Abstract
Deep learning network models are widely applied to fault diagnosis of planetary gearboxes. However, the multi-coupling fault characteristics, accompanied by data fuzziness and distribution differences, present certain challenges to diagnostic research. Under variable operating conditions, the fault data to be diagnosed becomes more prominently inconsistent in distribution, leading to suboptimal fault recognition rates in diagnostic models. A deep transfer learning method for planetary gearbox fault diagnosis based on a Selective Kernel Networks (SKN) attention mechanism is proposed. First, an input dataset is constructed through overlapping sampling, and a deep neural network diagnostic model is established to automatically learn features and perform diagnostics on the data. Second, a dynamic selection mechanism for convolution kernels is embedded in the deep neural network, enabling each neuron to adaptively adjust its receptive field size based on multi-scale input information. This mechanism extracts common features between the source and target domains, enhancing the network's feature extraction capability. Third, the Local Maximum Mean Discrepancy (LMMD) is used to perform sub-domain adaptation on the features of the source and target domains, reducing the distribution discrepancy between the two domains and constructing an end-to-end transfer adaptation model. This enables deep transfer learning fault diagnosis of planetary gearboxes under varying operating conditions. Finally, through experimental analysis and validation of 8 variable operating condition tasks, the fault identification accuracy of the diagnostic method proposed in this paper reached an average of 92.9%. Compared with traditional deep transfer learning diagnostic methods, it demonstrates higher diagnostic precision.