Research On Bearing Fault Diagnosis Method Based On Structured Depth Feature Representation

As an important part of mechanical,rolling bearing is prone to occur multiple faults under complex working conditions such as heavy load and strong impact,which directly leads to health deterioration of the whole mechanical equipment.In recent years,with the rapid development of artificial intelligence theory,the intelligent fault diagnosis techniques for rolling bearings based on vibration signal have been widely concerned.Among them,deep learning methods,which depend on a large amount of data,have been proven with better performance in bearing fault diagnosis filed.In actual applications,however,data collection is easily limited by various aspects,resulting in inadequate data available.In this case,the representation ability of feature extracted by autoencoder,which is a typical deep model,is insufficient and thus gets a low diagnostic accuracy.At same time,the diagnosis result is of great randomness and instability which restricts the practicability of deep learning methods in fault diagnosis.To solve the above problems,this paper focuses on the structural information among bearings fault states.We introduce the discriminant information of the output into the unsupervised feature extraction process to overcome the randomness and numerical instability generated by deep learning techniques when training data is insufficient.The final target of this thesis is to achieve a fast and stable diagnosis of multiple bearing faults.Specifically,the main contents of this paper are as follows:(1)Although the fault diagnosis methods base on deep learning techniques can adaptively extract more representative features from bearing fault data,they are generally computationally expensive with slow convergence speed.Although some deep learning algorithms like Multi-Layer Extreme Learning Machine(ML-ELM)can get fast training speed,there is sort of randomness inevitably.To solve this problem,a new bearing fault diagnosis method based on deep output kernel is proposed in this paper.The method firstly utilizes autoencoder to adaptively extract deep features,and then uses them to construct an objective function with output kernel regular terms.Then obtains output kernel matrix by optimizing that function,and fuses the matrix with the multi-dimensional output of the fault classifier as the final diagnosis model.Experimental results on CWRU and IMS bearing data sets show that the proposed method can effectively improve the accuracy and stability of bearing fault diagnosis in an acceptable time.(2)Focused on the traditional autoencoder couldn't extract discriminative feature and ignore the structure relationships among data,an autoencoder algorithm based on discriminative information fusion is proposed in this paper.The method firstly adds discriminant loss restricted by maximum correlation entropy loss.Then symmetric constraints of category information are added to modeling structured information between data and types.Finally,the loss function of autoencoder based on discriminant information fusion is constructed and is optimized by gradient descent algorithm.The simulation experiment results on CWRU and IMS demonstrate that the effectiveness in terms of accuracy.Moreover,the results from the Kruskal-Wallis Test also indicate the proposed method has good numerical stability.In conclusion,the proposed method has improved the accuracy and numerical stability of bearing fault diagnosis in the case of insufficient samples by utilizing the structured information between data and labels.It is of significant engineering application value for providing a new and effectiveness solution of rolling bearing fault diagnosis.