(191c) Application of the Matrix Profile for Efficient Large-Scale Data Mining of Operational Time Series Data in Process Manufacturing

The increasing rate of collection of time series data from sensor networks and control systems during the operation of process manufacturing plants motivates the utilization of efficient computational approaches to analyze this data. A novel technique for analyzing time series data, the computation of a meta time series called the matrix profile, has recently been developed [1]. This approach annotates a time series with the distance of all subsequences from their nearest neighbor. By efficiently computing the matrix profile common tasks in time series data mining become computationally trivial.

The tasks of identifying frequently recurring patterns, finding outlying subsequences, and semantic segmentation of time series data streams are considered in this paper. These tasks have been shown to be implementable as algorithms utilizing the matrix profile. Identification of outlying subsequences supports fault detection, and the identification of recurring patterns or division of time into operating regimes is valuable for process understanding and visualization. The utility of the matrix profile in these engineering and plant operations relevant tasks is demonstrated by evaluating the performance with data generated from a rigorous dynamic simulation of a chemical plant model for vinyl acetate monomer production [2].

References

[1] Yeh, Chin-Chia Michael, et al. "Matrix profile I: all pairs similarity joins for time series: a unifying view that includes motifs, discords and shapelets." Data Mining (ICDM), 2016 IEEE 16th International Conference on. IEEE, 2016.

[2] Machida, Yuta, et al. "Vinyl acetate monomer (vam) plant model: a new benchmark problem for control and operation study." IFAC-PapersOnLine 49.7 (2016): 533-538.