BSEMD-Transformer: A New Framework for Rolling Element Bearing Diagnosis in Electrical Machines Based on Classification of Time–Frequency Features

Rolling Element Bearing (REB) failures represent a critical challenge in rotating machinery maintenance, accounting for approximately 45% of industrial breakdowns. Considering the variable operating conditions of speeds and loads, vibration fault signatures are generally masked by noises. Consequently, traditional diagnostic methods relying on time and frequency analysis or conventional machine learning often fail to capture the nonlinear interactions and phase coupling characteristics essential for accurate fault detection, particularly in noisy industrial environments. In this study, we propose a framework that synergistically combines (1) Empirical Mode Decomposition (EMD) for adaptive handling of non-stationary vibration signals, (2) bispectrum analysis to extract phase-coupled features while inherently suppressing Gaussian noise, and (3) Time-Series Transformer with attention mechanisms to automatically weight discriminative feature interactions. Experimental results based on five different benchmarks show that the proposed BSEMD-Transformer framework is a powerful tool for REB diagnosis, reaching a classification accuracy of at least 98.2% for all tests regardless of the used dataset. The proposed approach is judged to be consistent, robust, and accurate even under variable conditions of speed and loads.