Fault diagnosis with reconstruction-based contributions for statistical process monitoring

Fault detection and diagnosis are important for the safe operation and control of a process, and for the reduction of its operation and maintenance costs. The implementation of mechanisms for the early detection and diagnosis of faults is called process monitoring. Due to the size and complexity of industrial processes, multivariate statistical methods are finding wide application in process monitoring. Some popular methods are principal component analysis (PCA) for linear processes, and kernel principal component analysis (KPCA) for nonlinear processes. In statistical process monitoring, faults are detected with fault detection indices that trigger alarms when an index has violated its control limit; after a fault is detected, a diagnosis method is used to find its root cause. A popular method for fault diagnosis has been contribution plots. This method uses the idea that a variable with a high contribution to a faulty index is likely the cause of the fault; however, there is no guarantee that a faulty variable would have the largest contribution. For the case of nonlinear processes, very few methods are available for fault diagnosis with KPCA models.;In this dissertation, a new diagnosis method is proposed based on the reconstruction of a detection index along an arbitrary direction. The method is called reconstruction-based contributions (RBC) and is able to provide contributions along single variable, univariate and multivariate directions.;An analysis of the diagnosis power of the RBC method shows that it guarantees correct diagnosis of sensor faults with large magnitudes while the traditional contributions do not. Analysis of the RBC and other diagnosis methods shows that several of these methods can be unified into some general methods; further analysis shows which of them fail to guarantee correct diagnosis for sensor faults with large magnitudes.;The RBC method is extended to the diagnosis of faults in nonlinear processes using KPCA models, where the RBC values are calculated along univariate and multivariate directions using nonlinear optimization methods.;Analysis and simulations show the effectiveness of the RBC method for the diagnosis of simple and complex faults, with univariate or multivariate directions that are known or unknown; and that happen in linear or nonlinear processes.