Research On Data-Driven Fault Diagnosis Method Based On Deep Convolutional Neural Networks

With the rapid development of automation and information technology,the complexity and intelligent level of equipments in aviation,aerospace,energy,man-ufacture and other fields are increasing gradually.The traditional fault diagnosis methods are difficult and even unable to satisfy the requirements of complex sys-tem fault diagnosis.In order to improve the safety and reliability of complex equipments,under the background of the current big data driven intelligence era,it is of great significance to study the intelligent fault diagnosis methods for com-plex systems.Convolutional neural networks(CNN)is a typical deep learning model,it has been widely used and achieved brilliant results in image recognition,video classification,speech separation,face recognition and other fields due to the characteristic of deep extraction and mining for data features.It is worth noting that deep convolutional neural networks have great potential in fault diagnosis due to the excellent performance in dealing with non-linear and non-stationary signals.Therefore,aiming at the problems of traditional diagnostic methods,e.g.they can not achieve deep feature extraction and have the limitations to deal complex operation conditions,the research on a data driven fault diagnosis method based on deep convolutional neural networks has been carried out.In view of the shortcomings of the shallow fault diagnosis methods,a data driv-en fault diagnosis method based on deep convolution neural networks is proposed in this paper,named EHHT-CNNs.Remarkably,the proposed EHHT-CNNs can not only automatically achieve deep feature extraction considering that there is no expertise in vibration analysis,and owns good robustness for different levels of noise.Specifically,the vibration data are processed to obtain the multi-band data by EHHT,which realizes primary feature extraction by the mapping from a low-dimensional space to a high-dimensional one.And then,the multi-band data are reconstructed to be taken as the input and train the model built by CNNs via supervised and implicit learning,which realizes the deep feature extraction of data.Besides,this paper studies the effects of various parameters on the perfor-mance of EHHT-CNNs based on the principle of single variable,and summarizes the qualitative law of the parameter adjustment of this method.Based on the rolling bearing data,the fault diagnosis experiments are carried out by EHHT-CNNs,and the experimental results show that EHHT-CNNs has high diagnostic accuracy and good robustness.In view of the limitation of traditional methods in dealing with complex con-ditions,this paper proposes a condition identification method based on K-means,named MK-means.This method can identify the complex conditions and has anti-interference ability to noise.MK-means obtains the window data by moving window to process time series first,and then K-means is used to complete the clustering of window data in an unsupervised way.To verify the performance of MK-means,the experiments are carried out separately based on the analog signals and the historical data of hydroelectric units.The experimental results show that MK-means can effectively realize the condition recognition of complex systems,with high recognition accuracy and strong anti-interference ability.