Research On The Framework Of Mining Abnormal Pattern On Multiple Correlative Time Series

As the rapid development of data collecting and technique of World-Wide-Web and computer, many corporations and organizations have built real-time surveillance systems.There come into being so many data(can be regarded as time series) by them. How to make use of these data to achieve real-time decision-making for corporations and organizations has become a hotspot in academic and industial community. The putting forward of the real-time data warehousing has pointed out a new way to solve the hot problem brought above. Active decision-making is just about the direct motive. However, the main problem when to make active decision-making is how to give an alarm and carry out the next forecasting when discovering the abnormity. It is very different from the traditional ways for diagnosing the abnormity, and it is an systemic analysis method. It contains three parts, namely fiding abnormal patterns, analyzing abnormal patterns, and the reasons cause the abnormity.Therefore, this thesis presents an framework of mining abnormal patterns on multiple correlative time series which is based on the reaserch on the three parts mentioned above. The framework plays an important role in the actual production and lives.This thesis deals with the data produced from technical flow in hot rolling as time series, and uses the mean of statistical process control to find abnormal patterns in single time series (or in a single monitored variable). Then, we use the ways for mining sequential patterns and association rules to mine the relations among the abnormal patterns. We can use these patterns or rules to online forecast the abnormal which is possible to appear during the real time production later. Thereby, combining with data mining and statistical process control approach, this thesis mainly discovers the abnormal on time series, then puts forward a process framework of analyzing abnormal patterns. It contains two steps:(1) mining the abnormal patterns based on SPC method on single time series, and (2) discovering sequential pattern and association rules on multiple correlative time series. At the end, we can make use of the association rules or fruquent abnormal patterns to discover the relation amang the abnormal patterns. Experiments show that the framework is feasible, and it triumphantly discovers the relations among abnormal patterns and provides the bases for diagnosing and analysing abnormal patterns.The characteristic of the framework is capable of annalyzing and dealing with multiple correlative time series. It mainly focuses on finding interested patterns over time series, such as abnormal patterns, frequent sequential abnormal patterns, abnormal pattern association rules. At the same time. So, to achieve the above goals, we present the framework of mining abnormal patterns on multiple correlative time series. The framework can carry out active decision-making in real-time data warehousing.