Research On Health Prediction Of Rotating Machinery Based On Deep Learning

Rotating machinery mainly through rotating action to complete specific functions,has been widely used in manufacturing,transportation,energy,aerospace and chemical industry and other industrial fields.Rotating machinery is the core operating component of various largescale high-end equipment.Due to the harsh operating environment and complex working conditions,rotating machinery is a failure-prone component.To ensure the long-term stable and safe operation of machinery and equipment,it is necessary to predict the health of rotating machinery to find the existing safety risks in time,so as to formulate maintenance strategies in advance to prevent serious accidents.Health prediction of rotating machinery refers to condition monitoring,fault diagnosis and remaining useful life prediction by using advanced sensors and various intelligent means.Deep learning can automatically extract features from massive high-dimensional data through superimposed multi-layer neural network and nonlinear transformation.Health prediction of rotating machinery based on deep learning has been rapidly developed.This dissertation utilizes deep learning as the technical means,focuses on the key issues in the field of rotating machinery health prediction,the supervised condition monitoring and fault diagnosis,unsupervised abnormal condition monitoring,domain adaptive unsupervised fault diagnosis and remaining useful life prediction are deeply studied.The main work completed and the research results obtained are listed as follows:(1)Aiming at the problem that the supervised condition monitoring and fault diagnosis models of rotating machinery in complex operating environments cannot effectively distinguish the importance of features information at different time points and sub-bands,a time-frequency attention residual network is proposed.The method takes the original vibration signal as input.The time-frequency coefficient matrix is obtained by short-time Fourier transform and then fed into the model.Using the time-frequency attention residual network to extract features and distinguish the importance of different time points and subbands,the proposed model can dynamically assign different weights to different time points and frequency bands of different samples.The model has the capability of distinguishing features,so as to effectively identify more important feature information.The experimental results of spindle rotation error monitoring and bearing fault diagnosis show that the proposed method can improve the prediction accuracy and has stronger anti-noise ability.(2)A multi-branch parallel attention convolution autoencoder is proposed to address the problems of difficult access to abnormal data and labels during abnormal condition monitoring of rotating machinery and insufficient feature extraction and reconstruction capability of existing autoencoder methods.The method takes the original vibration signal as input,and the frequency domain features are obtained by the fast Fourier transform and then fed into the model.Two parallel attention convolution autoencoder branches are used to extract feature information from different scales.The one-dimensional channel attention mechanism embedded after the convolution and deconvolution layers can adaptively calibrate the importance of the feature information of different channels.Euclidean distance and Pearson correlation distance loss function are used to measure the difference between input data and reconstructed output from two dimensions.The proposed method can effectively improve the ability of feature extraction and reconstruction,and can describe the similarity of complex feature spaces.The effectiveness and superiority of the proposed method is verified through bearing abnormal condition monitoring experiments.(3)Aiming at the problem of unsupervised fault diagnosis model of rotating machinery,which is difficult to extract domain-invariant features with the backbone network and a single training strategy,a multi-scale attention joint adaptive adversarial transfer model is proposed.The model consists of a shared multi-scale attention residual network as the feature extraction backbone module to learn more important multi-scale features from source and target domain data.In the model training,the joint distribution discrepancy of highdimensional features and labels between source domain and target domain data is considered comprehensively.The hybrid optimization strategy of joint maximum mean discrepancy and domain adversarial learning is adopted to effectively improve the domain confusion and reduce the domain distribution discrepancy.The proposed method can align the feature distribution of source domain and target domain while learning multi-scale features,and enhance the ability of the model to extract domain invariant features.The bearing unsupervised fault diagnosis experiment proves that the proposed method can improve the diagnosis accuracy.(4)Aiming at the problem of large number of parameters and difficulty in accurately predicting the remaining useful life of rotating machinery under different degradation modes through multi-scale convolution models,a lightweight multi-scale attention residual network based on feature reuse is proposed.The method first constructs a one-dimensional crosschannel maximum pooling layer to focus on the more important feature information.Then,based on the idea of grouping convolution,the feature reuse unit with hierarchy structure is constructed by reusing the feature information of the previous group.After the feature reuse unit,a one-dimensional efficient channel attention is constructed to highlight the important feature information.One-dimensional cross-channel maximum pooling layer,feature reuse unit and one-dimensional efficient channel attention unit are all extremely lightweight modules.The model can significantly reduce the number of parameters,which is conducive to easing the burden of parameter optimization.Through the experiments on bearing remaining useful life prediction,it is demonstrated that the proposed method can accurately capture the degradation trend of bearings and improve the model prediction performance while reducing the number of parameters.In this dissertation,the health prediction methods of rotating machinery are studied.The proposed methods effectively improve the prediction performance of rotating machinery in supervised condition monitoring and fault diagnosis,unsupervised abnormal condition monitoring,unsupervised fault diagnosis,remaining useful life prediction,and provide a new means and ideas for health prediction of rotating machinery.