Learning models of plant behavior for anomaly detection and condition monitoring

Providing engineers and asset managers with a tool which can diagnose faults within transformers can greatly assist decision making on such issues as maintenance, performance and safety. However, the onus has always been on personnel to accurately decide how serious a problem is and how urgently maintenance is required. In dealing with the large volumes of data involved, it is possible that faults may not be noticed until serious damage has occurred. This paper proposes the integration of a newly developed anomaly detection technique with an existing diagnosis system. By learning a hidden Markov model of healthy transformer behavior, unexpected operation, such as when a fault develops, can be flagged for attention. Faults can then be diagnosed using the existing system and maintenance scheduled as required, all at a much earlier stage than would previously have been possible.