Research On Early Fault Intelligent Diagnosis Of Engine Bearing Based On Improved CNN

Rolling bearing is an important supporting component of civil aviation engine rotor system.Because of the bad working environment of the bearing,the background noise of the vibration signal of the rolling bearing in the actual operation which leads to the low signal-to-noise ratio(SNR)of the measured signal.In this study,a diagnosis model combining signal processing method and convolution neural network is used to diagnose the low SNR signal and early vibration signal of rolling bearing.The main work of this paper is as follows:1.Research on the bearing testing equipment and process,select the fault type to be simulated in the bearing test,build the rolling bearing test bench according to the test requirements,determine the bearing model and size parameters used in the bearing test,and determine the vibration signal acquisition conditions.After testing the processing speed of different signal analysis methods,the traditional vibration signal analysis method,short-time Fourier transform and continuous wavelet transform with short processing time and strong universality are adopted to process the bearing signal of CWRU.The applicability of each method is analyzed.In order to simulate the early fault signal with low signal-to-noise ratio,high-strength noise was added to the bearing vibration data of CWRU,and the traditional signal analysis method was used to process the data.Continuous wavelet transform(CWT)is used to process the self-collected signals,and a low SNR,multi class fault diagnosis data set is established.2.In order to realize the diagnosis of early rolling bearing fault signal with low signal-tonoise ratio,a rolling bearing fault diagnosis method based on one-dimensional convolution neural network is proposed by combining the characteristics of time-domain signal and frequency-domain signal as one-dimensional vector.The control variable method is used to optimize the one-dimensional diagnosis model from three parameters: the size of convolution kernel,the number of convolution kernel and the proportion of lost output.The relatively optimal structural parameters are obtained.And the diagnosis effect of the model is verified on the open bearing data set.The results show that the diagnosis accuracy of the frequency domain signal after windowed Fourier transform is 99.2%,which is better than the original time domain signal model.Compared with the traditional intelligent diagnosis models such as BPNN and DNN,the diagnosis accuracy of BPNN is about 70% and that of DNN is about 30%,which shows that the convolutional neural network model has obvious advantages in fault diagnosis.3.It is found that the diagnosis accuracy can only reach about 80% when testing on selfcollected bearing data,which indicates that single time domain or frequency domain signal has limitations in fault diagnosis task.Therefore,in order to realize bearing fault diagnosis under small sample,multi-channel and multi fault types,a fault diagnosis method based on wavelet time-frequency diagram and two-dimensional convolution neural network is proposed,and then the diagnosis results of each type of fault in 2dcnn model are analyzed by using confusion matrix.In order to improve the feature extraction ability of the model,the vector capsule layer is used to optimize the two-dimensional convolution neural network,which is compared with the deep two-dimensional convolution diagnosis model.Because the traditional convolutional neural network is hierarchical structure,it is unable to extract multi-scale features.Therefore,the capsule layer of matrix is added in parallel for depth feature extraction,and the adaptive feature fusion method is used to fuse the extracted multi-scale deep features.The model is verified on the self-collected bearing data set,and the diagnosis accuracy reaches 98.3%.The diagnosis results are analyzed by the recall rate and F1 value.Then the bearing data under different working conditions are used as the training set and test set of the diagnosis model.The diagnosis accuracy of CNN＿Caps S can reach about 94%,which also shows that the model has strong generalization ability and robustness.