A Three-branch Incremental Learning Method For Frequent Patterns In Multivariate Temporal Sequence

Multivariate time series data exists widely in all walks of life.People obtain the data reflecting the system characteristics by deploying sensors,and obtain useful results through data mining.Multivariate time series data has the advantages of simple structure and strong interpretability,which can intuitively reveal the time series relationship between data.However,the data is usually huge and dynamically updated,how to quickly mine the latest sequential pattern is an important issue worth studying.This thesis first defines the problem of incremental mining of State Transition Pattern with Perioidc Wildcard Gaps on multivariate time series data.Secondly,under the guidance of the three-way decision theory,an accurate and efficient three-way incremental learning method(3IU-STAP)is proposed.Thirdly,the classic association rules are improved to the time scries association rules applicable to STAP,and the corresponding mining algorithms are given.Finally,a rapid prototyping system including the timing prototype rules of STAP and its mining algorithm is designed and implemented.The 3IU-STAP algorithm constructs candidate patterns based on raw data,frequent STAP,and incremental data,and uses thresholds to divide candidate patterns into positive,negative,and boundary domains.The patterns in the positive domain can be directly regarded as frequent patterns and stored;the patterns in the negative domain are regarded as infrequent patterns and discarded;only the candidate patterns in the boundary domain need to delay decision,that is,scan the data set to judge whether it is frequent.In terms of accuracy,an incremental data supplement technology is designed,through which the accurate number of occurrences of candidate patterns can be obtained.In terms of efficiency,the down-closed nature is used to control the number of candidate patterns,and then three-way decision models are used to avoid repeated scanning of the data,thereby further saving time.Although high-frequency sequences can reflect some basic rules existing in the data,it is difficult to reflect the interdependence and correlation among data.Therefore,according to the characteristics of multivariate time series data,an association rule mining method based on state transition mode is proposed,and a reasonable confidence calculation formula is designed.Association rules can more intuitively reveal the timing relationship between various states,thereby providing users and experts with richer decision information.In order to reflect the results of association rule mining,the mining results of STAP frequent patterns and the visualization results after screening association rules are displayed on four real data sets.For the 3IU-STAP algorithm,experiments were also designed on four real data sets and two man-made data sets.The total time of the incremental algorithm,the incremental time and related factors that may affect the algorithm are discussed.The results showed that,compared with the non-incremental method,the 3IU-STAP algorithm can significantly improve the time performance while obtaining accurate results.