Research On Fault Recongnot Method Of Rolling Bearing Based On Convolution Neural Network

Rolling bearing is one of the key to ensure the healthy operation of mechanical equipment,and rolling bearing faults occurs frequently in industrial field.Therefore,it is of great significance to carry out the research of efficient and accurate rolling bearing condition monitoring and fault diagnosis methods and timely find the abnormal and diagnose for ensuring the safety of mechanical equipment production.The traditional diagnosis methods rely on expert diagnosis knowledge and experience.The diagnosis methods based on deep learning gets rid of the dependence on expert diagnosis knowledge to a certain extent,but it relies heavily on the labeled fault data training network which seriously restricts the practicability of its industrial field.At the same time,in the massive big data environment,the diagnosis performance needs to be further improved.In view of the above problems,combining with convolutional neural network(CNN),this paper carries out the research of rolling bearing fault recognition method.Determining the best input of CNN through analysis and comparison different input feature,improving the supervised-CNN model to improving the model diagnosis performance,and proposing semi-supervised and unsupervised transfer bearing fault recognition method based on CNN to reducing the dependence of network training on labeled data:Firstly,analyzing the commonly used input features of rolling bearing fault recognition method based on CNN.The performance of the rolling bearing fault recognition model based on CNN under different input features is compared from the parameters of the model to be learned,the training time and recognition accuracy to determine the best input feature of CNN for rolling bearing fault recognition.The test data of rolling bearing show that the best input feature is spectrum amplitude sequence.Secondly,based on the best input feature of CNN for rolling bearing fault recognition,the CNN model is improved from the aspects of activation function,loss function,network structure,etc,and the performance of rolling bearing fault recognition model based on supervised-training CNN is improved.The rolling bearing test data is used to verify the effectiveness of the rolling bearing fault recognition method based on improved CNN.Then,based on the idea of self-training in semi-supervised learning,using a small number of and a small category of labeled data pre-training network.Combining the kmeans clustering method and pre-training network to determine the pseudo label of unlabeled data,and based on the labeled data and pseudo label data to jointly train the network to determine the best pseudo label of unlabeled data.In other words,the pseudo label is regarded as the final label of unlabeled data,so as to identify the fault category of unlabeled data and reduce the dependence of model training on the labeled fault data.The rolling bearing test data and intershaft bearing test data verify the effectiveness of the semi-supervised fault recognition method of rolling bearing based on CNN with few fault samples.Finally,based on the idea of domain adaptation,the research of transfering between different working conditions and different test bench is studied.By adapting of two-domain features with CNN,the transfering between the two domain label is realized,and solve the problem that the target domain unlabel data can not train the network.The effectiveness of the unsupervised fault recognition method based on CNN and transfer learning is verified by using the test data of common rolling bearing and intershaft bearing.