Adaptive Wavelet-Based Physics-Informed CNN for Bearing Fault Diagnosis

With the increasing expansion of data science into various fields, the application of deep neural networks in the fault diagnosis of rotating machines has attracted significant attention from researchers. However, in the methods available in the literature, the physical characteristics of the problem are not incorporated into the structure of deep networks. In most existing methods, fault diagnosis is performed solely based on features extracted by convolutional layers, with no additional layers utilized to enhance or refine these features. This work introduces a novel physics-based neural network for bearing fault diagnosis, in which specific layers are designed based on signal processing methods to extract the physical features of faults. These layers, referred to as physics-based layers, are constructed using adaptive analytical wavelet filterbanks. The features extracted by these layers are then classified using convolutional layers, enabling the diagnosis of bearing faults. A key advantage of this physics-based network is that it does not rely on a fixed architecture for feature extraction and classification. Instead, the characteristics of the network layers adapt to the fault characteristics present in the bearing vibration signals. The classification accuracy of the proposed method has been evaluated using experimental data from two studied cases. The results demonstrate that the newly introduced network achieves higher accuracy in classifying bearing signals with different faults compared to similar methods.