Physics-Informed Time-Frequency Fusion Network With Attention for Noise-Robust Bearing Fault Diagnosis

Abstract

We propose an accurate and noise-robust deep learning model to diagnose bearing faults for practical implementation in industry. To achieve high classification accuracy in a noisy environment, we designed a time-frequency multi-domain fusion block, incorporated bearing-fault physics into the model parameters, and employed attention modules. The proposed model individually extracts essential features from the time-domain vibration signal and the corresponding spectrum in a parallel pipeline. Subsequently, multi-domain feature maps are fused to capture a wider representation of bearing fault signals. The performance was enhanced by incorporating physical knowledge of fault frequencies in the design of the frequency-domain feature extraction network. The employment of an attention mechanism to selectively focus on high-importance fault characteristics on the multi-domain feature maps further improved the accuracy under high noise levels. Experiments on bearing datasets with artificially added noise demonstrated the effectiveness of the proposed model compared to other benchmark models. INDEX TERMS Additive white gaussian noise, Ball bearing, Condition monitoring, Convolutional neural networks, Fault diagnosis, Feature extraction, Machine learning algorithms, Prognostic and health management, Reliability, RobustnessRecently, the rapid growth of computing power and the exceptional performance levels facilitated by advanced model architectures have brought deep learning models to greater prominence in studies of bearing diagnostics [12]. In contrast to traditional machine learning approaches that require the