A multi‐sensors nonlinear industrial process fault detection method based on Procrustes analysis transfer learning

Abstract The primary challenge currently faced lies in insufficient and unevenly distributed data, stemming from operational mode changes that degrade model training accuracy and fault detection precision. To address these challenges, this paper proposes a multi‐sensor nonlinear industrial process fault detection method based on Procrustes analysis transfer learning (PATL). Firstly, the Gaussian kernel encompassing multiple nonlinear operational processes is spread via Tucker decomposition (TD). Subsequently, a correlation matrix is derived based on correlation analysis (CA). Through Procrustes analysis, knowledge transfer between domains is achieved by translating, rotating, and scaling the correlation matrices of the source and target domains. After transfer, the correlation matrix is projected onto a shared space, which is then partitioned using the canonical component decomposition (CPD) method into a common subspace for fault detection in the target process and a specific subspace. However, traditional statistics may generate false alarms due to the impact of mode switching. To address these issues, this paper introduces the DPCA residual statistic. Finally, fault detection is performed using the derived detection metrics. Experimental results validate the effectiveness of the proposed method.