable systems possibly subject to actuator or sensor faults are considered. Based on subspace identification techniques, state filtering, and appropriate dedicated observers, we construct some robust fault indicators. This work aims at detecting first the occurence of a fault and then identifying the affected component. Subsequently, an action can be decided to isolate the fault.