Research On Bearing Fault Diagnosis Method Based On Adaptive Convolutional Neural Network

Rolling bearings play an irreplaceable role as the core components of rotating machinery.Due to the continuous improvement of production requirements in the mechanical industry,bearings often work in complex environments,which are prone to failure in the long run.If the fault cannot be removed in time,it may cause major safety accidents and lead to inestimable economic losses.Therefore,it is necessary to realize accurate and efficient bearing fault diagnosis.Compared with traditional fault diagnosis methods that rely on time-domain and frequency-domain characteristics for fault analysis,fault diagnosis methods based on deep learning can achieve better fault diagnosis results by relying on the powerful learning ability of the neural network.However,the transfer learning ability to existing neural networks is weak,so it may only be suitable for a specific class of bearings to achieve fault diagnosis.If the fault model needs to realize the fault diagnosis task of different types of bearing data,the structure of the neural network model needs to be adjusted,and the optimal model structure can be determined after several experiments,or the model can be applied to different bearing data by means of transfer learning,intelligent optimization algorithm,etc.,Still,the use of such methods will also increase the computational complexity of the model and the overall training time.To solve the above problems,a fault diagnosis method based on the Adaptive Depthwise Separable Dilated Convolution and Multi-grained Cascade Forest model(ADSD-gc Forest)is proposed to achieve fault diagnosis with different bearing data.Firstly,vibration signals are converted into Symmetry Dot Pattern(SDP)images.Three dilated convolutions with different dilation rates are combined with convolution attention mechanisms to achieve multi-scale feature extraction of fault samples.Meta-Activate or Not(Meta-ACON)activation function is used as the activation function of all convolutional layers.The output mode of the neural network of this layer can be adjusted by Meta-ACON function according to the current input samples to realize the adaptive optimization of the model structure based on different bearing data.Then the gc Forest is introduced to sample the output of the final fully connected layer again to output the final fault diagnosis results.The experimental results show that the model can better complete the fault diagnosis task of different bearings under complex working conditions.Due to the lack of theoretical guidance in the design of structural parameters of the fault diagnosis model,it is necessary to rely on a large number of experiments and continuous trial and error.At the same time,there are few fault sample data,it is rare to explore solutions to the problem of unbalanced sample distribution from within the model.To solve this problem,a diagnostic model based on the Dynamic Adaptive Structural Parameter Optimization Model(Dy-ASPO)is proposed.Inception Convolution(IC-Conv)is used to construct the trunk network.By combining pre-training supernet with statistical optimization,appropriate dilation modes of the convolutional layers can be selected.By estimating the contribution of each layer of convolution to loss reduction,Dynamic diverse branch block(Dydbb)is introduced to expand the convolution layer with the largest contribution to reducing losses.Before the final classifier,Batch Attention is applied to make the model focus on the relationships between samples of different categories during training.Thus,the problem of the unbalanced distribution of sample categories can be effectively addressed.Experimental results show that the fault diagnosis method can select appropriate convolutional neural network parameters according to different bearing data,and carry out targeted multi-branch structure enhancement,and the optimized network can better realize bearing fault diagnosis under the condition of unbalanced sample distribution.