Machine learning for model-based diagnosis

Process diagnosis is the task of classifying observed process measurements as consequences of specific process faults. Model-based techniques such as state and parameter estimation, combined with statistical hypothesis testing, have been a common theme in the literature on fault detection and diagnosis. Pattern classification techniques such as template matching, discriminant analysis and cluster analysis are among machine learning techniques that are suitable when accurate theoretical process models become not readily available. One major difficulty with applying pattern classification techniques has been the selection of an appropriate feature space. Automatic feature selection techniques can generate only linear features, while neural networks which generate a large number of empirical nonlinear features, require large data sets and tend to overfit process data. Furthermore, most applications of pattern classification techniques have been limited to stationary data. The paper addresses the integration of a parsimonious machine learning technique and a model-based scheme for feature selection. The advantages of the proposed approach include: (a) robust generalization using complexity-based regularization, (b) automatic extraction of intuitive fault signatures from both transient and steady-state data, (c) ability to handle compound faults, (d) a confidence measure for fault diagnosis decisions is readily available in terms of prediction error of process models and a measure of complexity