Mining Dynamic Association Rules From Multiple Time-series Data Streams

Multiple time-series data streams exist in industrial processes, commercialactivities and natural science,etc. Therefore, it is meaningful to do the research ofthe knowledge discovery of multiple time-series data streams. An importantresearch direction of the streams mining is mining the association relationshi psbetween multiple time-series data streams. Time-series data stream has the similarcharacteristics with data stream like massive, continuity and liquidity,etc. Data andknowledge contained in the data of time-series streams changes over time. Inpractice knowledge in the new data may be more interesting than that in old. Amethod of mining dynamic association rules which change over time and whosemeta-patterns which have only one item maybe have different length of time slot indifferent rules in a sliding window from multiple time-series data streams isproposed.So far there are a number of association rules mining researches overtime-series data. These researches mine a variety of the types of rules. But most ofthe researches mine the rules whose basic elements, that is the meta-patterns whichhave only one item and have the same length of time slot. And most of them do notconsider the case that sometimes knowledge in the new data may be moreinteresting.In this paper when streams flow, the data of streams are preprocessed for rulediscovery. The preprocessing includes piecewise linear approximation, segmentinglinearized time series to let each stream have only one line segment in one time slot,and then incrementally cluster these line segments, symbolic representation of thedata, and merging preprocessed streams into transactions from which rules aremined. After preprocessing we use a rule finding method to obtain rules. Thepatterns in the sliding window are stored in a summary data structure SWFI-tree.Through periodically pruning the obsolete and infrequent patterns are deleted. Todifferentiate the patterns of the latest generated transactions from those of historictransactions, a time decay model is introduced. In this model there exists a decayfactor to reduce the weight of patterns of historical transactions. Through the analysis of the experimental results of the actual data of thermalpower plant we can get that the method is effective. It can mine the relatedassociation rules.