Intelligent diagnosis of compound faults in rolling bearings based on ensemble window AR-PSD

Under complex operating conditions involving multiple faults, load variations, speed variations, and noise interference, rolling bearing vibration signals show strong non-stationary and nonlinear characteristics. Traditional feature extraction methods often have difficulty in representing complex fault information effectively. To address this problem, a fault diagnosis method combining ensemble window auto-regressive power spectral density (EWAR-PSD) and efficient channel attention mechanism into VGG16 (ECA-VGG16) is proposed. EWAR-PSD introduces a sliding window and noise perturbations into the conventional auto-regressive model. It converts one-dimensional vibration signals into two-dimensional power spectral features. This method preserves local spectral variations, enhances fault-related frequency components, and reduces the influence of random noise. The extracted spectral features are then fed into ECA-VGG16 for multi-class fault identification. The proposed method is evaluated under variable-load, variable-speed, and noisy conditions and compared with several representative fault diagnosis methods. The results show that it achieves high and stable diagnostic accuracy under complex operating conditions and exhibits strong robustness to noise. These results demonstrate its effectiveness and potential for engineering applications in rolling bearing fault diagnosis.