Multiple incipient fault classification approach for enhancing the accuracy of dissolved gas analysis (DGA)

Multiple incipient faults are practically known to exist in transformers. They tend to produce suddenly changing ratio limits in ratio-based methods or oscillation of fault location in graphical methods. In consequence, the energy associated with them lies in-between low and high severity single faults. Hence multiple fault detection needs to be addressed appropriately which may otherwise pose the serious constraints during transformer condition monitoring. In this study, novel and intelligent classification approach is proposed to upgrade the classical dissolved gas analysis (DGA) technique to cater the requirement of multiple fault diagnosis. This consists of Duval-triangle-based optimised fuzzy inference system and neural network models sensitive to both single and multiple incipient faults. Both models have been rigorously trained and tested using dataset credited to field and literatures to achieve high fault recognition and isolation rates, alternatively low false detection and no-detection rates. Both parameters are combined into single index to determine the accuracy in terms of F1 score which is evaluated to be >97%. The diagnostic ability of the scheme is highly promising and can improve reliability of transformer fault forecasting by DGA.