Fault Diagnosis Of Rotating Machinery Based On Deep Residual Network

As one of the important industrial infrastructures,rotating machinery is widely used in the production activities of various industries and plays a very important role in promoting the development of the national economy.Therefore,it is necessary and meaningful to carry out long-term condition monitoring of mechanical equipment and timely diagnosis when mechanical equipment fails.With the development of machinery and equipment towards large-scale,precise,intelligent and automated,the maintenance of the health of machinery and equipment has also entered the "big data" era.How to effectively carry out condition monitoring and fault diagnosis of machinery and equipment has become an essential system engineering.Traditional fault diagnosis techniques based on data-driven methods,whether based on signal analysis techniques or machine learning methods,require strong human experience and the participation of human engineering.However,in the face of increasingly complex mechanical equipment and with increasing data,the disadvantages of traditional methods have becomed increasingly prominent.In response to these problems,this paper applies a deep learning method with "end-to-end" learning characteristics to the field of fault diagnosis.Taking the rolling bearings and gears,which are important components of rotating machinery,as the research object,an intelligent fault diagnosis method based on the deep residual network model is proposed.The deep residual network is a deep neural network model with a new network structure.It proposes a residual connection structure based on the classic convolutional neural network.The network uses residual blocks as the basic unit to build a deep network model.The residual network solves the two major problems of traditional convolutional neural networks: the gradient vanishing problem and the deep network training degradation problem.In this paper,the deep residual network is introduced into the fault diagnosis of rotating machinery.In view of the working characteristics of rotating machinery often operating in constant speed,variable speed and strong noise interference situations,the following researches are carried out:(1)Aiming at rotating machinery and equipment under constant speed conditions,a bearing fault diagnosis method based on one-dimensional deep residual network model is proposed.This method directly uses the one-dimensional time-domain signal of bearing faults for “end-to-end” learning and training,allowing the model to directly learn from the time-domain signal and perform feature extraction,avoiding the participation of artificial feature engineering,and finally realizes the failure pattern recognition and classification diagnosis,and the effectiveness of the method is verified through the bearing test data set.The test results show that this method can achieve refined fault classification and identification capabilities for bearing faults with different damage levels.(2)Aiming at rotating machinery and equipment under variable speed conditions,a fault diagnosis method combining Synchroextracting Transform analysis and twodimensional residual network model is proposed.Because the vibration signal of rotating machinery under variable speed conditions has strong non-stationary characteristics,Synchroextracting Transform is a time-frequency analysis method that can effectively process non-stationary signals.Combining it with the deep residual network model can effectively ensure the stability and generalization performance of the model during fault diagnosis.At the same time,the validity of the model is verified by the rolling bearing fault data set under variable speed conditions,and the high-accuracy fault diagnosis of rolling bearing is realized.(3)Aiming at the situation of mechanical fault data lacking labels and industrial background noise in the actual industry,an unsupervised learning fault diagnosis model based on temporal residual denoising auto-encoder network is proposed.The model is mainly divided into two parts: an unsupervised training module and a classification finetuning module.The unsupervised training module is used to encode and decode the original vibration signal data.Through this training,an encoder with robust feature is obtained;The classification fine-tuning stage uses the encoder trained in the previous step and a small amount of labeled data to fine-tune the model to achieve fault classification and recognition.Finally,the test part verifies the model with the data collected by the planetary gearbox failure test bench,which proves the validity and reliability of the method.Finally,the research work of the full thesis is summarized and the research directions of the next step are forecasted.