Diagnosis of continuous processes using a qualitative and quantitative model.

Summary In this thesis, we propose a model-based approach for the diagnosis of continuous processes. The objective is to detect process malfunctions and to search for their primary causes. The process is represented by two models with two different levels of abstraction: a quantitative causal graph, whose nodes represent the relevant variables of the process and whose arcs symbolize the causal relationships between these variables, and a qualitative sign graph that exploits the principles of sign algebra. The causal graph constitutes the model of the good behavior of the process which is used as reference for the detection of defects. The simulation of the dynamic behavior consists in the propagation of disturbances from exogenous variables through the causal graph. Fault detection is based on the comparison of simulated and measured data, through time windows. The diagnostic task allows to search for the set of primary causes that explain the drift, by going through the causal graph step by step backwards from the symptom variable. The consideration of time in the modeling allows us to manage the temporal link of causes and effects between variables. Each step of the research consists of two phases. First, the use of the sign graph to perform a qualitative filtering in order to prune the search space of candidate explanations and thus reduce the algorithmic complexity of the diagnostic task. In a second step, the use of the causal graph to perform a local numerical simulation in order to validate or reject definitively the candidates retained by the qualitative analysis. We applied a specific treatment to the variables belonging to the loops. The results of the diagnostics performed show the effectiveness of the combination of qualitative and quantitative analysis to reduce the search space.