Bearing Fault Diagnosis Based On Improved Stacked De-noising Auto-encoders Network

Rolling bearing as a common part in mechanical systems,influenced by various operating conditions,is susceptible to damage.From the perspective of the pattern recognition,bearing fault identification is a process of recognizing the fault states through machine learning approach.From the perspective of network depth,the development of intelligent diagnosis has experienced the transition from shallow learning to deep learning.From the perspective of application,due to the depth of the shallow network expansion difficult,easy to fall into the local optimal,complex fault for the lack of good diagnostic results.And complex fault problems tend to need more abstract,naturally corresponding to the deep network,which can achieve an abstract representation through the combination of low-level features.Compared to the shallow network deep network has more powerful features extraction and fault classification capabilities.However,the effect of deep network diagnosis is greatly influenced by the hyper-parameters of the network.Therefore,a hyper-parameters optimization method is proposed to introduce the Stacked De-noising Auto-encoders(SDAE)network into the fault diagnosis and state recognition of rolling bearings in order to realize the diagnosis and diagnosis of complex faults.The SDAE network can extract the deep features from the original data of the high dimension.The characteristics of the SDAE network directly determine the classification effect of the network.The characteristics of the SDAE network depend on the hyper-parameters of the SDAE network,such as the number of hidden layers,The number of visible layer nodes,the number of hidden layer nodes,the sparse parameters,and the ratio of corrupted of inputs.Because SDAE network is trained based on reconstruction error under unsupervised teaching,the smaller the reconstruction error,the deep feature of SDAE network extraction can represent the original data to the greatest extent,and the network has the ability to restore the original data.In order to determine the hyper-parameters of SDAE network under the original data input,the network reconstruction error is used as the criterion to determine the reasonable range of the hyper-parameters selection The rationality of the hyper-parameters selection of SDAE network is verified by the classification accuracy and anti-noise performance analysis.However,when the original time domain signal is input,the number of network nodes is large,the structure is complex,and the efficiency of the hyper-parameters is determined.In order to improve the efficiency of SDAE network hyper-parameter selection,a hyper-parameters optimization method of PSAE-SDAE network is proposed by using particle swarm optimization algorithm to select SDAM network hyper-parameter.At the same time,the paper explores the influence of different training sets,the proportion of fine-tuning samples and the degree of imbalance of training set on the classification of faults,and provides guidance for the rational distribution of sample sets.Firstly,KPCA is used to weight the multi-source characteristics of the bearing,and the fusion characteristics are input into the PSO-SDAE network to evaluate the bearing state.The proposed method can accurately identify the early fault type and fault bearing of the bearing.