Rotating Machinery Fault Diagnosis Based On Time-frequency Representation And Capsule Network

Research on fault diagnosis of rotating machinery has important practical significance for reducing maintenance cost and improving safety of industrial production.Since deep learning has the ability to automatically extract fault features from massive data,it can overcome the defect of manual feature extraction to some extent.However,traditional deep learning-based fault diagnosis methods still have certain limitations: 1)the background noise in the original vibration signal will interfere with the feature learning performance of the model;2)the problem of data imbalance will reduce the diagnostic reliability of the types with small samples;3)the problem of variable working conditions will lead to aliasing in the feature space of different types of samples,which will affect the accuracy and generalization performance of diagnostic model.Aiming at the above problems,this paper takes the key components of rotating machinery as the research object,takes time-frequency representation of vibration monitoring signals as the entry point,combines the theoretical methods of signal decomposition,deep learning and multi-sensor information fusion,and focuses on the research of rotating machinery fault diagnosis under data imbalance and variable working conditions.The main works of this paper are as follows:Aiming at the problem that the feature learning performance of classification model is easily affected by the noise characteristics of original vibration signals,a fault diagnosis method based on frequency slicing wavelet transform(FSWT)and capsule network is proposed.Firstly,the energy ratio index is used to select the observation frequency band of FSWT adaptively,and the time-frequency features of the original vibration signals are extracted and further stored as two-dimensional time-frequency images;then,combined with multi-scale network and channel attention mechanism(CAM)network,a multi-scale feature enhancement capsule network(Ms FE-Caps Net)model is proposed,which provide method support for the subsequent research on fault diagnosis under imbalanced data and variable working conditions.Aiming at the problem that the data imbalance will affect the reliability of fault diagnosis,a fault data augmentation method based on multi-scale residual generative adversarial networks(Ms R-GAN)is proposed.Firstly,aiming at the training instability and gradient disappearance of the original generative adversarial networks,a Ms R-GAN model is proposed based on residual network and hybrid optimization loss;then,the trained Ms R-GAN model is used to balance the distribution among different types of time-frequency image dataset;finally,the fault pattern recognition is realized combining with Ms FE-Caps Net model,which effectively solves the problem of fault diagnosis under imbalanced data.Aiming at the problem that variable working conditions will reduce the accuracy and generalization performance of diagnostic model,a fault diagnosis method based on dynamic weighted Ms FE-Caps Net is proposed.Firstly,the original data from different sensors are converted into time-frequency images by FSWT and concatenated at the channel level;then,trainable normalized weights are assigned to different channels through CAM network to achieve adaptive data fusion,and to reduce the information redundancy among multiple sensors meanwhile;finally,the fused data is input into the Ms FE-Caps Net model for fault pattern recognition,which effectively improves the accuracy and generalization performance of diagnostic model under variable working conditions.