Incremental Gaussian Mixture Model for Time-varying Process Monitoring

Abstract

With the increasing complexity of industrial production, data-driven based monitoring methods attract more attention. However, the conventional static process monitoring methods may show poor performance for the time-varying processes since they fail to track the time-varying characteristics. As Gaussian Mixture Model (GMM) has been widely used for process monitoring, this paper presents a new incremental GMM model for monitoring time-varying processes. First, an incremental GMM (IGMM) model is proposed, which can recursively update model parameters, adaptively add new Gaussian components and discard the irrelevant component based on the shifting samples online. Then the Bayesian Inference Probability (BIP) is introduced for monitoring statistics and a two-level partition strategy that can separate normal shifting samples from fault samples is proposed, which reduces the possibility of adding fault samples to the model. On the basis of IGMM model, an adaptive monitoring scheme is developed, which can track the time-varying characteristics of processes. Finally, a time-varying numerical example and the Tennessee Eastman process are adopted to validate the feasibility of the proposed monitoring model. Experimental results clearly demonstrate the adaptiveness of the monitoring model to time-varying processes and the ability to avoid false updates.