A probabilistic model with spike-and-slab regularization for inferential fault detection and isolation of industrial processes

Abstract This article develops a Bayesian latent variable model for inferential fault detection and isolation using a spike-and-slab regularization technique. Different from conventional probabilistic models like probabilistic principal component analysis (PPCA), a mixture prior with spike-and-slab component is introduced to indicate whether the latent variables are sensitive to the abnormal state. The relevant information of latent variables is selected by the probability of assigning the effect to the slab component, while the influence of the non-informative ones is eliminated by the spike component. Furthermore, a Bayesian inference scheme based on the expectation-maximization framework is developed for the parameter learning procedure of the model. The fault detection and diagnosis method based on the inferential model is subsequently developed. The performance of the proposed method is illustrated by the benchmark Tennessee Eastman process and an application to an industrial methanol distillation process.