Lightweight two-stream heterogeneous neural network approach for bearing fault diagnosis

In order to solve the problem of high resource consumption and insufficient feature sensitivity of traditional two-stream neural network in bearing fault diagnosis due to the redundancy of isomorphic algorithms, this paper proposes a lightweight two-stream heterogeneous neural network method (LQCKAN). This method realizes multi-dimensional information complementation through heterogeneous feature extraction modules: the time domain branch uses a quadratic convolutional neural network (QCNN) to capture the nonlinear features in the time series signal, and the time-frequency domain branch uses a convolutional Kolmogorov-Arnold network (CKAN) to extract the complex spatial features of the time-frequency graph. Furthermore, the model is lightweighted by dynamically pruning redundant weights and neurons, which reduces resource consumption while ensuring performance. Experiments on the CWRU public dataset and the self-built bearing fault dataset show that the fault diagnosis accuracy of LQCKAN reaches 99.99% and 98.87% respectively, and the resource consumption is reduced by 20.88% and 26.87% after lightweight processing, which is significantly better than the mainstream deep learning model.