Font Size: a A A

Research On Fault Diagnosis Of Mechanical Transmission System Based On Deep Learning Model

Posted on:2018-06-22Degree:MasterType:Thesis
Country:ChinaCandidate:X J GuoFull Text:PDF
GTID:2322330542463342Subject:Control engineering
Abstract/Summary:PDF Full Text Request
The deep neural network is an abstract computational model to allow the computer to automatically learn useful features in terms of the perspective of information processing.Its excellent feature extraction function has a promising application in the fault diagnosis of mechanical equipment.The fault diagnosis method based on the traditional shallow learning model requires a priori knowledge about the mechanical signal and the fault type,and cannot comprehensively extract the fault features for accurate classification.In this paper,the fault diagnosis of mechanical transmission equipment based on the depth model is described in detail.The fault diagnosis system based on the frequency domain signal and the time domain signal can extract the characteristics automatically,efficiently and objectively to complete the fault diagnosis and fault size prediction target.Firstly,an adaptive learning rate deep neural network algorithm is proposed to solve the problem that ubiquitous learning step is not easy to choose.The optimal step size is chosen according to the gradient direction of the adjacent two iterations:if the two gradient directions are the same,the weight optimization speed is increased,or the gradient is slowed down.In order to verify the effectiveness of the method,a convolution neural network is taken as an example.The comparison between the fixed learning rate algorithm and the adaptive learning rate algorithm is carried out.It is proved that the adaptive learning rate algorithm can reduce difficulty in convergence due to unsuitable initial learning rate.Secondly,an Integrated Stacked Denoise Auto-encoder Model(ISDAE)is proposed for the frequency domain signal.The model first converts the original time domain signal into the frequency domain and extracts the noise from the signal and extracts the feature through the noise reduction layer and the feature extraction layer.Experiments were carried out using the rolling bearing data set and the gearbox data set.The results show that ISDAE has higher recognition accuracy than other fault diagnosis models(such as DBN).Especially in the case where the original signal has random noise.Its noise reduction capability makes the model still achieve high precision diagnosis.Furthermore,in order to directly use the original time domain signal for classification,a hierarchical fault diagnosis model based on Adaptive Deep Convolution Neural Network(ADCNN)is proposed.The model can extract and classify the original time domain signals and solve the problem that the signal feature is lost in the frequency domain transform process and it will increase the complexity of the diagnosis process,so as to achieve high recognition accuracy.The reliability and superiority of the proposed model are proved by comparing with the classical fault diagnosis method,SVRM model.Finally,what have been done in this research is summarized and some suggestions for the follow-up study are put forwarded.
Keywords/Search Tags:Deep learning, ISDAE, ADCNN, Fault diagnosis, Feature extraction, Mechanical rolling bearings
PDF Full Text Request
Related items