An improved dynamic latent variable regression model for fault diagnosis and causal analysis

The advancement of industrial techniques has imposed a high demand for powerful machine learning algorithms to model the increasingly complicated relations in the data. Among them, dynamic models are widely studied to capture the inevitable temporal relations. However, most existing methods only focus on the dynamics between input and output data, failing to exploit other valuable information in the output. In this article, an improved dynamic latent variable regression (LVR) method is proposed to capture both auto-correlations and cross-correlations between input and output with an auto-regressive exogenous model, which is referred to as dynamic regularized LVR with auto-regressive exogenous input (DrLVR-ARX). Further, a DrLVR-ARX-based fault detection and diagnosis framework is designed to identify the root causes of a detected fault. The framework systematically integrates reconstruction-based contribution, time-domain Granger causality, and conditional spectral Granger causality to determine and locate the assignable causes. The effectiveness of the proposed algorithms is demonstrated with two industrial processes.