Abstract
Fault monitoring plays a vital role in ensuring operating safety and product quality of industrial manufacturing processes. However, modern industrial processes are generally developing towards the direction of large scale, diversification, and individuation, complexity, and refinement, exhibiting strong non-linearity and dynamically time-varying characteristics, leading to a great challenge in fault monitoring. This paper addresses a fault monitoring method based on a dynamically-recursive kernel principal component analysis (DRKPCA) model with a variational Bayesian Gaussian mixture model (VBGMM), called DRKPCA-VBGMM, for the continuous, time-varying process monitoring. Specifically, a computationally efficient DRKPCA scheme is derived for the anomaly/fault detection of time-varying processes. Successively, a variational inference-induced optimal Gaussian mixture model, called VBGMM, is introduced for the fault type identification, which can automatically converge to the real number of Gaussian components based on the empirical Bayes approach to achieve the optimal probability distribution model. Extensive confirmatory and comparative experiments on a benchmark continuous stirred tank reactor process and a continuous casting process from a top steelmaking plant in China have demonstrated the effectiveness and superiority of the proposed method. Specifically, the proposed method can effectively improve the fault detection and identification accuracies while reducing false alarm rates, laying a foundation to ensure stable and optimized production of complex manufacturing processes.