Research On Bearing Fault Identification Method Based On Two-Dimensional Convolutional Neural Network

Rolling bearings are one of the key components of rotating machinery.When the bearings run at high speed,due to the complex and harsh working environment,the bearings are easily damaged,and their small size causes failures to be directly identified.How to effectively identify bearing faults has always been a problem in industry.The bearing fault recognition method based on signal processing relies on manual feature extraction and expert experience,and lacks adaptability in the complex and changing working environment,which cannot meet the needs of increasingly large and high-speed electromechanical data.Aiming at this problem,this paper studies the "end-to-end" method of bearing fault recognition based on two-dimensional convolutional neural network(2DCNN),which can automatically complete the process of feature extraction and fault recognition.The details are as follows:(1)Based on the 2DCNN algorithm,its classification-level structure of fully connected neural network is used for bearing fault recognition research.According to the characteristics of bearing vibration signals,a preprocessing method of two-dimensional image form samples is studied,and then the model parameters are designed and tested for training.The test results show that the fully connected neural network can directly achieve a high recognition rate when acting on the time-domain signal,but there is still room for improvement.On this basis,a 2DCNN model for bearing fault identification is designed.The optimal training parameters of the model were obtained through the crosscomparison experiment between the optimizer and the learning rate,and the structural parameters of the model were adjusted according to the deviation problems that occurred during the experiment.In addition,the Batch-Size and the total number of samples of the model training are optimized,the influence of the image resolution of the signal on the model feature extraction is studied,and the stability of the model is improved by the batch normalization algorithm.The experimental results of the test set show that the designed 2DCNN model can achieve a recognition rate of 100% on the CWRU bearing data set.(2)In view of the problem that the model recognition rate will decrease when the training data is disturbed under variable load and noisy environment,the 2DCNN model is improved.In order to enhance the feature mining capability of the model and reduce its computational burden,different combinations of filter size and the structure and number of fully connected layers are studied.In addition,an adaptive batch normalization algorithm is introduced to enhance the model’s adaptability in different sample fields,an improved moving average model is adopted and the Dropout algorithm is used to improve the generalization performance of the model.The integrated learning through the ensemble learning reduces the probability of error of the classifier,and finally the model’s anti-interference ability is tested by training in mini-batch.The test results show that the improved 2DCNN model has strong noise resistance and ability to resist load changes.(3)Aiming at the problem of fewer samples of measured bearing data,a 2DCNN model for bearing fault recognition based on transfer learning is designed.Taking the CWRU bearing data set as the source field and the measured bearing data as the target field,the 2DCNN model after transfer learning is used to perform engineering verification on the fault identification of the measured bearing data.The experimental results show that the model of transfer learning can accurately identify the fault type of the measured bearing data.