In order to operate thermal power plants safely, early detection of equipment failure signs is one of the most important issues. To detect the signs before an alarm is issued in the existing monitoring system, we developed a fault diagnosis system based on the Adaptive Resonance Theory (ART). The vigilance parameter, which is a design parameter in the ART model, was shown to influence the diagnosis accuracy. Fixing the value of the vigilance parameter also had problems: we needed to use time-consuming trial and error, and we needed to have empirical knowledge of the parameter tuning. In this paper, using simulations we demonstrated the relationship between the vigilance parameter and diagnosis accuracy. Furthermore, to overcome the problems of the vigilance parameter tuning, we have proposed an auto tuning algorithm to make the parameter the optimum value. The performance of the proposed algorithm was evaluated in several case studies using gas turbine plant data. The effectiveness of the proposed algorithm was confirmed by the obtained results.

No abstract

No abstract

Early detection of anomalies is crucial for maintaining high productivity in industrial plants. To meet that requirement, anomaly detection systems based on adaptive resonance theory (ART) have been developed. Although various anomaly detection methods based on ART have been proposed, their performances have not yet been evaluated systematically. A new anomaly detection criterion is proposed, and the performances of ART-based anomaly detection systems, using di erent anomaly detection criteria and di erent anomaly detection structures, are evaluated. The performance evaluation results show that distributed model based systems, which use an ART model for each part of the plant, attain higher anomaly detection performance than that of systems using an ART model for the entire plant. They also show that an anomaly detection system using a new anomaly detection criterion, based on the distance between samples, attains almost the same anomaly detection performance as that of a system using generation of new categories.

No abstract