Application of Frequency Aware Mechanism-Based Physical Information Convolutional Neural Network in Rolling Bearing Fault Diagnosis

Bearing fault diagnosis is critical for equipment health monitoring, as bearings are key components in mechanical systems. However, traditional deep learning methods, when processing vibration signals with complex physical phenomena, often rely solely on data-driven approaches, neglecting the integration of physical information, which leads to a lack of interpretability in the model's decision-making process. To address this, we propose a frequency-aware physical information convolutional neural network (PICNN) that enhances fault diagnosis by effectively embedding physical information. The vibration signal is first transformed into a time-frequency map using continuous wavelet transform (CWT). A frequency-aware convolutional (FAC) layer is then designed to generate sinusoidal waveforms corresponding to the bearing's fault frequencies, forming a frequency response kernel. This allows the integration of physical information into the model. The original convolution kernel is multiplied element-wise with the frequency response kernel to form an enhanced kernel, enabling the model to focus on fault characteristic frequencies. To improve network efficiency and representation, residual concatenation is used to enhance training. Additionally, the efficient channel attention (ECA) mechanism is introduced, boosting the network's focus on key fault features by adjusting the importance of feature channels. PICNN outperforms existing methods in fault classification accuracy, robustness, and generalization, particularly under varying loads and noise. Additionally, PICNN reveals the physical basis of the model's decision-making process through visualization, offering a clear explanation for fault diagnosis.