Research On Gearbox Fault Diagnosis Based On Particle Swarm Optimization And Improved Deep Residual Networ

With the continuous improvement of modern mechanization levels,once mechanical equipment fails,it can seriously affect industrial production,causing casualties and economic losses.Therefore,it is of great practical significance to study the fault diagnosis method of mechanical equipment.Gear faults are a common cause of gearbox failure,and the characteristics of gear faults are often submerged by noise under complex working conditions.Traditional neural networks for fault diagnosis have problems such as difficulty in identifying fault features from noise,difficulty in obtaining fault data sets,and slow model convergence.In view of these problems,the main research work of this dissertation is as follows:(1)Wavelet transform:First,analyze the collected fault dataset,obtain the arithmetic time series according to the sampling frequency,then use Morlet wavelet to perform continuous wavelet transform on the input vibration signal,so as to obtain the fault timefrequency map as the input of the network.(2)Adaptive soft threshold denoising:Secondly,the adaptive soft thresholding subnetwork is added on the basis of the deep residual network(ResNet).Copy the input fault time-frequency map into two identical copies,one is used to learn the appropriate threshold,and the other is multiplied by the learned threshold to achieve the noise reduction effect.(3)Split Attention and Group Convolution:Due to the periodicity of the gearbox fault vibration signal,the fault features are often concentrated in the local area.Therefore,the introduction of the attention mechanism module enables the network to extract local fault features more effectively.Then,group convolution is used to optimize the convolution layer,so that the network can avoid a large increase in the number of model parameters while deepening the number of layers,thereby improving the computational performance of the model.(4)Transfer Learning:At the same time,the method of pre-train in Transfer Learning is also used to solve the problem of difficult collection of fault data sets in production and life,and compared with the original network,the pre-train model can converge faster.(5)Particle swarm optimization:The hyperparameters of the neural network have a decisive influence on the learning performance of the network.Traditional manual parameter adjustment is often difficult to achieve the optimal solution.This dissertation uses the particle swarm optimization method to obtain the optimal solution of the hyperparameters of the above-mentioned improved network,thus solving the limitations and inefficiencies of traditional manual parameter tuning.Finally,the fault time-frequency diagram is input into the improved network for fault classification.The experimental results show that the model based on particle swarm optimization and improved deep residual network can effectively identify and extract gear fault features,and the accuracy of identification and classification is high,reaching 99.7%.