Induction motor process monitoring in power plants based on multi-step reconstruction-based PCA

Abstract

Fault diagnosis of induction motors is crucial for enhancing the reliability of industrial processes. Reconstruction-based principal component analysis (RB-PCA) is commonly used in fault diagnosis for industrial equipment because it effectively solves the problem of smearing effects. However, RB-PCA encounters challenges related to temporal inconsistency in the industrial processes. This issue arises in the early stages of a fault, where fault indicators fluctuate around the control threshold. Such oscillations can cause the model to switch intermittently between reconstruction and non-reconstruction states, which diminishes diagnostic accuracy and model stability. This paper provides a multi-step reconstruction-based principal component analysis (MS-RBPCA) algorithm that integrates a moving time window. Additionally, spatial distance reconstruction and sequence floating forward search are introduced to improve the computational efficiency of fault isolation. The effectiveness of the MS-RBPCA is demonstrated through one simulation study and one industrial case involving fault samples from induction motors in a power plant. The results show that MS-RBPCA can significantly reduce computational time, achieving a speed improvement of up to 50% while maintaining the fault detection rate above 97% and the false alarm rate below 1.5%, providing a viable solution for industrial process monitoring.