Research On Rotating Machinery Fault Diagnosis Methods Under Multi-Source Data And Different Transfer Scenarios

Rotating machinery plays a central role in modern industries,and its accurate and timely condition monitoring and diagnosis are of great significance for avoiding production interruptions and reducing maintenance costs.The multi-source sensing information can more comprehensively reflect the operating condition of mechanical equipment.However,since industrial equipment is not allowed to operate with faults,it is still difficult to obtain comprehensive fault type data.Coupled with the complex and changeable production conditions,these factors greatly limit the engineering application of data-driven intelligent fault diagnosis technology.Based on the fusion of multi-source monitoring signals,this paper gradually reveals and solves several problems in the construction of diagnosis models for different scenarios under multi-source incomplete data,from constant equipment working conditions and changing working conditions to diagnosis scenarios where state classes are missing and state space is unknown.Firstly,in view of the problem that the insufficient feature fusion in traditional models,a multi-level fusion dual convolutional neural network(CNN)model is proposed,which performs deep level feature mining and fusion diagnosis of the data layer and feature layer from the time and frequency domain space of the multi-source signal.Secondly,in view of the significant distribution differences in heterogeneous sensor data and the difficulty in directly fusing state features,a multi-scale deep coupling CNN model is constructed.Through multi-scale convolutional modules and Maximum mean discrepancy(MMD)to map the multi-scale heterogeneous features to common subspaces for full fusion and diagnosis.Validation on motor bearing and gearbox datasets has shown that the proposed data fusion model can effectively fuse multi-source complementary information,significantly improving diagnosis accuracy and reliability.A multi-source subdomain adapation intelligent fault diagnosis method is proposed to address the problems of simultaneous transfer of multi-source diagnosis information under variable working conditions and the fuzzy marginal feature distribution matching.By constructing a multi-branch CNN network structure to match the feature distribution of each pair of source and target domains,and utilizing local MMD to accurately transfer state knowledge in the same subdomains.Finally,the weights are combined with multi-source classifiers for joint diagnosis of equipment states.In the variable working condition fault diagnosis experiment,the proposed model can effectively transfer and integrate multi-source condition information,and its diagnostic performance is superior to single source and global domain adaptation(DA)methods.An open set multi-source DA intelligent fault diagnosis method is proposed to address the problem of identifying fault forms that are not visible in model training data.By integrating the consistency,confidence and stability of multi-source classification results,a complementary transferability metric is proposed to quantify the similarity between each target sample and the known state class.and then allocate sample level weights in adversarial learning mechanisms to promote the transfer of shared knowledge.Finally,the unknown pattern detector is trained by selecting high-confidence samples to detect unknown samples.The proposed method can accurately identify shared state classes and effectively isolate unknown fault classes in variable condition fault diagnosis tasks with different degrees of openness.A universal set multi-source DA fault diagnosis model is proposed to address the diagnosis problem of the unknown relationship between the state label spaces of source and target domains in the training data.By using unsupervised clustering algorithms to analyze the internal structure of target data,a composite consensus metric that combines single-domain and cross-domain and consistency matching rule is constructed to identify shared and unknown state classes in the target domain.Subsequently,an MMD-based class-level DA algorithm is applied to expand the difference between classes while reducing the intra-class shift.Finally,state monitoring is performed by measuring the spatial distance between the test sample and the target feature clustering center.In fault diagnosis tasks in various scenarios,the proposed method can accurately match shared fault classes and simultaneously detect the presence of multiple target unknown classes.