Research Of Fault Diagnosis Method Based On Transfer Learning

The healthy and stable operation of key components of electromechanical equipment is the guarantee of high quality and efficient production in intelligent manufacturing process.The research of intelligent fault diagnosis method is an important technical support to improve the safety of electromechanical equipment.With the continuous innovation of computer technology and the wide application of digital sensors,the equipment health monitoring system has accumulated massive data resources.Through deep learning,we can fully explore the potential characteristics in the big data of equipment running state,which can overcome the lack of analytical model and expert experience despite rich data.The effectiveness of the deep learning method depends on the quantity and quality of samples.In the big data of equipment health monitoring,the sample number of early incipient faults is small.And multi-rate sampling,random packet loss,diversity of sensor types and different storage forms of collected information all lead to low sample quality.Aiming at the problem of small sample number and low sample quality in equipment health monitoring,the dissertation,based on transfer learning,solves difficult problems such as the incipient fault feature extraction,deep feature extraction of structural inconsistency samples,and making full use of multi-source heterogeneous information.This dissertation focuses on the key issues of transfer between deep learning models established by applying transfer learning to fault diagnosis with different fault degrees,different structural samples and multi-source heterogeneous information,so as to achieve the full use of information both in the field and between fields.The major work and innovations of the dissertation are as follows:1.To solve the problem of the small number of labeled incipient fault samples,an incipient fault diagnosis method based on deep transfer network is proposed,aiming to optimize the deep learning model of early incipient fault samples with a large number of labeled significant fault samples.The main work is to design the transfer mechanism from the significant fault diagnosis model to the incipient fault diagnosis model under different working conditions,so as to transfer the significant fault feature extraction model to the incipient fault deep learning model,to solve the problem of small sample size and low diagnosis accuracy.Therefore,the innovation of transfer between deep learning models established under different working conditions and different fault degrees in this field is realized.2.To address the problem of structural inconsistency samples at different times under multi-rate sampling,we present a fault diagnosis method based on transfer learning under multi-rate sampling,aiming to establish a bidirectional transfer mechanism from part to whole and from whole to part by making full use of samples with inconsistent structure.The main innovation is to establish a bidirectional transfer mechanism from the fault diagnosis model of structural incomplete samples to the fault diagnosis model of structural complete samples and from the fault diagnosis model of structural complete samples to the fault diagnosis model of structural incomplete samples by using a large number of structural incomplete samples and a small number of structural complete samples.On this basis,a real-time fault diagnosis system with multi-rate sampling is constructed to realize the online fault diagnosis of structural inconsistent samples and improve the accuracy and real-time performance of the deep-learning fault diagnosis model with multi-rate sampling.In this way,the innovation of transfer between deep learning models with different structural samples in the field is realized.3.To resolve the problem of insufficient utilization of multi-source heterogeneous information and data in other fields,a fusion fault diagnosis method based on multi-source heterogeneous information transfer learning is developed.This method aims to establish a transfer mechanism for multi-source heterogeneous information fusion by using multi-source heterogeneous information and data in other fields.The main work is to optimize the convolutional neural network model of screen capture images by transferring the VGG16 network model trained on the natural image data set,so as to solve the inaccurate problem of fault diagnosis model caused by a small amount of screen capture images.Then a feature extraction model of one-dimensional sequence signals is constructed,and finally a deep fusion network is designed to better extract the fusion features of one-dimensional sequence signals and two-dimensional screen capture images.It,thus,realizes an innovation at the level of feature transfer between fields and sufficient utilization of multi-source heterogeneous data.4.Aiming at the problem of poor real-time performance of fault diagnosis algorithm when multi-source heterogeneous information and data in other fields are fully utilized,we propose a fault diagnosis method of multi-source heterogeneous information fusion based on two-level transfer learning.This method,by making full use of multi-source heterogeneous information and data in other fields,constructs a two-level transfer mechanism to integrate multi-source heterogeneous information,which avoids convolution calculation to achieve real-time fault diagnosis.The main work is to construct the feature extraction network model of screen capture images,to design the transfer mechanism from the feature extraction model of screen capture images to the deep learning model of one-dimensional sequence signal,and to realize the transfer from the convolutional neural network to the deep neural network.The two-level transfer fault diagnosis model not only fuses the features of one-dimensional sequence signal and screen capture images,but also avoids convolution operation and has low time complexity,which realizes the innovation of real-time transfer mechanism for fault diagnosis.