Research On Intelligent Fault Diagnosis Method Based On Deep Learning

Increasingly complex mechanical equipment is an indispensable part of modern society and plays an important role in many fields such as petrochemical industry,aerospace and power supply.Accurate and timely fault diagnosis helps to curb fault germination and prolong the stable operation time of equipment,which is an important premise to improve production efficiency and ensure personnel safety.The traditional fault diagnosis technologies are often subject to the non-stationary characteristics of the signal and the research of fault mechanism.Besides,those methods also suffer from strong subjectivity in feature extraction procedure,making it difficult to meet the growing needs of diagnosis.In recent years,deep learning has been widely used in fault diagnosis field because of its superior nonlinear characterization ability and excellent pattern classification performance,and has gradually become a research hotspot.However,due to its origin in the field of image recognition,the existing deep learning fault diagnosis methods still lack the consideration of equipment operating conditions and data attributes,resulting in limited practical effects.In view of the above defects,the research on intelligent fault diagnosis method based on deep learning is carried out.The research of this paper takes deep learning as the theoretical basis,and mainly focuses on the complex working conditions and the harsh data attributes of the equipment,including strong noise,speed fluctuation,few-shot and label-free.The main research contents are as follows:(1)In view of the limitation of the lack of denoising link in current deep learning model,a novel deep rational attention network is proposed on two different levels,including the channel level and the spatial level.In the developed deep rational attention network,the idea of threshold denoising in signal processing is introduced into the deep architecture,and a pseudo soft threshold function is developed based on the classical threshold function,which overcomes the risk of gradient disappearance during model training.Besides,the threshold value of the pseudo soft threshold function can be automatically optimized in the back-propagation process,which breaks through the limitation of traditional methods that the threshold is required to be set in advance rely on experience,and can endow the model with the ability to further ignore noise features.The essence of threshold processing in the model is discussed from the perspective of deep learning.It is found that its function is similar to that of the activation function,which can effectively add nonlinear transformation to the deep architecture,so as to improve the performance of the model.The effectiveness of the proposed method is verified by the rolling bearing data of Case Western Reserve University and the actual gas valve data of reciprocating compressor.(2)In view of the limitation of inflexible information integration under variable speed condition in conventional deep learning model,a novel deep branch attention network is proposed.The developed branch attention network constructs two parallel and relatively independent forward propagation channels in the model,and extends the attention mechanism to the outside of the model to coordinate the contribution of different branch features,which can realize the deep autonomous fusion of features within the model,and break through the limitations of single feature input or external information fusion.Besides,in the case of no speed signal,the extreme multi-scale entropy is developed as the alternative characterization parameter of speed information on the basis of the proposed deep branch attention network.Extreme multi-scale entropy can suppress the information aliasing phenomenon of the classical multi-scale entropy,and can be used as an auxiliary feature in the case of no speed signal.It gets rid of the dependence on tachometer and further improves the flexibility of variable speed fault diagnosis.The effectiveness of the suggested method is verified by the variable speed rolling bearing data of Ottawa University.(3)Aiming at the problem that small fault sample can not meet the requirements of model generalization in practical application,a novel “generation-first and transfer-second” strategy towards few-shot fault diagnosis is developed.The proposed diagnosis strategy introduces variational auto-encoder into few-shot fault diagnosis,and the fine-tuning transfer mechanism is adopted to further improve the utilization of the generated data,which provides a new idea for few-shot fault diagnosis.In the diagnosis procedure,the variational auto-encoder is first employed to obtain the latent distribution characteristics of the data in high-dimensional space,and expands the sample space through random sampling and decoding,which breaks through the limitation of sample number on model training.Then,the residual network classification model is pre-trained by a large amount of generated data to learn the general basic features.Finally,after the pre-training converges,a small number of actual samples are utilized to fine-tune the pre-trained classifier model,which can make up for the difference between the generated data and the actual data and improve the actual diagnosis performance.The effectiveness of the proposed method is verified by the rolling bearing data of Case Western Reserve University.(4)Aiming at the problem of training data label missing and difficult to label in practical application,a novel sub-domain joint distribution adaption capsule network is proposed.The developed sub-domain joint distribution adaption capsule network integrates the idea of sub-domain adaptation into the joint maximum mean discrepancy,and adopts multi-kernel strategy to enhance algorithm robustness,which overcomes the one-sidedness of global or marginal distribution matching.By minimizing the local joint maximum mean discrepancy,the method of this paper can match the joint distribution of the feature space and the prediction space in different domains at the sub-domain level,which ensures the refinement and comprehensiveness of domain adaptation.In addition,the introduction of the vectorization capsule adaption layer can accommodate the difference between source domain and target domain to a certain extent,which allows the model to get rid of the limitations of the scalar mode of the original fully connected adaption layer,and endow it with stronger generalization ability.The effectiveness of the developed method is verified by the rolling bearing data of Paderborn University and the general industrial gearbox data of PHM2009 under multiple label-free transfer tasks.Through the above research,the deep learning fault diagnosis models suitable for complex working conditions and harsh data attributes of mechanical equipment are established,which improves the diagnosis performance under restrictive conditions such as strong noise,speed fluctuation,few-shot and label-free.It enriches the theoretical basis of mechanical fault diagnosis,and provides guidance for practical engineering application.