Research On Intelligent Recognition Of Equipment Monitoring Anomaly Based On Domain Adaptation

In the era of industry 4.0,the scale of industrial equipment continues to expand,abnormal detection technology to detect equipment problems in advance is becoming more and more important to ensure the normal operation and production of equipment.Data-driven device anomaly detection technology has become the mainstream technology of device health check.Among them,the introduction of deep learning technology with powerful data processing ability has promoted the rapid development of equipment anomaly detection technology.However,the introduction of deep learning technology also brings the following two problems:(1)The feature mapping between the distribution of the training source domain and the application target domain of the equipment abnormal data is biased.(2)The anomaly category of the equipment anomaly data training source domain is often different from that of the application target domain,which makes it difficult for the device anomaly detection model to identify the exception category of the target domain.The above problems limit the effective application of depth model in equipment anomaly detection.To solve the above problems,this thesis takes one-dimensional vibration signal data of rotating machinery equipment as an example,and studies how to solve the mapping deviation between source domain and target domain when the target domain label is unknown.At the same time,it also discusses how to detect new health category samples under the condition of target domain label extension.Specific research contents are as follows:(1)For the mapping deviation between source domain and target domain,an Unsupervised Adversarial Domain Adaptive for Fault Detection Based on Minimum Domain Spacing(MDS-ADAN)model.In this method,the importance of weight parameters in classification stage for fitting edge feature distribution is considered when reducing the distribution difference between domains in feature extraction stage.By adding the weight parameter training of the classifier,some features of the source domain are transferred to the application target domain,which reduces the difference of feature distribution between the two domains,which is reflected in the reduction of the maximum mean difference distance between the two domains,and improves the fitting characteristics of the data distribution of the source domain and the application target domain.Two experimental platforms,rolling bearing and planetary gear box,were used to carry out validation tests,including 6 diagnostic tasks.The MDS-ADAN model was comprehensively analyzed and compared with four general diagnostic frameworks and a model with only one domain alignment.The results show that compared with the Domain Adaptive Neural Networks(DANN)model,the detection accuracy of the new model can reach more than 99% when the number of parameters is reduced by 33.66%.(2)In this thesis,an Active Domain Adaptation Method for Label Expansion Problem(LDE-ADA)is proposed for label expansion.In this thesis,it is found that there is a lack of transferable knowledge of new health categories in the source domain,and the domain invariant features extracted by the unsupervised domain adaptive model only have a greater correlation with the source domain health categories,but lack the key features to distinguish the new health categories.Moreover,through feature visualization,it is found that the detection results of such samples are mostly at the boundary of the health category of the source domain,which means that the newly added health category samples have a high amount of information.Therefore,this thesis uses active learning to select the sample auxiliary model with new health categories in the target domain for training,and designs the LDE-ADA framework in three stages:the first stage,adaptive model in the pre-training domain,invariant features in the learning domain;In the second stage,active learning was used to select new health category samples in the target domain for labeling.In the third stage,the model is trained again with the labeled fusion sample set.Finally,six migration tasks were analyzed and compared on the rotating machine data set.When there is a new category of health,the accuracy of LDE-ADA can be improved by about 9.39% when 3 samples are labeled in each round and 20 rounds of training are conducted,which verifies the validity of the results.The method proposed in this thesis can effectively solve the mapping deviation between the source domain and the target domain when the target domain label is unknown.At the same time,it can cope with the expansion of the target domain label and realize the anomaly detection of cross-domain samples with good robustness and accuracy.This thesis has 18 figures,12 tables and 89 references.