The Research Of Intershaft Bearing Fault Diagnosis Based On Kernel Auto-encoder

Intershaft bearing is one of the critical components of the aeroengine and its runing condition closely relates to the overall performance of the engine.Due to the complicated and changable operating conditions of the double rotor system of the engine,the vibration of the intershaft bearing is frequently troubled.Therefore,it is of great significance to study the fault state of aeroengine intermediate bearing and make diagnosis.This article takes the vibration signal of the intershaft bearing as the research object,and uses the deep neural network as the diagnostic method.Aiming at the deficiency of deep auto-encoder neural network to analyze the original signal with large noise,a kernel auto-encoder is proposed as an improved method.The research content of the thesis is summarized as follows:(1)Discuss the research background and the significance of the topic.Introduce the research progress of the fault diagnosis method of the rolling bearing and the aeroengine intershaft bearing,as well as analyze the application of the deep learning method in the field of rolling bearing fault diagnosis.Three typical deep learning basic models are introduced,and an in-depth study of stacked auto-encoders is conducted to explore the pre-training and fine-tuning network architecture.Using the frequency domain feature as input,a stacked auto-encoder neural network was used to diagnose the typical failure of the intershaft bearing.The results show that the deep neural network can obtain better diagnostic results than the traditional shallow model.(2)To deal with the inadequacy of the stacked auto-encoder model in handling the noise-containing original signal capability,a kernel auto-encoder was proposed in combination with the kernel function method.A stacked kernel auto-encoder network was constructed with the original signal,frequency domain features,and mixed features respectively to validate model.To solve the problem of poor generalization and parameter selection of the proposed kernel auto-encoder,a stacked kernel denoising auto-encoder model with L2 penalties is proposed,and an improved firefly algorithm based on chaos is used to optimize the parameters.The intershaft bearing failure test data verified that the proposed method can extract better features from the original data than the typical deep neural network,and the diagnostic accuracy is higher.(3)Aiming at a large number of nonlinear components in the original signal of vibration,an unsupervised feature extraction model is proposed combining with wavelet kernel function and sparse auto-encoder,and a kernel extreme learning machine is used to diagnose the faults.Firstly,a stacked morlet wavelet kernel sparse auto-encoder is constructed,and the features of the running state of the intershaft bearing are extracted from the original vibration signal unsupervisedly.Then,the second fusion of the extracted features can be performed by the LPP method,which can reduce the redundant components in the features.Finally,the kernel extreme learning machine was used to identify and diagnose the faults.The proposed method can obtain great results with the verification of the typical fault data of the intershaft bearing.(4)The application of the algorithm is implemented by LabVIEW combined with MATLAB software based on the previous research.Through the development of the aeroengine intershaft bearing condition monitoring and fault diagnosis system,the deep learning method was applied to the practice of real-time diagnosis of intershaft bearings.The system can not only monitor the real-time running status,save the data and analyze it offline,but also can use the deep learning model to diagnose the fault in real time.