This paper describes a cluster assessment (CA) method for automatic detection, assessment, and logging of significant fault gas production events by analysis of multi-gas online monitor data. When large numbers of transformers are monitored, automatic screening interpretation of the data is necessary. The automated interpretation of dissolved-gas data must be quite skilled at discriminating between exceptional and unexceptional patterns to provide high sensitivity and high specificity, i.e., to detect incipient problems reliably while generating very few false alarms. The CA method is applied to a moving time window
of the most recent 30 to 90 days of multi-gas monitor data for each transformer. It relies upon three innovative elements:
1. Statistical estimation of measurement uncertainty from in-service data. A paper deriving the statistical estimators and evaluating their performance is in preparation.
2. A fault energy index calculated from the gas concentrations. The index is the normalized energy intensity (NEI) described in "Thermodynamic Estimation of Transformer Fault Severity."
3. Cluster analysis applied to the NEI data.
An uncertainty estimate derived from the data is used for adaptive tuning of the cluster analysis and for discriminating between insignificant and significant increases. Clusters and transitions between clusters are subjected to diagnostic assessment. Any that appears to represent the production of a significant amount of new fault gas is reported as an event with a severity score and an apparent fault type. The severity score can be based on the amount of change in survival probability as illustrated in this paper.