Robust Fault Detection in Rotating Machinery under Variable Load Conditions Using Integrated Multiscale and Multiway PCA

Abstract

Various methodologies based on frequency domain and multivariate statistical analysis are employed for the interpretation of vibration data. However, the impact of load fluctuation on these diagnostic techniques has not been thoroughly delineated. This study investigates the impact of loading variations on vibration response and proposes that a combined wavelet and multiway principal component analysis (MPCA) methodology can enhance the sensitivity of vibration-based anomaly detection techniques for rotating machinery. The methodology utilizes multivariate vibration signals acquired from several sensor locations and under diverse load situations, resulting in a three-dimensional dataset. It is employed to denoise PCA on wavelet-transformed data. The dataset is subsequently converted into a two-dimensional matrix using a novel unfolding technique, wherein each individual batch recorded under a certain load is regarded as an observation for further PCA analysis. This approach facilitates the analysis of vibration signals in relation to baseline data obtained from standard operational conditions under varying applied loads, therefore differentiating load-induced variability from genuine mechanical flaws. Numerous simulation studies on bearing defect detection have been conducted to validate the suggested methodology.