Partial Periodic Pattern Mining Method For Time Series Of Multi-source Data

The identification of periodic patterns hidden in time series data has become one of the the research hotspots in the field of data mining.Periodic pattern is a set of events or items that appear periodically in a sequence of events or transactions,and includes full and partial periodic patterns.Compared with full periodic patterns,partial periodic patterns are more flexible in that they can capture periodic changes at specific points in the time series data and are more commonly used in practice.Traditional partial periodic pattern mining algorithms tend to mine patterns that satisfy the conditions in univariate time series,and are mostly based on tree structures to mine dynamically updated periodic patterns,resulting in significant consumption of time and memory resources due to frequent tree structure adjustments.With the increasing diversification of data acquisition methods and application environments,the sources of time series data are constantly being expanded.The time series data obtained in practical application scenarios are no longer univariate,each time series often contains different sampling points,and multi-source time series from different sources(e.g.data collected from different processes in industrial production processes)are increasingly common.It is important to find the periodic patterns of interaction between multiple and multi-source time series data from dynamic time series data.This thesis presents an in-depth study of partial periodical pattern mining methods for multi-source time series by introducing multi-scale theory,vertical data structure and locally sensitive hash correlation calculation methods.The main research contents are as follows:1.In response to the problems of high computational complexity and poor scalability of partial-periodic pattern mining for dynamic time series data,a partial periodic pattern mining algorithm(MSI-PPPGrowth)combining multi-scale theory is proposed.The MSI-PPPGrowth takes advantage of the objective temporal multi-scale nature of time series data,and uses scale transformation to mine patterns in incremental datasets.The missing counts in the PJK-Estimate Count function based on the Kriging method are used to efficiently complete the frequent period counts that are missing in the scale transformation process,avoiding the cost of rescanning the dataset and constant adjustment of the tree structure.Experimental results show that the MSI-PPPGrowth algorithm performs well on dense datasets,with a nearly 12%improvement in operational efficiency.2.Based on the classical Eclat algorithm,the multivariate time series partial periodic pattern mining algorithm(MS-Eclat)is proposed to address the unacceptable time and memory loss caused by frequent tree structure adjustment.This avoids the large amount of time and memory lost in the process of generating and adjusting the huge tree structure.Experimental results show that the algorithm reduces time and memory consumption by nearly 8% and 7%respectively,effectively improving the efficiency of mining frequent partial periodic patterns from multivariate time series.3.In order to address the pattern correlation in multi-source data time series,a relevant partial periodic pattern mining method for multi-source time series is proposed.A new periodic pattern is defined,i.e.,the relevant partial periodic pattern,furthermore the density ratio and the average periodic rate are presented to portray the periodicity and correlation between multi-source time series.Finally,the correlation calculation algorithm PW-Min LSH based on the locally sensitive hash function by considering the influence of the weight factor is given.The experimental results show that PW-Min LSH can accurately extract the relevant partial periodic patterns form multi-source data,thus effectively inscribing the interactions between different time series in multi-source data.