Self-Attention Sliding Window Enhanced Canonical Correlation Analysis for Incipient Fault Detection in Dynamic Industrial Processes

Thanks to its excellent nonlinear representation ability, deep neural network (DNN) has been designed to aid canonical correlation analysis (CCA) for stable kernel representation. These DNN-aided CCA methods make use of correlation as the optimization objective and exploit its change to distinguish system states. However, the maximum correlation-based training strategy lacks robustness to effectively tackle system false alarms caused by changes in operating conditions. To this end, this work proposes a self-attention sliding window enhanced CCA (SaECCA) for incipient fault detection in dynamic industrial processes. The main novelties of this work include the following: first, with the aid of DNNs, an adaptive weighting mechanism with self-attention is developed to amplify the incipient fault information; second, residual estimation-based, a novel and robust optimization objective for nonlinear CCA is formulated; third, an SaECCA-based fault detection algorithm is designed, whose convergence and detectability are illustrated via theoretical analysis. Studies on a three-tank system simulation and a multiphase flow industrial process are presented to verify the effectiveness of the proposed SaECCA method.