(19g) Semi-Supervised Anomaly Detection for Production Oil Wells

Anomaly detection and diagnosis (ADD) plays an important role in preventing process faults and failures. ADD is a challenging area of process monitoring due to lack of a priori information regarding anomalies. Traditionally, either fault class information is required or an unsupervised (i.e., one-class [1]) approach is adopted. Process monitoring as applied to modern chemical systems is further complicated due to processes that operate continuously, perhaps in several modes, with feedback control. To tackle these issues, we extend the semi-supervised approach of Blanchard et al. [2] for static data to accommodate time series data. Motivated by the work of Fulcher and Jones [3], an expert feature space is proposed as the input data. A likelihood ratio test is employed with the nominal and anomalous distributions modeled using kernel density estimates with Gaussian kernels. Early operations are leveraged as semi-supervised training data under the premise that nominal operating days are more easily labeled than anomalous days. Results are visualized using a monitoring plot similar to a contribution plot [4-6]. The success of the approach is demonstrated on operational data collected from real production oil wells.

[1] D. M. J. Tax (2001) One-class classification. Ph.D. Thesis, Delft University of Technology.

[2] G. Blanchard, G. Lee, and C. Scott (2010) Semi-supervised novelty detection. Journal of Machine Learning Research, 11, 2973-3009.

[3] B. D. Fulcher and N. S. Jones (2014) Highly comparative feature-based time-series classification. IEEE Transactions on Knowledge and Data Engineering, 26, 3026-3037.

[4] J. A. Westerhuis, S. P. Gurden, and A. K. Smilde (2000) Generalized contribution plots in multivariate statistical process monitoring. Chemometrics & Intelligent Laboratory Systems, 51(1), 95-114.

[5] X. Zhu and R. D. Braatz (2014) 2D contribution map for fault identification. IEEE Control Systems, 33(5):72-77.

[6] B. Jiang, D. Huang, X. Zhu, F. Yang, and R. D. Braatz (2015) Canonical variate analysis-based contributions for fault identification. Journal of Process Control, 26, 17-25.