Fault Diagnosis For Imbalanced Data Based On Time-frequency Sample Generation And Application In Bearings

With the development of the mechanical system towards the direction of intelligence,a large number of state signals can be collected for mechanical systems,which are used to reflect the health of the machinery and to diagnose faults on the machinery.Mechanical equipment is usually working in a normal state,and it is difficult for sensors to collect fault state signals.Only in a few cases can equipment collect fault state signals when the mechanical equipment fails.This results in the problem of data imbalance in fault diagnosis,i.e.,the normal state signals are far exceeded than the fault state signals.In this case,the performance of the fault diagnosis model will significantly decrease,leading to misdiagnosis and missed diagnosis.As a hot data generation model in the field of deep learning,the generative adversarial network provides a new solution for data imbalance.This article mainly includes the following contents:Aiming at the problem of bearing data imbalance causing the performance degradation of fault diagnosis model,a fault diagnosis method based on time-frequency images and deep convolutional generative adversarial network(DCGAN)is designed to improve the performance of fault diagnosis model classifiers.Continuous wavelet transform is employed to convert the original vibration signal into time-frequency images,and then the time-frequency images are inputted into DCGAN for training to generate pseudo-fault time-frequency images.Finally,the generated fault time-frequency images are added to the original imbalanced data set for data augmentation,so that the imbalanced data set reaches a balanced state,and the classifier is trained to achieve fault diagnosis.Results in the fault diagnosis for the bearing dataset of CWRU University show that the accuracy of the diagnosis is improved by the proposed method.Aiming at the problem of training failure caused by the unstable training of the generative adversarial network,based on the spectral normalization theory and the gradient penalty term,an improved fault diagnosis method of the generative adversarial network is proposed.First of all,spectral normalization is introduced in the network structure to replace the original batch normalization and constrained the Lipschitz constant of the discriminator by constraining the spectral norm of the network weights of each layer to achieve stable training.Effect: Secondly,a gradient penalty term is introduced,and the objective function of the generative adversarial network is modified to replace the original gradient clipping.Results in the fault diagnosis for the bearing dataset of Paderborn University show that the accuracy of the diagnosis is improved by the proposed method.Aiming at the problem that the generative adversarial network cannot understand the internal relationships of images and the diagnosis accuracy is not improved due to the low quality of the fault time-frequency picture generation,based on the self-attention mechanism,a self-attention mechanism generative adversarial network fault diagnosis method is proposed.The self-attention mechanism module is introduced into the network framework to focus on the internal relationship of the fault time-frequency images so that the generative adversarial network can effectively model the relationship between the distant regions inside the images.Results in the fault diagnosis for the bearing dataset of Paderborn University show that the accuracy of the diagnosis is improved by the proposed method.