Research On Fault Detection And Diagnosis Of Mechanical Equipment With Typical Limited Data

As the core of the equipment manufacturing industry,mechanical equipment is characterized by complex structure,harsh operating environment,and difficulties in operation and maintenance.Innovative development of key technologies for fault detection and diagnosis of mechanical equipment has significant economic value in addressing the current problem of safe and reliable operation of mechanical equipment.In recent years,with the widespread deployment of industrial sensors and the rapid development of artificial intelligence algorithms,data-driven methods have provided a technical means for fault detection and diagnosis of mechanical equipment.However,data-driven methods typically require monitoring data to be from the same distribution,abundant data,and complete data.In practical engineering applications,typical limited data in the mechanical equipment such as time-varying operating conditions,poor quality of monitoring data,and small fault samples,have posed numerous challenges to data-driven methods.Therefore,this dissertation systematically researches intelligent fault detection and diagnosis techniques in typical limited data scenarios,develops a prototype system for fault detection and diagnosis,and conducts engineering application verification.The main research work of this dissertation includes:(1)Aiming at mechanical equipment without fault samples and with implicit noise in the monitoring signal,a fault detection method based on shrinkage transformer relation network(STRN)is proposed.In this method,the residual shrinkage network is introduced to suppresse the impact of noise disturbances and random fluctuations,so that the intrinsic state information in the the normal state is captured.And,feature-sample pairs are built to reveal the uniqueness and commonality of normal state.Then,the transformer relation network is constructed to mine the similarity relations between the feature-sample pairs to evaluate the health states of mechanical equipment.In addition,an auxiliary sample library is built to enhance the capability of the STRN to characterise normal state.The effectiveness,noise immunity and superiority of the proposed method are verified by experiments conducted on a ship propulsion shafting and a industrial robot.(2)Aiming at the problem of variable distribution in monitoring data of mechanical equipment,a cross-multi-domain fault diagnosis method based on Res Net deep subdomain adaptive network(RDSAN)is proposed.An local maximum mean discrepancy is employed to align the distribution of each subdomain in the source domain and the target domains,thereby weakening the influence of the subdomain distribution differences caused by the change of working conditions.This enables RDSAN to extract fault features highly correlated with multiple domains,thereby improves the generalization ability of acrossdomain fault diagnosis.Two rolling bearing rigs case are usesd to the verify the adaptive diagnostic performance of the proposed method across multiple operating conditions.(3)Aiming at the problem of limited fault samples in mechanical equipment,a crossequipment fault diagnosis method based on relational conduction graph network(RCGN)is proposed,which can transfer the fault diagnosis knowledge learned from the experimental equipment to the engineering equipment.In this method,the sample pairs is used as relation nodes to construct a relation graph neural network.Then,the relation conduction rules are designed to propagate,aggregate,and update the relation nodes information in the relational graph network,thereby assisting RCGN in mining the most discriminative relation features in sample pairs.Finally,a similarity function is utilized to evaluate the health states of engineering equipment based on prior knowledge of a small number of fault samples from engineering equipment.Experimental results on two cross-equipment fault diagnosis experiments from experimental equipment to engineering equipment,the proposed method achieves significant improvements in diagnostic performance compared to transfer learning method and few-shot learning method.(4)Aiming at the zero-sample problem in fault diagnosis of mechanical equipment,a cross-unknown domain fault diagnosis method based on deep convolutional attention relation network(DCARN)is proposed.First,in the known domain,the feature extraction module is used to extract the high-dimensional features from the monitoring signal,and the features from different state types are randomly concatenated to form a feature pair.Then,a convolutional attention module is designed to mine and locate the most discriminative regions between different categories in feature pairs.The relation module is built to reveal the similarity relationships among feature pairs,enabling DCARN to have a higher capability for similarity evaluation.Finally,DCARN calculates the similarity relationship scores between unknown domain features and known domain features to evaluate the health states of unknown domains.The effectiveness and advancedness of the proposed DCARN method are verified by cross-unknown domain fault diagnosis experiments from known working conditions to unknown operating conditions and from experimental equipment to engineering equipment.(5)A prototype system for mechanical equipment fault detection and diagnosis is developed,integrating the fault detection and diagnosis models developed in this thesis.The feasibility of the proposed fault detection and diagnosis method for the engineering application is verified by the hydraulic screw pump,and the application effectiveness of the prototype system is tested.