Classification And Recognition Of Transmission Diagnosis Data Based On Deep Learning

As an important part of the vehicle,the transmission always operates under the complicated and diverse conditions,which makes it prone to failure.Therefore,before the transmission is delivered,the transmission is needed to detect by the on-line diagnostic system,which is the last stage to find transmission faults.This important role of the on-line detection system also requires it to have high fault recognition precision.Currently,the on-line detection system completes the task by the assistant of engineer,while people is subjective,fatigue,and poor consistency.Thus,It is significantly important to study the fault classification and identification of the transmission online detection system.The vibration signals of the transmission under non-steady conditions are nonlinear and non-stationary,while the deep neural network has better recognition capability for complex and nonlinear signals.Therefore,Deep convolutional neural network(CNN)and deep belief network(DBN)were used to solve the classification problem of transmission complex fault signals.The main research contents were as follows:(1)our datasets were based on the vibration data collected by the transmission online detection system.Firstly,the data were accurately labeled.And then the vibration data were analyzed by different signal processing methods,so the appropriate signal processing methods were selected according to certain criteria.Finally,the original vibration signal dataset,the time-frequency domain statistical dataset and the colormap diagrams data set were established.Each type of data set contained six kinds of labels,including OK,gear bumps,input shaft bumps,scream,shifting interference and compound faults.(2)In order to improve the classification accuracy of transmission datasets,the research of AlexNet,GoogLeNet and ResNet networks on the colormap diagrams data set was explored.And the classification of CNN based on residual module in the original vibration signal dataset and the DBN based on the time-frequency domain statistical dataset were also researched.By contrast,the deep neural network which was the best performance was selected.And the corresponding transmission data set were selected.(3)In order to further improve the accuracy of the classification model of transmission fault data set,we visualized the hidden layer of the network,analyzed the problem and proposed the optimization idea.Then multiple sets of optimization experiments were conducted.Finally we got the better model ResNetC-v2.(4)For the problem that the two kinds of faults vibration data were small,the method of transfer learning was adopted to make the ResNetC-v2 complete the classification of the above six data.And the experimental was carried out.Experiments show that:(1)Compared with the DBN network,the CNN network has better recognition ability for the transmission fault data set.The CNN network was used to complete the classification and identification of four categories in the colormap data set.The best network is ResNetC,and the best test accuracy is 92.62%.(2)Using the migration learning method,the network completed the learning of the six types of data.The final test accuracy is 92.3%.