Research On Fault Location And Diagnosis Of Multiple Bearings In Shafting Based On Deep Learning

Rolling bearings are widely used in machinery equipment,their health can often affect the equipment running status.In practical engineering applications,rolling bearings in shafting often exist in pairs of multiple installations.When the fault occurs,it is necessary to locate the fault bearing before further diagnosis and maintenance.Therefore,it is of great significance to locate the fault bearing in the shafting.In this paper,multiple rolling bearings under coaxial installation in shafting are taken as the research object,and a method of locating faulty bearings in shafting by using convolutional neural network on the basis of bearing samples which are expanded by generative adversarial networks is proposed.Then the diagnostic model is optimized with transfer learning,and its diagnostic accuracy is enhanced under the conditions of noise or variable load.This study first provides a data basis for the diagnosis model based on convolutional neural network.After filtering and denoising the bearing experimental signal,the two-dimensional time-frequency distribution of the experimental signal is obtained by Smooth Pseudo Wigner-Ville Distribution.The rolling bearing fault data samples are generated by the deep convolution generative adversarial networks embedded with gaussian mixture model,and the rolling bearing fault data set is expanded.The feasibility and effectiveness of this method to extend the fault data set are verified by the classification experiment of the generated fault data samples and real experimental data samples.Aiming at the fault location problem of rolling bearings in shafting,a fault location diagnosis method based on convolutional neural network is proposed on the basis of time-domain shock phenomenon and frequency-domain characteristic frequency analysis of bearing signals in shafting.The time-frequency processing is carried out for the rolling bearing signals at various positions in the shafting,then the time-frequency diagram is splited according to the position of the corresponding bearing in the shafting.And the resulting time-frequency diagram is used as the input of convolutional neural network for fault location diagnosis.The experimental results show that the method for locating fault in the shaft bearing has good accuracy.Based on the principle of transfer learning,a method of fine-tuning the pre-training network Alex Net is proposed to solve the problem that the fault signal of rolling bearing changes with the working condition,which results in the decrease of the accuracy of the diagnosis model based on convolutional neural network.The basic feature extraction layer in the early stage is retained,which accelerates the training speed in the early stage of the network.And the generalization ability and training efficiency of the model are improved by optimizing the activation function and back propagation algorithm in the high-level network.The original Alex Net and the fine-tuned Alex Net are tested under the experimental conditions of adding noise and variable load respectively.The experimental results show that compareing with the original Alex Net,the fine-tuned Alex Net has a higher diagnostic accuracy under the conditions of adding noise and variable load.It is verified that the Alex Net whose structure and parameters are fine-tuned has better generalization ability to locate and diagnose the fault bearing in the shafting under the experimental conditions of adding noise and variable load...