Hierarchical Sparse Causal-gated LSTM Networks for Root Cause Identification of Industrial Processes

Accurately identifying the root cause of faults is of great significance for the maintenance and health management of industrial processes. However, the intricate causal relationships among multiple variables in complex industrial processes make the root cause elusive, which brings challenges to root cause identification (RCI). To address this issue, this work proposes the hierarchical sparse causal-gated long short-term memory (LSTM) networks. First, the causal-gated matrix with sparse penalty which combined with LSTM is designed to extract significant causal relationships during time series prediction. In addition, by decomposing the entire system into multiple groups, the proposed method enables hierarchical causal modeling at both coarse and fine granularities. Experimental results on the Tennessee Eastman process (TEP) simulation platform demonstrate that the proposed method can capture the concise causality and obtain more explicit root cause.