(646i) Performance Evaluation of Anomaly Diagnosis System Based on Adaptive Resonance Theory

Early detection of anomalies is crucial for maintaining high productivity at industrial plants. Clustering algorithms are promising methods for detecting anomalies. In particular, adaptive resonance theory (ART) can be used to make a powerful clustering algorithm [1], and anomaly detection systems based on ART have been installed in an industrial waste liquid treatment plant [2], thermal power plants [3], and distillation tower [4].

The anomaly detection systems of these studies were developed using a single model and a single algorithm. When applying ART to anomaly detection and identification problems, various models and algorithms can be used for detection and identification, and the model and algorithm selections affect the detection and identification performance. These effects, however, have not been evaluated yet.

In this study, we examined the performance of anomaly detection and identification of diagnosis systems based on ART by using the Tennessee Eastman process data. For anomaly detection, the effect of model division and multiple use of the ART model on the anomaly detection performance was evaluated. We examined a case in which an ART model was applied to the whole plant and a case in which an ART model was applied to each subsystem of the plant. We also evaluated how the performance of the system changes with the selection of vigilance parameters of the ART model. The performance of anomaly identification was examined by comparing the kinds of anomalies and the categories created by the ART models. The results of a case study indicate that model division can improve the anomaly detection performance and multiple use of ART models can reduce the false alarm rate. The results also show that anomaly diagnosis systems based on ART are applicable to anomaly identification.

References

1. Carpenter, G., and Grossberg, S., ART2: Self-Organization of stable category recognition codes for analog input patterns, Applied Optics, 26(23), 4919-4930 (1987)

2. Yamashita, Y., A Clustered Class Distribution Approach for Process Monitoring and Fault Detection, Chemical Engineering Communications, 191, 302-313 (2004)

3. Sekiai, T., Kusumi, N., Hori, Y., Shimizu, S., and Fukai, M., Auto tuning algorithm for vigilance parameter in the adaptive resonance theory model and its application to fault diagnosis system of thermal power plants, Proc. of the ASME 2011 Power Conference (POWER 2011) and the International Conference on Power Engineering 2011 (ICOPE-11), July 12–14, Denver, Colorado, USA, 227–234 (2011)

4. Hori, Y., Yamamoto, H., Suzuki, T., Okitsu, J., Nakamura, T., Maeda, T., Matsuo, T., Zabiri, H., Tufa, L., and Marappagounder, R., Online anomaly detection of distillation tower system using adaptive resonance theory, Proceedings of 7th International Symposium on Design, Operation and Control of Chemical Processes (PSE Asia 2016), Tokyo (2016)