时间序列对称模式挖掘

随着信息化和工业化的融合,物联网和工业互联网蓬勃发展,由此产生了以时间序列为代表的大量工业大数据.时间序列中蕴含着很多有价值的模式,其中,对称模式在各类时间序列中广泛存在.挖掘对称模式对于行为分析、轨迹跟踪、异常检测等领域具有重要的研究价值,但时间序列的数据量往往高达几十甚至上百GB.使用直接的嵌套查询算法挖掘对称模式...

Time Series Symmetric Pattern Mining

With the integration of informatization and industrialization, the Internet of Things and Industrial Internet are booming, resulting in a large amount of industrial big data represented by time series. There are many valuable patterns in time series, among which symmetric patterns are widespread in various time series. Mining symmetric patterns has important research value in the fields of behavior analysis, trajectory tracking, anomaly detection, etc. However, the amount of time series data is often as high as tens or even hundreds of gigabytes, and direct nested query algorithms may take several months or even several years, and typical acceleration techniques such as indexing, lower bounds, and triangular inequalities can only produce an acceleration of one or two orders of magnitude at most. Therefore, based on the inspiration of the dynamic time warping algorithm, this paper proposes a method that can mine all the symmetric patterns of the time series within the time complexity of O(WX|T|). Specifically, given the symmetric pattern length constraint, the symmetric subsequences can be calculated based on the interval dynamic programming. Then the most non-overlapping symmetric subsequences can be selected according to the greedy strategy. In addition, this paper also studies the algorithm for mining symmetric patterns in the time series data stream, and adjusts the window size to ensure the integrity of the symmetric pattern data according to the data in the window the feature dynamically. Using one artificial data set and three real data sets to experiment with the above method under different data volumes, it can be seen from the experimental results that compared with other symmetric pattern mining methods, this method has a good performance in terms of pattern mining results and time overhead.