Discovery Of Temporal Association Rules In Multivariate Time Series

With the development of the Internet of Things and wireless sensor networks,more and more various sensors are deployed into production and life.A great deal of time series data is generated by wireless sensor networks,which is an opportunity and a challenge for most manufacturing companies.Whether the potential knowledge contained in these time series data can be fully utilized has become a problem that many business decision makers and engineers need to consider.According to the engineers of a large auto parts manufacturer,the associations contained in the data produced by sensors on various mechanical devices in the factory may be helpful for understanding the working state of machine and for machine failure detection.Therefore,the research of mining temporal association rules from the time series data produced by multiple sensors are important and meaningful.The purpose of this thesis is to mine temporal association rules but not classical association rules from the data produced by multiple sensors,which is more meaningful in the context of time series data.However,classical association rule mining algorithms are not suitable for time series data mining but only for transactional data sets.After analyzing and studying several algorithms,this thesis find that the Apriori algorithm can be extended and then be applied to mine association rules on time series data.The Apriori algorithm needs to scan the whole data set each time when it calculates the support count for a candidate frequent pattern,which makes it inefficient.The improved-Apriori algorithm proposed in this thesis avoids the repeated scanning of the dataset by using a position list calculation method,which improves the efficiency compared to the Apriori algorithm.An experiment is performed on the time series data generated by 14 sensors,and by comparing the experiment to another experiment using Apriori algorithm,this thesis proves that the proposed algorithm is more efficient than Apriori algorithm.There are many redundant rules in the generated temporal association rules,and the number of rules is huge.Thus,this thesis applies two new steps into the general process for temporal association rule mining,which is the pattern pruning step and the pattern clustering step.After completing the pattern mining,the new mining process prunes and clusters the patterns,and then tries to find the temporal pattern among the pattern clusters.The temporal association rules are generated from the temporal patterns found.Based on the new mining process,an experiment is performed and the result shows that the pruning and clustering steps play a significant role in reducing the number of rules and improving the mining efficiency.Finally,this thesis analyzes and evaluates the temporal association rules found,and proves that the rules found are of high interest.Through the demonstration and analysis of some rules,it shows that the rules are of great significance to the prediction and detection of machine failure.