Research On The Classification Method And Application Of Mahalanobis-taguchi System For Time Series Data

As one of the research hotspots in time series data mining technology,time series classification problem has been applied in all walks of life.Massive time series data classification,mining useful information and knowledge,can be used to guide life practice and quickly solve practical problems.However,the characteristics of time series data have high dimensionality and change with time.Its generation process is very susceptible to environmental factors and there is a certain amount of noise.For such complex data,studying how to obtain effective information and knowledge from it is of great theoretical research value and practical significance for scientific research and social production events.The Mahalanobis-Taguchi System(MTS)is a quantitative pattern recognition method for classification,diagnosis,and prediction of multivariate data.MTS has many advantages,e.g.,MTS is a data-based analysis method and can really reduce the number of variables,which can simplify problems and improve the accuracy and efficiency of classification.MTS builds a co ntinuous measurement scale and calculates the deviation degree of test sample from the refere nce space,which is conducive to adopt corresponding solutions and improve the flexibility of problem solving.However,there are still some deficiencies in the theory and application of M TS.This paper is oriented to time series data.In view of the problems existing in traditional MTS,it is improved by using multiple Mahalanobis distance method,improved quadratic loss function,multi-tree theory,etc.The goal is to develop MTS into a Efficient method for time series classification.Due to the high-dimensional nature of the time series itself.In practical applications,it is necessary to perform local feature extraction or global feature decomposition on the time series,thereby reducing the dimension of the original time series data.Different time series data types have different feature extraction methods.Therefore,this article discusses the extraction of time series features from different types of time series.The research work of this paper mainly includes the following aspects:(1)Research on MTS algorithm optimization.When MTS is oriented to time series data classification,a series of problems usually occur,such as the problem of large number of features,large amount of data and many types of categories.In view of the limited number of traditional MTS feature variables,the concept of feature subsets and a new measurement scale—Multiple Mahalanobis Distance(MMD)are introduced to improve MTS.At the same time,under the background of big data,the error rate when selecting training data redefines the loss value in the quadratic loss function,so that the loss is no longer subjectively defined.Therefore,by improving the system threshold,the recognition accuracy of the algorithm is improved.With the help of multiple classification rule(MCR)and multi-branched tree(MT)theory,the application of MTS in the field of multiclassification is broadened.Research shows that after the improvement of the above methods,the classification performance and application breadth of traditional MTS have been improved.(2)Research on MTS classification method and application for time series data with typical features.Taking the classification of time series data with typical features as the research background,the discussion on feature extraction and MTS classification research is carried out.First,considering its typical characteristics,a time series hybrid model that can characterize its characteristics is constructed.The parameters of the mixed model are estimated through periodic graphs,least squares,step discrimination and statistics.Thus characterizing the typical characteristics of the original time series.Secondly,the typical features extracted from the time series are combined with the MCR-MTS algorithm with improved threshold,and the advantages of MCR-MTS in feature optimization and multi-classification are used to realize fast and efficient classification of time series.Finally,in order to verify the classification ability and effect of the mixed model feature extraction combined with the MCR-MTS algorithm.The control chart data set of UCI database was used for experiment,and compared with other commonly used classification methods.The results show that the mixed model based on time series and the MCR-MTS algorithm can effectively characterize the time series with typical characteristics,simplify the classification system,and have high recognition accuracy.It is a more effective classification method for typical feature time series.(3)Research on MTS classification method and application of signal-oriented time series data.Taking the time series data classification of signal class as the research background,the discussion on feature extraction and MTS classification research is carried out.First,through variational modal decomposition,the signal data is decomposed into multiple modal components,and multiple signal features in each component are extracted.At this time,for each original vibration signal,the number of features is huge,and at the same time,the time series classification problem of the signal class is usually a multi-classification problem.Therefore,the MMD-MT-MTS algorithm is introduced to solve the problem of large number of features and multiple categories.Moreover,the advantages of MTS orthogonal table and signal-to-noise ratio in optimizing the system are used to select the sensitive modal components in each category that are helpful for classification,so as to truly realize the unified combination of variational modal decomposition and MTS classifier.Finally,in order to verify the classification ability and effect of the signal decomposition feature extraction combined with the MMD-MTMTS algorithm,the experiments were carried out using the bearing data collected by the rolling bearing fault simulation test bench of the Electrical Engineering Laboratory of Case Western Reserve University.Then compare the experimental results with the algorithm diagnosis results in other literatures.Studies have shown that based on variational modal decomposition combined with MMD-MT-MTS algorithm can effectively extract the time series features of the signal class,and the classification performance and dimensionality reduction effect are better.It is a more effective classification algorithm for signal-like time series.(4)Research on MTS classification method and application for general time series data.Taking the general time series data classification as the research background,the discussion on feature extraction and MTS classification research is carried out.First,the time series is divided into multiple sub-sequences by dividing the time series into equal lengths,and multiple time-frequency domain features in the sub-sequences are extracted.At this time,for each original time series,the number of features grows multiples with the increase of the number of segments.At the same time,the time series classification problem is usually a multiclassification problem.Therefore,for the general time series,the feature extraction is not as accurate as the above research.In order to improve the classification accuracy,the MMD-MCRMTS algorithm is introduced to solve the problem of a large number of features and multiple categories.Using the advantages of MTS orthogonal table and signal-to-noise ratio in optimizing the system,the discriminating subsequences in the original sequence are selected.Finally,in order to verify the classification capability and effect of the segmented feature extraction combined with the MMD-MCR-MTS algorithm,the time series data set provided by the University of California Riverside was used for experiments.The results are compared with the algorithm diagnosis results in other literatures.Studies have shown that segmented feature extraction combined with MMD-MCR-MTS algorithm can effectively extract general time series features,can find discriminative subsequences and has better classification performance.It is a more effective classification algorithm for general time series,and has good practical application value.In summary,this paper takes the time series classification problem as the research object,aiming at the deficiencies of MTS,taking MTS improvement as the main line,taking multiple Mahalanobis method,multi-tree algorithm and various feature extraction methods of various types of time series as the main Means,the goal is to develop MTS to become a practical and efficient classification method suitable for time series,and applied to the study of real problems.