Shapelet Dictionary Learning Methods And Applications

Effective industrial big data analysis methods are the core elements of the manufacturing intelligence.As an important part of industrial big data analysis methods,the time series classification has become a hot topic in recent years.However,existing time series classification methods are difficult to balance the relationship between the accuracy and efficiency,and need to be improved to meet the requirements of real-world scenarios.The equipment health assessment technology including fault diagnosis,anomaly detection and remaining useful lifetime prediction is an important application scenario of time series classification in the industrial field.There are some individual problems in each content of the equipment health assessment technology.For fault diagnosis methods,the uncertainty of working conditions are not considered,which wildly exist in many real-world applications.For anomaly detection methods,anomaly samples are needed for hyper-parameters turning,which are scarcest in real-world applications.For remaining useful lifetime prediction methods,the lack of supervision information is adverse for feature extraction.Aiming at the above problems,a Shapelet dictionary learning（SDL）algorithm is proposed,which can balance the relationship between accuracy and efficiency.Taking SDL as a unified theoretical framework,this paper studies fault diagnosis methods under uncertain working conditions,end-to-end anomaly detection methods and self-supervised remaining useful lifetime prediction methods.For the individual problems in various fields,adaptive improvements are made respectively,and an improved method is applied to real-world problems.The main work of this paper is as follows:Aiming at the common problem of time series classification methods,an SDL algorithm is proposed to give consideration concurrently to accuracy and efficiency.Firstly,the idea of dictionary learning is introduced into Shapelet exploration.By optimizing the proposed learning model,Shapelets are found in a generative way,which improves the operation speed and feature generalization ability.Then,a support vector machine based ensemble classifier is proposed.Multiple support vector machine models are constructed by using randomly selected features and classifier parameters,final results are determined by majority voting,which can improve the generalization ability of the classifier.Finally,the accuracy and efficiency of the method are verified in a benchmark containimg 45 time series classification data sets.According to uncertain equipment conditions problems in real-world scenarios,this work improves the SDL algorithm,proposes a L₂-Shapelet dictionary learning algorithm,and implements it in fault diagnosis problems.By analyzing the sparsity mechanism,the sparsity induction method of SDL is improved,and the L₁ norm regularization term in the original model is replaced by the L₂ norm regularization term,which captures cross working condition features and improves the feature robustness.Through a single uncertain working condition case study collected in a laboratory setting and a multiple uncertain working conditions case study collected in a real-world scenario,the efficiency of the proposed method is verified.Results show that the proposed method can solve the problem of fault diagnosis under uncertain working conditions.To deal with the scarcest of anomaly samples in real-world scenarios,an end-to-end one-class classification classifier named one-class Shapelet dictionary learning algorithm is proposed.Firstly,a support vector data describtion term is introduced into the original SDL model to describe the soft boundary,which can obtain the features and decision boundary by a joint learning.Then,a hyper-parameters setting strategy of multi-scale Shapelets is proposed.By learning multi-scale Shapelets,the hyper-parameters turning is avoided,and anomaly samples are not necessary in the training stage.Finally,through a wind turbine main bearing monitoring data set from a real-world scenario,the efficiency of the proposed method is verified.Furthermore,the early fault diagnosis ability of the proposed method is also verified.Aiming at the lack of supervision information in the remaining useful lifetime prediction,a self supervised Shapelet dictionary learning algorithm is proposed.Firstly,a self supervised framework for remaining useful lifetime prediction is constructed,which designs a pretext task for the division of equipment degradation stages.Then,by introducing the contrastive learning into the SDL model,a self supervised Shapelet dictionary learning algorithm is proposed which can divide degradation stages effectively.Finally,the efficiency of the proposed method is verified through a case study of machine tool remaining useful lifetime prediction from a real-world scenario.The engineering application value of the proposed algorithms is verified through a case study covering all degradation stage of equipment.The efficiency of the proposed method in engineering application is verified by a bearing case study with all degradation stages.A prototype system is developed to deploy the proposed method.Finally,the contributions and innovations of this work are summarized,and the future work based on the limitations of this work are discussed.