Study Of The Method For Fault Diagnosis Of Rolling Bearing Based On Graph Embedding Auto-encoder

As one of the most common parts of rotating machinery systems,rolling bearings are widely used in many industrial fields such as aviation,aerospace,intelligent manufacturing,transportation,petrochemicals,etc.They are responsible for supporting,fixing or guiding,reducing friction and wear,and so like,which are called ‘Industrial joints'.Therefore,the effective health monitoring and fault diagnosis of the core components of rotating machinery such as rolling bearings have positive research significance for ensuring the safe and stable mechanical equipment operation.With the rapid development of measurement,sensing,computing,and other technologies,the industrial data monitored and collected presents a high-dimensional and massive distribution status.This also brings about industrial problems such as ‘dimensionality disaster',unbalanced categories,shortage of labeled information,sparse valuable information,and unlabeled information,etc.For this reason,a data-driven fault diagnosis method for rolling bearings has emerged,which can dig out useful and beneficial fault diagnosis information from a large amount of the industrial data.In recent years,deep learning has become one of the widely popular methods in the field of data-driven rolling bearing fault diagnosis,especially the auto-encoder(AE)model,which can automatically learn the sensitive hidden feature information from high-dimensional data.Thereby the accuracy of classification,clustering,or prediction will be improved.However,due to many model parameters of deep learning,the excessive demand for training samples and label information is too large,resulting in a more complicated training process.Simultaneously,the traditional deep learning is challenging to deal with the actual industrial scenes such as new samples,fewer training samples,and unbalanced categories data.This requires the introduction of a more powerful data-driven rolling bearing fault diagnostic technology.As a new type of graph neural network,graph embedding auto-encoder(GEAE)can combine graph theory with deep learning.The established graph relationship can extract the local,discriminant,sparse,and other structural information of the data to assist in different industrial tasks.To sum up,this paper introduces the theory of GEAE into the field of rolling bearing fault diagnosis.Aiming at the real industrial problems in rolling bearing fault diagnosis,the following research work has been carried out:(1)Based on auto-encoder,starting from the principle of regularized auto-encoder,a new deep Laplacian auto-encoder(DLap AE)algorithm is proposed by combining with the Laplacian local graph embedding framework.This algorithm mainly embeds the Laplacian local graph embedding into the auto-encoder model to form the Laplacian auto-encoder.Then the multiple layers of Laplacian auto-encoder are stacked to form the proposed DLap AE algorithm.To deal with the problem of the rolling bearing health data being categories imbalanced,a DLap AE-based rolling bearing fault feature extraction method is proposed.Firstly,this method improves auto-encoder performance for enhancing the manifold smoothing and feature extraction of imbalanced data through Laplacian local graph embedding.Then,the extracted sensitive features are fed to the classifier for fault diagnosis and recognition.Experimental bearing data verifies the feasibility of the proposed method.Experimental results show that the proposed method improves the feature extraction performance and fault diagnosis accuracy of the imbalanced data.(2)Based on the sparse auto-encoder model,starting from the principle of semi-supervised sparse auto-encoder,a new semi-supervised deep sparse auto-encoder(Semi-supervised deep sparse auto-encoder,SSDSAE)algorithm is proposed by combining local-non-local graph embedding and semi-supervised learning framework.The SSDSAE algorithm mainly uses the local and non-local graph embedding constraint matrix to construct the data's unlabeled information.The weighted cross-entropy can be employed to define the labeled information of the data.Finally,semi-supervised learning can be combined to realize the joint optimization of the labeled information and unlabeled information.Based on SSDSAE algorithm,to solve the fewer labeled samples,a method for extracting fault features of rolling bearings based on the SSDSAE algorithm is proposed.Firstly,the collected vibration spectrum signals are fed into the SSDSAE algorithm for feature extraction.Then the extracted sparse discriminant features are provided into the backpropagation classifier for diagnosis and identification.The analysis based on the experimental data of rolling bearings illustrates the advanced advantage of the proposed method.The analysis results demonstate that compared with other semi-supervised learning,this method can fully use the labeled and unlabeled information of the data.The extracted fault features are more detachable,and the diagnosis result is more stable.(3)Based on the research of contractive auto-encoder,starting from the principle of sparse contractive auto-encoding,an adaptive sparse contrative auto-encoder(ASCAE)algorithm is proposed by combining with sparse graph embedding and homotopy regularization framework.This algorithm uses sparse graph embedding to improve the sparse performance of the contrative auto-encoder,and then homotopy regularization is applyed to achieve adaptive optimization of core parameters.Furthermore,to overcome the defect of sparse valuable information in rolling bearing data,a new fault diagnosis method for rolling bearing based on ASCAE combined with optimized unsupervised extreme learning machine(ASCAE-OUSELM)is proposed.This method first inputs the vibration spectrum signal of the bearing to the ASCAE model to extract the multi-layer sensitive features.Then the extracted features are fed to the OUSELM classifier for unsupervised fault diagnosis and separation.The experimental data of rolling bearings validates the effectiveness and adaptability of the proposed method.The experimental results show that the method realizes the adaptive optimization of the diagnosis model's parameters,and the degree of automation of the diagnosis model can be effectively improved.(4)Based on extreme learning machine-autoencoder(ELM-AE)research,starting from the principle of unsupervised ELM-AE,a multiple-order graph embedding deep extreme learning machine-autoencoder(MGDELM-AE)algorithm is proposed by combined with multi-order graph embedding and unsupervised learning framework.Then,combined with Fuzzy C Clustering(FCM),a rolling bearing fault diagnosis method based on MGDELM-AE algorithm is proposed,which realizes the intelligent diagnosis of the rolling bearing without label information.The MGDELM-AE algorithm can use the first-order proximity embedded in the Cauchy graph embedding to extract the vibration signal's local structure information,and the second-order proximity is employed to mine the global structure information of the vibration signal to achieve unsupervised feature extraction.The extracted features are input to the FCM algorithm for unsupervised fault clustering.The measured data of the bearing verify the efficiency of the proposed method.Finnaly,the analysis results demonstrate that compared with other latest related methods,MGDELM-AE has achieved a specific competitive rapid,and accurate diagnosis effect.(5)The above-mentioned research methods are validated by using bearing failure simulation experimental data and industrial petrochemical bearing data.Firstly,the experimental overview of the rotor-bearing system fault simulation test bench is introduced.Then,the experimental data and engineering data through the above four-fault diagnosis methods(i.e.,DLap AE,SSDSAE,ASCAE-OUSELM,MGDELM-AE)can be analyzed and discussed.The analysis results demonstrate that the four diagnostic methods' effectiveness and their different application scenarios have been further verified and supplemented.Namely,DLap AE is more suitable for the diagnosis of unbalanced health data,and SSDSAE is ideal for the diagnosis with a few category labels.At the same time,ASCAE-OUSELM and MGDELM-AE are suitable for unsupervised fault diagnosis.