A Learning Based Stochastic Approach for Fault Diagnosis in a Continuous Stirred Tank Reactor

Many approaches have been developed to detect and diagnose the different types of faults that may occur in a complex process. Most of these approaches have traditionally been based on linear modeling techniques, which restricts the type of practical situations that can be modeled. Recently, many learning based non linear modeling using neural and other on-line approximation models have been developed. This paper presents a learning based stochastic methodology for constructing automated fault detection and diagnosis architecture for Continuous Stirred Tank Reactor (CSTR). The main idea behind our approach is to use Hidden Markov Models (HMMs). By applying off-line training procedures, different fault models may be constructed using only system measurements. Based on these models, on-line fault diagnosis may be achieved. A Maximum Likelihood approach is used for fault detection. Some practical aspects of the different algorithms are discussed and simulation results using Scilab are presented.