面向旋转机械迁移诊断的分层并行网络模型自动创建方法

In view of the current deep learning-based fault diagnosis methods of rotating machinery relying on manual modeling experience, requiring manual parameter adjustment and continuous trial-and-error iteration, re-creation of diagnosis models for different diagnostic tasks, an automatic creation method of hierarchical parallel network model for transfer diagnosis of rotating machinery is proposed, which can quickly and automatically search high precision diagnosis models with heterogeneous transfer performance according to different diagnosis tasks. Based on neural architecture search (NAS) and modular design ideas, two types of foundation blocks of parallel structure containing multiple layers are designed, which is different from the traditional NAS method to search layer by layer, but searches based on the foundation blocks. The controller outputs decision sequence to determine the foundation blocks’ structure and stacks them to form a hierarchical parallel candidate model. Then according to the verification results of the candidate model on the diagnosis task, the controller is optimized using the strategy gradient algorithm, and the diagnosis accuracy of the candidate model is continuously improved by iterating the above process. The hierarchical parallel structure of the candidate model supports its good heterogeneous transfer performance. In addition, in order to solve the time-consuming bottleneck problem of NAS method, the weight sharing mechanism is set in the candidate model training process to improve the efficiency of automatic modeling. The proposed method is used to conduct automatic modeling and heterogeneous transfer diagnosis experiments for four different rotating machinery fault datasets, and the results show that the proposed method can efficiently create 100% accuracy diagnosis models for four different diagnosis tasks, consuming 313 s to 1 601 s, and the candidate model can achieve a transfer diagnosis accuracy of more than 95% for the target diagnosis task with only 10% of the target domain data and 100 s of fine-tuning.