Improved Dynamic Latent Variable Modeling for Process Monitoring, Fault Diagnosis and Anomaly Detection

Abstract

Due to the rapid advancement of modern industrial processes, a considerable number of measured variables enhance the complexity of systems, progressively leading to the development of multivariate statistical analysis (MSA) methods to exploit valuable information from the collected data for predictive modeling, fault detection and diagnosis, such as partial least squares (PLS), canonical correlation analysis (CCA) and their extensions. However, these methods suffer from some issues, involving the irrelevant information extracted by PLS, and CCA’s inability to exploit quality information. Latent variable regression (LVR) was designed to address these issues, but it has not been fully and systematically studied. A concurrent kernel LVR (CKLVR) with a regularization term is designed for collinear and nonlinear data to construct a full decomposition of the original nonlinear data space, and to provide comprehensive information of the systems. Further, dynamics are inevitable in practical industrial processes, and thus a dynamic auto-regressive LVR (DALVR) is also proposed based on regularized LVR to capture dynamic variations in both process and quality data. The comprehensive monitoring framework and fault diagnosis and causal analysis scheme based on DALVR are developed. Their superiority can be demonstrated with case studies, involving the Tennessee Eastman process, Dow’s refining process and three-phase flow facility process. In addition to MSA approaches, autoencoder (AE) technology is extensively used in complicated processes to handle the expanding dimensionality caused by the increasing complexity of industrial applications. Apart from modeling and fault diagnosis, anomaly detection draws great attention as well to maintain the performance, avoid economic losses, and ensure safety during the industrial processes. In view of advantages in dimensionality reduction and feature retention, autoencoder (AE) technology is widely applied for anomaly detection monitoring. Considering both high dimensionality and dynamic relations between elements in the hidden layer, an improved autoencoder with dynamic hidden layer (DHL-AE) is proposed and applied for anomaly detection monitoring. Two case studies including Tennessee Eastman process and Wind data are used to show the effectiveness of the proposed algorithm.