Online conditional anomaly detection in multivariate data for transformer monitoring

Retrofitting condition monitoring systems to aging plant can be problematic, since the particular signature of normal behavior will vary from unit to unit. This paper describes a technique for anomaly detection within the context of the conditions experienced by an in-service transformer, such as loading, seasonal weather, and network configuration. The aim is to model the aged but normal behavior for a given transformer, while reducing the potential for anomalies to be erroneously detected. The paper describes how this technique has been applied to two transmission transformers in the U.K. A case study of 12 months of data is given, with detailed analysis of anomalies detected during that time.