Research On The Prediction Method Of Equipment Remaining Life Based On FastDTW-LSH-Shapelet

With the continuous development of the industrial intelligent management system,the stable operation of the core components of industrial equipment has become very important,and the health monitoring,fault diagnosis of industrial equipment and the prediction of the remaining using life has also become a research hotspot.In the field of equipment remaining life prediction research,the current mainstream method is a data-driven remaining life prediction method,which is based on the analysis of historical data to establish a corresponding data model to make predictions.In recent years,the remaining life prediction methods based on data-driven analysis are mainly divided into two categories.One is the data-driven method based on machine learning theory which predict the remaining useful life of equipment through different network models are established by analyzing data.The other is the data-driven method based on the time series similarity theory,which uses the local characteristics of historical data to make predictions.Among them,the remaining life prediction method based on shapelet is a research hotspot.The shapelet template that carries the degraded information is filtered from the historical data.And match with the current data to predict the remaining life.The main problem encountered is that the size of the candidate subsequences is too large,which makes the shapelet screening time-consuming.In this regard,this dissertation adopts the locality-sensitive hash mapping to screen shapelets.During the mapping process,it is proposed to use FastDTW distance to measure sequence similarity and classify them.After screening,increase the diversity of candidate subsequences and improve the recognition degree of shapelet,enhance the quality and interpretability of the RUL-Shapelet template library in the prediction process,improve the matching performance of the template library and the current state,and improve the prediction accuracy.（1）There was obvious subsequence clustering in the hash bucket after LSH mapping,while random screening in the LSHST（LSH Shapelet Transform,locality-sensitive hashing shapelet transformation method）caused more subsequences with different characteristics to be missed,which affected the quality of shapelet,a FastDTW-LSH shapelet screening method was proposed.The subsequences in the high-dimensional space are mapped to the corresponding hash buckets through the locality-sensitive hash function.The subsequences are classified,and candidate subsequences are randomly selected from each class for the next hash calculation.After multiple hash operations,shapelets with more interpretability are screened.The experimental results show that compared with LSHST,the shapelets selected by the FastDTW-LSH are significantly better than the classification accuracy of LSHST on 8data sets under the premise that the time of shapelet transform is not increased significantly,and the classification accuracy is increased by 11.57、9.16 percentage points under Cricket Z、Fifty Words.（2）The above FastDTW-LSH method is applied to the remaining life prediction,and a screening method of the RUL-Shapelets template library based on this method is proposed.During the FastDTW-LSH mapping screening process,the diversity of subsequences is increased and the template library is improved.In the matching stage,the FastDTW distance between the template library shapelets and the current state is calculated and given different weights,and the remaining life of the current state is predicted according to the carried degraded information,which improves the matching performance between the template library and the current state and improves the prediction effect.The experimental results show that,compared with the LSHST method for predicting the remaining life,the prediction algorithm proposed in this dissertation not only reduces the time consumption,but also improves the prediction performance.