Fault diagnosis and fault-tolerant control of nonlinear dynamic systems using artificial intelligence techniques

(English) This thesis attempts to reduce the gap between the approaches developed by researchers to build a fault tolerant system and their implementation in a real system. The most widespread methodologies are model-based approaches which require a representative description of the behaviour of the system to use them. This model is usually not available, either because the system corresponds to a complex process or is poorly understood, or the performance of the system is not static in the sense that the behaviour depends on the current state and the situation (environment) of the system. Thus, obtaining a model based on the knowledge of the system will require a high effort and the accuracy of the resulting model will often be not sufficient to apply the approaches successfully. In this situation, considering generic or adaptable models will allow to reduce the time required to design a fault tolerant system, creating a more flexible process. In this thesis, artificial intelligent (AI)-based models are considered to obtain adaptable models. Moreover, a procedure to train these structures (pa-rameter estimation), in order to achieve an accurate and representative model, is also proposed. The main issue of AI-based models is their incompatibility with linear-like techniques which have been developed to identify or accommodate a fault in a nonlinear system. The transformation of these models into a promising structure which embeds the nonlinearities in the varying parameters, the so-called linear parameter varying (LPV) model, is considered in this thesis as a bridge to combine the classical linear or nonlinear model-based approaches with the learning ability of AI techniques. The polytopic representation of an LPV model, obtained from an AI-based model, will lead to a high number of vertex controller/observer gains, which may make the implementation of an LPV controller/observer into a micro-controller to become infeasible. This representation is used to reduce the controller/observer conditions to a finite-dimensional problem that can be solved using linear matrix inequalities (LMIs). The real-time implementation of a fault-tolerant control approach should be feasible using the heuristic algorithm developed in the thesis. By means of reducing the number of controller/observer gains, defining a common controller/observer for a set of vertices (vertex reduction), the algorithm will be able to provide a compact controller/observer that will be realizable and