Time-Frequency Information Multilevel Integration and Multiscale Parallel-Fusion Network for Bearing Fault Diagnosis

The frequency-domain features of the vibration signal contain rich fault information. However, how to extract multidimensional features from it to achieve failure and complicated conditions is still a key challenge in the field. This article proposes a fault diagnostic network based on parallel modules, which, respectively, build local feature capture modules and global dependence to build modules. Local feature capture module by converting the Fourier transform (FFT) signal into a time-frequency diagram, using a 2-D convolutional network to extract the local time-frequency characteristics. The global dependence to build modules block directly performs time convolution modeling on FFT signals and combines the constructed adjacency matrix to input the graph convolutional network to capture the global correlation of frequency-domain signals. Finally, the features extracted by the parallel module are fused to generate a joint feature representation, which is then input into the Softmax classification layer to complete the classification task. The experiment was validated on the bearing fault datasets of Beijing Jiaotong University and Huazhong University of Science and Technology, achieving high-precision fault recognition for 13 and 12 categories, respectively, with diagnostic accuracies of 99.40% and 99.91%, respectively. The results indicate that the proposed network exhibits excellent classification performance and robustness in complex fault modes. This study provides a new method for efficient and accurate fault diagnosis, laying the foundation for health monitoring and intelligent maintenance of complex rotating machinery systems.