Classification Of Rotating Machinery Fault Based On Deep Belief Network

Effective fault diagnosis of rotating machinery plays an important role in ensuring the safe and continuous operation of equipment,where satisfactory fault features play a key role in diagnosis accuracy.At present,most fault diagnosis methods based on signal processing and shallow learning model depend on artificial feature extraction,which needs not only the relevant signal processing prior knowledge,but also complex extraction to obtain the suitable features,leading to uncertain feature extraction and poor fault diagnosis result.In this paper,the concept of deep learning is introduced and the characteristic of deep learning is used that the deep learning can automatically learn useful fault features from fault signals by simulating the multi-layer abstract learning mechanism of human brain.As a representative deep learning model,deep belief network is applied to the fault diagnosis of key parts(bearing and gear)of rotating machinery.The theory of deep belief network is described in detail.To overcome difficulty in model parameter setting due to no clear theoretical guidance,the influence of initial parameter setting on results is discussed in detail,providing a reference for the model parameter tuning.Firstly,the IADLR-NM algorithm is proposed by combining Nesterov momentum and self-individual adaptive learning rate to solve blind gradient descent caused by the momentum method in stochastic gradient descent and the problem that learning rate is difficult to choose.Nesterov momentum is used to predict the next position of parameter.Thus,positive or negative compensation for gradient descent is obtained to ensure that the parameters are always heading to the optimum values at the suitable speed for avoiding missing the optimum values.On this basis,the self-individual adaptive learning rate is used to adjust the decrease step length in the process of parameter decline,whose judging criteria is different from the Nesterov momentum,and to some extent,avoid model training slow down caused by conservative judgement of Nesterov momentum.Then,for frequency domain signals,the NM-based ADDBN model is proposed for fault diagnosis of the key components of rotating machinery.The model is verified by the gearbox failure data set and locomotive bearing failure data set.The result shows that the NM-based ADDBN model obtain a higher recognition accuracy compared with support vector machine and the standard deep belief network.Besides,five optimization methods,namely,Adam,Adadelta,Momentum,NM and IADLR,are used to optimize the deep belief network respectively,and the results are compared with the results of IADLR-NM optimization.Based on the decrease of RBM pre-training error and the recognition accuracy of the fine tuning,IADLR-NM can effectively accelerate model training speed and improve the generalization ability of the model without overfitting compared with the above five optimization algorithms.Furthermore,considering the lack of deep belief network in multi-fault diagnosis,hierarchical fault diagnosis model and single-layer fault diagnosis model based on NMbased ADDBN are established to automatically extract deep features from frequency domain signals,and to complete the bearing fault type and size classification and identification.The advantages and disadvantages of the two models are compared from the perspective of model recognition accuracy and computational cost,which provides a certain reference for the selection of multi-fault diagnosis model.Finally,what have been done in this research is summarized and some suggestions for the follow-up study are put forwarded.