Research On Small Sample Fault Identification Method Of Transmission Based On Deep Separable Convolution Capsule Network

Due to the different power sources of new energy vehicles and traditional vehicles,new energy vehicle transmissions have some new characteristics compared to traditional transmissions.The new energy vehicle transmission is directly driven by the motor,with high precision,high speed,large instantaneous acceleration,and more severe gear and bearing working conditions.If the transmission has a shock fault,when the transmission operation instantaneous acceleration increases,it will enhance the source of the fault impact components,causing the transmission noise,or even lead to its unstable power output,failure and other problems.Therefore,it is important to identify faults in critical parts of the transmission.However,the transmission installation location is compact,the multi-sensor arrangement is difficult,the shock fault is accompanied by shock modulated signal generation and accompanied by high noise,and the acquired sample data components are complex.Therefore,it is important to study the transmission fault identification method under small sample for transmission fault early knowledge and fault location.In the paper,we study on the theory of convolutional neural network and capsule network,and improve the capsule network by using depth-separable convolutional method and residual connection to construct depth-separable convolutional capsule network model to solve the problem of fault recognition accuracy of new energy transmission under small samples.Simulation signals are used to determine the optimal configuration of this network parameters,open-source data from bearings are applied to verify feasibility,and actual data from a transmission failure test bench are used for intelligent fault diagnosis.The details of the study are as follows:（1）By analyzing the transmission gear and bearing vibration response mechanism,a simulation signal containing shock fault components is established.Since the transmission contains fault parts with shock components,a transmission fault acquisition test bench is built and the transmission fault signal components are studied by analyzing the simulation signals.（2）The parameters of the Capsule Network are optimally configured using the simulation signals.Different learning rate values,convolutional kernel sizes,batch sizes,and dynamic routing iterations are used to analyze the accuracy of each parameter under the Capsule Network,and the results of the analysis are used to determine the network model structure.Case Western Reserve University bearing open source data was used to verify the feasibility of the optimized capsule network.Subsequently,gear,bearing fault data collected by the transmission fault acquisition testbed was used to improve transmission fault diagnosis accuracy.（3）In order to solve the problem of weak generalization ability of Capsule Network under small samples,the convolutional layer of Capsule Network is improved into a depthseparable convolutional layer.A multi-scale convolutional decomposition is applied in the depth-separable convolutional layer to form a depth-separable convolutional capsule network through residual connectivity.The performance of the Depth-Separable Convolutional Capsule Network with different residual blocks is analyzed to determine the number of residual blocks and to obtain the optimal network model structure.The DepthSeparable Convolutional Capsule Network model is used for fault identification of small sample transmission faults,and good identification results are obtained.This paper focuses on the capsule network with its capsule network convolutional layer improvement type to simulate the transmission small sample fault data,and to demonstrate the feasibility of transmission faults in small samples and improve the fault recognition accuracy through experiments.