Multivariate Time Series Temporal Association Rules Mining Based On Trend Feature Symbolization

Complex systems are characterized by continuous processes and complex structural relationships,which show nonlinearity and uncertainty in hierarchical structure,time process and functional composition,making it difficult to establish a mechanism analysis model that can accurately describe complex systems.In addition,because the operation process of complex systems presents significant time characteristics,its operation monitoring parameters are a typical kind of multivariate time series data.Guided by the idea of data-driven control,the operating conditions of complex systems are converted to data mining.The process data are analyzed by the control optimization algorithm to realize the condition identification,fault detection and health diagnosis of complex systems.In this thesis,we focus on data mining and related algorithms of time series,aiming at discovering important operational characteristics of object development,such as changing process,changing trend and changing law,fully considering how to efficiently and accurately mine hidden temporal correlation and trend relationship among parameters in the process of system operation from multi-time series data,and mainly studying multi-temporal sequence.Trend feature extraction and temporal association rule mining of column data include the following aspects.In order to mine the useful temporal association rules in multivariate time series,the time series data needs to be converted into the symbol sequence required for rule mining.Aiming at the characteristics of large amount and high dimension of multivariate time series data representing the operation conditions of complex systems,the existing symbolic representation is used to compress the data,and the trend characteristics of operation monitoring data and the mining accuracy of time series data are fully considered.A symbolic representation method is proposed to extract sectional trend.Three basic trend symbols are defined by extracting the trend changes in the sequence segments,which represent the trend characteristics of "up","down" and "steady".Thus,they are extended to the multi-level segment trend representation,and the similarity measure calculation method of trend features is given.To improve the efficiency of mining association rules with temporal characteristics in multivariate time series data.The classical association rule mining algorithm is improved.The transaction database in time interval is transformed into a Boolean matrix.Then a new tree of frequent itemsets is constructed by using the frequent 1-itemsets and frequent 2-itemsets generated by Boolean matrix.All frequent itemsets are found and the association rules with time sequence are generated.This method constrains the mining range of sequences,reduces the number of database scans,and can quickly mine frequent patterns and store corresponding temporal relationships.In addition,pruning redundant rules can effectively reduce the waste of system resources,thus improving the efficiency of the algorithm.Combined with the research results,the research algorithms proposed in this thesis are applied to the TE industrial process data set.The data is symbolized based on trend.The symbol sequence based on the trend representation is improved by the rule mining algorithm.The experimental comparison with the similar methods demonstrates the applicability of the proposed algorithm to characterize the operating conditions of complex systems,and effectively improves the accuracy and efficiency of rule mining.