Enhanced diagnostic method for rolling bearings using time-reassigned multi-synchro squeezing transform

In response to the significant challenges posed by strong non-stationarity and the vulnerability to intense background noise in rolling bearing signals, as well as the inherent limitations of conventional convolutional neural networks (CNN) when processing one-dimensional (1D) signals without fully leveraging the inter-data relationships, this study introduces an innovative diagnostic approach for rolling bearings. The method employs the Time-Reassigned Multi-Synchro Squeezing Transform (TMSST) to preprocess 1D vibration signals. By harnessing the temporal correlations across various intervals, TMSST generates a set of time-frequency feature maps that are subsequently fed into a CNN to adaptively extract and classify the fault characteristics of rolling bearings. To substantiate the efficacy of the proposed model, the Case Western Reserve University's bearing dataset serves as the benchmark for the fault diagnosis analysis. Moreover, the study incorporates several alternative data processing techniques for comparative evaluation of the classification accuracy. The findings reveal that the proposed model, when juxtaposed with other image encoding methods, consistently delivers superior diagnostic performance across a spectrum of load conditions and noise environments. It achieves an impressive global accuracy of 95.67 %, thereby facilitating robust end-to-end fault pattern recognition in rolling bearings.