Research On Bearing Fault Diagnosis Method Based On Semi-supervised Learning Under Data Imbalance

As an important component of machinery,the good working condition of bearings is a prerequisite to ensure the normal operation of the machine.As the bearings are often in bad and complicated working conditions,they are prone to failure,which may affect the working performance of the machine,or paralyze the whole machine and cause huge economic loss.Therefore,the bearing working condition monitoring and fault diagnosis technology is of great engineering significance for the safe and stable operation of the machine to avoid economic losses.In the actual industrial production process,the bearings are in normal operation most of the time,and the failure cycle is short.Therefore,the normal data collected by sensors is much more than the fault data,which leads to the data imbalance problem of bearing fault diagnosis.How to carry out effective fault diagnosis under the condition of imbalanced data of bearing fault samples is a challenge for the current data-based fault diagnosis method.In order to solve the data imbalance problem,the main research work of this paper is as follows:Firstly,we gain insight into the current status of fault diagnosis research under data imbalance,and describe the fault types of rolling bearings,common deep learning networks and evaluation metrics applicable to imbalanced data.Then,a deep convolutional self-encoder bearing fault diagnosis method based on semi-supervised learning is investigated.The input of the method is a one-dimensional time-domain vibration signal,and the feature extractor is a convolutional self-encoder encoder.A bidirectional long-and short-term memory prediction network with an attention mechanism is introduced into the convolutional self-encoder,and the prediction error is used as a penalty term to avoid classifier bias toward large classes.In the training unlabeled data phase,the nearest class prime pseudo-labeling method is used to enhance the generalization of the model using unlabeled data.The nearest class mass center pseudo-labeling method records class mass centers by introducing a storage unit,labels pseudo-labels based on the distance metric from the class center,and updates the class centers using an exponential moving average method as the pseudo-labeled data increases.The performance of the proposed model is validated using bearing failure datasets from Case Western Reserve University,USA and University of Paderborn,Germany.The validation results show that the proposed method has stronger generalization ability and higher accuracy than typical data processing models.Finally,a semi-supervised learning-based deep convolutional generative adversarial network rolling bearing fault diagnosis method is investigated.The one-dimensional time-domain vibration signal is converted into a two-dimensional frequency-domain grayscale map as the input of the network,and the SELU activation function and GN normalization function are added to the generator on the basis of the deep convolutional generative adversarial network,and the discriminator is changed from binary classification to multi-classification,and the semi-supervised learning is realized by combining the nearest class prime pseudo-labeling method.The results show that the semi-supervised DCGAN has good generalization ability and high accuracy diagnosis performance when dealing with data imbalance problems.The quality of the generator output data,the effect of the degree of data imbalance on the network accuracy,and the superiority of the improved pseudo-labeling method are further analyzed.