Power transformers' condition monitoring using neural modeling and the local statistical approach to fault diagnosis

On-line monitoring of electric power transformers can provide a clear indication of their status and ageing behavior. This paper proposes neural modeling and the local statistical approach to fault diagnosis for the detection of incipient faults in power transformers. The method can detect transformer failures at their early stages and consequently can deter critical conditions for the power grid. A neural-fuzzy network is used to model the thermal condition of the power transformer in fault-free operation (the thermal condition is associated to a temperature variable known as hot-spot temperature). The output of the neural-fuzzy network is compared to measurements from the power transformer and the obtained residuals undergo statistical processing according to a fault detection and isolation algorithm. If a fault threshold (that is optimally defined according to detection theory) is exceeded, then deviation from normal operation can be detected at its early stages and an alarm can be launched. In several cases fault isolation can be also performed, i.e. the sources of fault in the power transformer model can be also identified. The performance of the proposed methodology is tested through simulation experiments.