Research On Fault Diagnosis And Remaining Useful Life Prediction Of Rotating Machinery Based On Deep Learning

As one of the most widely used mechanical equipment,the safe and stable operation of rotating machinery is crucial to industrial production.However,the long-term operation of rotating machinery in the complex working environment causes malfunctions of the key components(such as bearings,gears and rotors),resulting in serious loss of life and property.Therefore,the research on fault diagnosis and reminding useful life(RUL)prediction of rotating machinery has become a hot field.Since deep learning models can adaptively extract and learn the features of faults and health degradation of components from monitoring data,it effectively overcomes the shortcomings of traditional fault diagnosis and RUL prediction methods,which rely on complex mathematical modeling and a large amount of prior knowledge.Therefore,the deep learning-based fault diagnosis and RUL prediction(DLFDRP)has gradually become the mainstream research direction to ensure the safe and stable operation of rotating machinery in modern industry.Based on deep learning methods,this thesis takes fault diagnosis to RUL prediction as research sequence,aiming at the problems in the field of DLFDRP such as unbalanced number of samples between different classes,incomplete feature representation of faults,and difficulty extracting features of health degradation process(HDP)from vibration signals and onesidedness of HDP feature extraction.The main research contents are summarized as follows:Aiming at the difficulty of fault feature extraction due to the unbalance data between classes,a deep learning framework combining small sample expansion and fault diagnosis is proposed in this thesis.This framework includes a novel dilated residual generation adversarial network and a deep learning fault diagnosis model integrates dilated residual network and extreme learning machine(ELM),which can realize high-quality virtual sample generation and accurate fault diagnosis.Moreover,the framework overcomes the disadvantage that the traditional down-sampling methods such as pooling layers is easy to lose the data information,and ensure the integrity of the representation information of the fault modes under the premise that the network has sufficient receptive field.In addition,a fault diagnosis module based on ELM is designed to diagnose fault classes instead of the traditional full-connection layer,which can accurately establish the nonlinear mapping relationship between features and fault classes without iteratively updating the weights,and simplify the training process.In order to make the proposed framework more general,the diagnostic results of laboratory single faults,laboratory compound faults and industrial mechanical faults are presented and discussed.To solve the problems of information loss caused by data domaintransformation and high complexity of deep learning models,a lightweight deep learning framework with a time-frequency feature fusion strategy is proposed to diagnosis faults in this thesis.In the stage of feature extraction,two lightweight temporal convolutional network branches are designed to extracted one-dimensional temporal features and two-dimensional time-frequency features of fault samples,respectively,and then the feature fusion is carried out to enable the framework to conduct more comprehensive representation learning of different fault modes.In order to further reduce the complexity of framework,a classifier based on broad learning system is designed to identify faults in the stage of fault diagnosis which can overcome the shortcoming of too many parameters caused by stacked dense layers in the traditional classifier on the premise of ensuring accurate fault diagnosis.Besides,additional noise is introduced in the datasets to verify the robustness of the framework.Experimental results show that the proposed framework has high fault diagnosis performance,strong robustness and low complexity compared with other fault diagnosis models.To solve the difficulty of temporal feature extraction of components’ HDP due to the complexity of vibration signals of rotating machinery,a deep learning-based RUL framework with strong capability of temporal feature extraction is proposed for RUL prediction in this thesis.The framework contains an adaptive extraction method of health degradation indicator based on temporal convolutional autoencoder,which strengthens the temporal features contained in vibration signals.The experimental results show that the health degradation indicator can effectively improve the representational learning ability of the proposed framework.Besides,the feature extraction module with strict time constraint structure in the framework can effectively capture the temporal information of the constructed health degradation indicator and assign multi-perspective attention weights for extracted features,which improves the feature expression ability of the framework for the HDP of components.Then,a residual network-based feature fusion module is designed to fuse the extracted multi-perspective features and mine the deep semantic information from the features for RUL prediction.The experimental results prove the structure rationality and performance superiority of the proposed framework.To solve the problems of incomplete HDP feature extraction caused by single receptive field and feature scale of traditional deep learning-based models for RUL prediction,this thesis proposes an end-to-end RUL prediction framework with capability of multi-scale temporal feature extraction.The feature extraction module in the framework contains three dilated causal convolutional network branches with different receptive fields,which can effectively extract multi-feature level information contained in the raw vibration signals and fuse the information without destroying the temporal relationship.Experimental results show that the multi-scale feature extraction structure can effectively improve the comprehensiveness of HDP feature representation.In addition,via introducing a self-attention mechanism at the beginning of each network branch,the globally dependent and diversified feature representation of vibration signals is realized.Compared with existing RUL prediction models based on multi-scale features,the proposed framework can improve the diversity of feature representation of samples,improving the RUL prediction performance.