Research On The Fault Diagnosis Method Of Rolling Bearing Based On Deep Learning

With the continuous development of modern industry,industrial equipments are becoming more complicated and more systematic,which put more requirements on operating and monitoring of industrial systems.Full-time monitoring at all aspects is the key to the safe and stable operation of industrial systems.However,the large amount of monitoring data produeed in monitoring proeess makes traditional methods based on manual analysis and expertise difficult to accomplish diagnosis work,and the structure complexity and non-stationary characteristics of the equipments operation state bring difficulties to the diagnosis and analysis of equipments.Rotaring machinery is a common equipment in modern industry.In this thesis,rolling bearing is chosen as the research object,which is the core component of rotating machinery.According to the data distribution difference of rolling bearing vibration data,bearing fault diagnosis is divided into balanced datasets diagnosis and unbalanced datasets diagnosis.The balanced datasets contains fault data and normal data with uniform distribution and equal quantity.This thesis systematically studies rolling bearing fault diagnosis under different data scenarios through both theoretical research and experimental verification,and proposes fault diagnosis algorithms based on deep learning.The specific research contents include the following three aspects.(1)A diagnostic model based on residual network and attention mechanism is proposed for fault diagnosis under balanced datasets.The diagnosis model takes one-dimensional vibration time series signal as input,completes feature extraction through residual network,and then achieves feature expression in time domain through Bi-LSTM(Bidirectioanl Long Short-Term Memory)unit with attention mechanism,assigns different weights,and outputs them to classifier to complete the end-to-end vibration signal diagnosis.Experiments show that the diagnostic accuracy of the proposed model is over 99.86%,and the diagnostic accuracy of each fault type is over 99%,and the extracted feature information is highly distinguished.The diagnostic accuracy of the model is superior to the diagnostic model based on feature engineering.Compared with the model based on deep learning,the model has better stability.(2)Changing working conditions in bearing operation are addressed in this thesis,and a fault diagnosis model of rolling bearings changing working conditions based on domain adverasial is proposed.The model mainly includes feature extractor,fault classifier and domain classifier.The feature extractor is composed of residual network and Bi-LSTM network,which is used to extract the characteristics of vibration signals;the fault classifier is responsible for the state classification of vibration signals;and the domain classifier is responsible for distinguishing whether the signals come from the source domain or the target domain.By adding gradient flip layer between classifiers to construct domain migration network,domain migration is used to complete domain adaptive work,and then labeled source domain data sets are used to realize the diagnosis and recognition of unlabeled target domain data sets under variable working conditions,and the fault diagnosis under variable working conditions is completed.Experiments show that the model extracts features suitable for migration between different working conditions and improves the diagnostic performance of various fault types under variable working conditions.The average diagnostic accuracy under chaing working conditions can reach 97.42%,which is better than the direct cross-domain diagnostic models.(3)The imbalance problem between fault data and normal data is addressed,which can lead to the undetection of fault data while the fault data belong to minority class is misclassified as majority class sample.This thesis proposes a model for imbalance fault diagnosis based on residual network and XGBoost(extreme gradient boosting)by improving the balanced diagnosis above and introducing the ensemble learning method for imbalanced data classification for fault diagnosis.The model retains residual network as feature extractor of vibration signals,and inputs the extracted feature into XGBoost for fault classification.Experiments show that the model can accomplish the fault diagnosis task well under different fault-normal data ratios with high fault diagnosis accuracy.