Research On Gearbox Fault Diagnosis Based On Signal Processing And Improved Residual Network

The gearbox is an indispensable key component in the production and operation of rotating machinery.Long-term operation will lead to damage to the components in the gearbox and cause various accidents.Therefore,the diagnosis of the fault type of the gearbox is imminent.Due to the complex mechanical structure inside the gearbox,it is difficult to analyze the vibration signal and frequency spectrum signal.In actual working conditions,strong noise and variable load conditions make it more difficult to diagnose the fault types of parts in the gearbox.In this paper,based on the combination of intelligent signal processing and improved neural network model,the diagnosis error is effectively reduced,and the accuracy of fault diagnosis under various working conditions is improved while ensuring the training accuracy.The main content of the thesis is as follows:First,aiming at the problem of nonlinear characteristics of gearbox fault signals,a method combining two-stream dimensionality reduction Ensemble Empirical Mode Decomposition(EEMD)and Convolutional Neural Network(CNN)is proposed.First,the EEMD algorithm is used to decompose,screen and dimensionally reduce the collected signals,and then use the convolutional neural network to perform model training on the screened intrinsic mode components(IMF)and original signals,and output the final accuracy of the model.It is verified by experiments that this method can achieve more accurate fault diagnosis when training both noise-free data and noise-containing data.Compared with other traditional network models,the model proposed in this chapter has higher accuracy in fault diagnosis results.Secondly,in view of the modal aliasing phenomenon in the process of decomposing the vibration signal in the EEMD algorithm,and the occurrence of endpoint effects cannot be avoided,a method using variational mode decomposition(VMD)and improved one-dimensional residual convolutional neural network failure is proposed.diagnosis method.By using the Krill Swarm Algorithm(KHA)to optimize the hyperparameters in the VMD algorithm,the influence of artificial experience settings is reduced.By decomposing and screening the vibration si gnal,the filtered IMF components that meet the requirements are input into the r esidual network model with an adaptive threshold,and feature extraction and classification are performed using the network model with a residual shrinkage module.Through experimental research,the robustness and generalization of the proposed method are verified.Finally,while satisfying that the model can fully extract fault features and further reduce training time,a gearbox fault diagnosis method with lightweight network combined with intelligent signal processing and inverted residual shrinkage module is proposed.The one-dimensional IMF component screened out by VMD is used as the input of the improved lightweight network module through two-dimensional matrix transformation,and the inverted residual shrinkage module built by depth-separable convolution is used to ensure high fault diagnosis accuracy,effectively reducing the model training time.The experimental results show that the proposed method has certain application value in the actual working environment.