Research On Data-driven Based Intelligent Fault Diagnosis Methods Of Transmission Rotating Parts

Due to the characteristics of large capacity,diversity and high speed shown in the era of mechanical big data,traditional fault diagnosis methods still rely on manual feature extraction and a large amount of prior knowledge,which cannot cope with the problems of accurate fault classification under the conditions of large capacity data,large speed fluctuation and unbalanced samples.Therefore,more advanced and intelligent ideas and methods are urgently needed to solve these practical problems in engineering.In this paper,in order to meet the diagnostic requirements of mechanical transmission rotating parts,we have conducted the in-depth research on data-driven based transmission rotating parts(gear,bearing and shaft)intelligent fault diagnosis methods based on deep learning theory.The contributions of this thesis are summarized as follows:(1)In order to overcome the weaknesses of traditional fault diagnosis methods on manually feature extraction and big data processing,we combine the characteristics of mechanical big data with the advantages of deep learning,stacked autoencoders(SAE)are employed to constract deep neural networks(DNN)and the bearing fault signal spectra are used as input,so that the automatic feature extraction and fault diagnosis under mechanical big data are achieved.Thus the proposed method overcomes the shortcomings of traditional methods in fault feature extraction and classification.Meanwhile,we have made a preliminary exploration of the hierarchical feature learning process of DNN.Through the visualization operation,the feature learning process is presented layer by layer.The results show that with the deepening of network layer,the learned fault features are gradually clear and obvious,and the characteristics of different fault types show obvious different trends.(2)Aiming at the problems of too long training time consumption caused by large fault data capacity and too many iterations of neural network,batch normalization(BN)is introduced into the intelligent fault diagnosis field of bearing and gear,so a novel method named batch normalized deep neural networks(BN-DNN)is proposed for achieving fast fault diagnosis.Firstly,SAE are used to build DNN,and then BN technique is applied to the ahead of nonlinear mappings(activation functions)in each layer.So that each layer of the network input can have a stable distribution,and become more conducive for network training.The analysis results of bearing and gear datasets show that the proposed method can achieve better diagnose results with less training samples,less time and iteration times than the DNN without BN.(3)In view of the shift and scale function on input signals of BN-DNN method,it is considered to be applied to the fault diagnosis research under large speed fluctuation.Due to the frequency shift and amplitude change of the fault signals in the frequency domain with large speed fluctuation,the general algorithms cannot accurately distinguish the fault signals.However,the shift and scale parameters of BN technology can be used to shift and scale the characteristic frequencies of different speed signal samples under the same health condition,so that the characteristic frequencies of all the same fault samples can be highly unified.Thus,the problem of accurately identify bearing faults under the background of large speed fluctuation is resolved.(4)Aiming at many deep learning models need supervised fine-tuning process of BP algorithm on fault diagnosis problems,an unsupervised learning method named sparse filtering(SF)is discussed.The method focuses on optimizing the sparsity of learning features and ignores the distribution of learning data.So there is no need of BP algorithm participating in the training process,and only one feature parameters need adjustment.Firstly,SF is used to extract the unsupervised features from the spectra of training samples,and then softmax is adopted as the classifier to realize fault diagnosis.Through a set of bearing fault test,the proposed method is proved to be powerful in fault feature extraction and classification.(5)In order to directly solve the problem of fault diagnosis under large speed fluctuation through the original time domain signals,the regularization technology in deep learning is considered,and L_1/2 norm is selected to regularize the objective function of SF.So a new method named L_1/2-SF is proposed,which allows sparse filtering to extract features from the overlapped vibration signals in large speed fluctuation.Then,softmax is used as the classifier to realize fault classification.Finally,the effectiveness of the proposed method is verified by two special designed bearing large speed fluctuation datasets.(6)In view of the problem that fault diagnosis effect is not so good under the imbalanced dataset,that is,more samples of health condition,fewer samples of fault conditions and the sample numbers of different fault conditions are different.Wasserstein generative adversarial networks(WGAN)are introduced in our research,and a new method named WGAN-SAE is present.Through WGAN simulating the real fault signals to generate artificial signals,so as to increase the number of generated samples in different fault conditions,so that the number of training samples in each health condition can achieve a balanced state.Then,the SAE is used to extract and classify the fault features layer by layer.Through the experiments of three different imbalanced datasets of gear and shaft,it is shown that the diagnosis accuracies of the balanced dataset by the WGAN-SAE method are much higher than that of the imbalanced dataset.