Machinery Fault Diagnosis Based On Deep Autoencoder

Mechanical equipment plays an important role in modern industrial production.Faults that occur may result in huge economic losses and even casualties.Therefore,effective condition monitoring and fault diagnosis are of great importance to guarantee the security of the machinery and avoid major accidents.Feature extraction is the key to fault diagnosis.Traditional fault diagnosis methods based on signal processing and analysis and shallow machine learning all rely on discriminative features extracted artificially,which requires abundant expertise about signal processing and analysis and fault diagnosis,and cannot effectively realize the integrated and intelligent feature extraction and fault diagnosis.Therefore,with the aim of integrated and intelligent feature extraction and fault diagnosis,novel machinery fault diagnosis methods based on deep autoencoder under the framework of deep learning are proposed in this thesis and detailed as follows:(1)Sparsity is an important aspect in measuring the extracted features.In view of the demand for sparsity enhancement of features and qualitative and quantitative diagnosis of fault types and fault sizes,a hierarchical diagnosis network based on deep sparse autoencoder for sparsity enhancement of features is proposed.The deep sparse autoencoder is used to extract essential features automatically and enhance their sparsity effectively,so as to avoid redundancy and make features more discriminative.The first and second layer of the hierarchical network is used to automatically realize the qualitative and quantitative diagnosis of the fault types and fault sizes.Bearing fault diagnosis experiment fully validates the effectiveness of the proposed diagnosis network.Compared to the shallow machine learning methods and the deep autoencoder,the proposed network can effectively enhance the sparsity of features,thus,it has stronger ability of feature expression and is superior in automatic feature extraction and fault diagnosis.(2)Robustness is another important aspect in measuring the extracted features.In view of the demand for robustness enhancement of features and convergence acceleration of the network,an adaptive fault diagnosis network based on deep contractive autoencoder for robustness enhancement of features is proposed.The deep contractive autoencoder is used to automatically extract essential features and effectively enhance their robustness,so as to reduce the noise interference.The RELU function and the proposed adaptive learning rate algorithm are combined to accelerate the convergence of the network.Gear fault diagnosis experiment fully validates the effectiveness and superiority of the proposed diagnosis network and the combination of the RELU function and adaptive learning rate in accelerating the convergence of the network.In addition,the comparison experiments under different signal to noise ratio further demonstrate the ability of the deep contractive autoencoder in effectively enhancing the robustness of features to make them more expressive even under the noise interference.(3)In view of the demand for the fault diagnosis using the original time domain signals in practical applications,a combination of feature robustness and sparsity enhanced fault diagnosis network is proposed.The deep contractive autoencoder and sparse autoencoder are combined.The deep contractive autoencoder is firstly used to enhance the robustness of features and reduce the noise interference in the time domain signals.Then the deep sparse autoencoder is used to enhance the sparsity of features and improve the representativeness of the features.The features can be extracted automatically and the robustness and sparsity can be enhanced at the same time,which makes them more expressive to automatically realize the fault diagnosis with the time domain signals as the inputs.Finally,bearing datasets collected by the self-made experimental platform are used to validate the effectiveness of the proposed network,and the powerful feature learning and fault diagnosis ability of the combined network are further verified by comparing with the shallow machine learning methods,the deep contractive autoencoder and the deep sparse autoencoder network.This thesis studies the fault diagnosis of machinery based on the deep autoencoder,and proposes specific fault diagnosis networks in view of different demands in practical applications,which can effectively realize the integrated and intelligent fault feature extraction and fault diagnosis of the machinery,and of certain theoretical and practical value.