Convolutional Neural networks based on parallel multi-scale pooling branch: A transfer diagnosis method for mechanical vibrational signal with less computational cost

Traditional fault diagnosis models based on machine learning technology are difficult to apply to data samples under different working conditions. In a working environment that can only provide less computational resources, the parameter scale of algorithms is restricted, and the difficulty of transfer diagnosis is further increased. To this end, this paper proposes a transfer diagnosis method based on PMSPB-CNN (Convolutional Neural Networks Based on Parallel Multi-scale Pooling Branch) to solve the mechanical vibrational signal fault diagnosis problem under multiple working conditions with less computational cost. PMSPB-CNN introduces a parallel multi-scale pooling branch (PMSPB) structure to replace the basic convolution module used in traditional 1D-CNN. The multiple parallel paths in PMSPB structure contain the pooling layers with different scales and pooling methods to mine high-level features with different granularities. There are no network parameters that need to be trained in this structure, which greatly saves computational resources and reduces the risk of overfitting. Based on the transfer learning strategy of freezing the pretrained feature mining unit and fine-tuning the parameters of the fault identification unit, PMSPB-CNN can perform high-accuracy fault diagnosis on similar fault samples under multiple working conditions. The experimental results show that the parameter number of PMSPB-CNN is 774, and the number of parameters that need to be re-optimized for transfer diagnose is only 360. However, compared with the existing methods, even if the pre-trained network is directly used to diagnose faults under the other working conditions, the accuracy of PMSPB-CNN still maintain a high level on the two verification datasets, reaching 73.2% and 97.8% respectively. After fine-tuning the fault identification unit, PMSPBCNN can achieve the 100% transfer diagnosis accuracy. In addition, the mechanism analysis experimental results show that when dealing with the data under different working conditions, the pooling layer in PMSPB structure with the best classification performance is not exactly the same. Furthermore, the output features of the frozen feature mining unit already highly recognizable before fine-tuning the fault identification unit. These conclusions proved that PMSPB structure provides sufficient fault tolerance and flexibility for the network, thereby improving the generalization of PMSPB-CNN.