Research On Rolling Bearing Fault Diagnosis Based On CNN

Rolling bearings are the core components of all kinds of machinery and equipment.In the industrial field,about 30% of the faults of machinery and equipment are caused by bearing damage.Whether the state of the bearing is normal or not is directly related to the operation of the equipment.The core of bearing fault diagnosis is to identify the bearing fault location and damage size according to the signal characteristics by analyzing the bearing vibration signal.The traditional solution is to analyze and process the bearing fault signal in time-frequency domain,which requires a lot of basic professional knowledge to assist in feature extraction,which will cause difficulty in feature extraction and low accuracy of fault diagnosis.With the rise of convolutional neural network,it can maximize the feature information extraction of the signal,so this paper uses convolutional neural network to identify and diagnose rolling bearing faults.The following is the main research content of this paper:(1)Research on signal preprocessing of rolling bearing and traditional bearing fault identification method.The effect of Fourier transform,wavelet transform,empirical mode decomposition and other methods in the pretreatment of rolling bearing vibration signal is studied and analyzed,and the advantages and disadvantages of traditional fault diagnosis methods such as wavelet transform,wavelet packet transform,blind source separation,and Hilbert-Huang transform in the fault recognition of rolling bearing are analyzed.(2)The one-dimensional signal is converted into two-dimensional vibration image.Aiming at the problem that the spatial characteristics of one-dimensional signal are not obvious and the spatial feature extraction is difficult,several methods of converting time series signals into two-dimensional vibration image matrix are studied and analyzed.The data set is processed by using the Gram matrix method,and the vibration image features are used to express the bearing fault information,so as to improve the feature information carried by the bearing vibration signal.(3)Bearing fault recognition based on convolutional neural network is studied.A multichannel feature extraction network structure is designed for the feature map generated by the time series signal,which broads the network and adds convolution kernels of different scales to improve the adaptability of the network model for bearing problems,so as to improve the efficiency of the model in extracting effective vibration image feature information,and increase the accuracy and stability of the model for bearing fault recognition.In conclusion,on the bearing data set of Case Western Reserve University in the United States,the bearing fault recognition method proposed in this paper successfully achieves 99.86%fault signal recognition accuracy,and the average accuracy of fault type location under different working conditions is 98.65%.And for the bearing signal with 30 db and 25 db Gaussian white noise,the fault-free recognition accuracy is 98.45%and 96.61%respectively,which is in line with expectations,indicating the advancement of the convolutional neural network in the fault diagnosis and recognition of rolling bearings.