An automated classifier for asynchronous diagnosis of partial discharge defects

One of the obstacles to widespread deployment of phase-resolved partial discharge (PD) diagnostics is the complexity of deploying and operating new monitoring systems. Typically, these rely upon synchronization with the electrical system frequency via a power line connection, introducing cost, inconvenience and risk into the system. For utilities to exploit partial discharge monitoring on a wide scale, a new approach is required that can work asynchronously which, looking forward, can be combined with wireless sensor network technology. This paper details some investigations of a novel PD pulse diagnosis method which obviates the need for power frequency synchronization, using an automated classifier trained on laboratory- generated PD pulse data, specifically intended for operation within the resource constraints of a wireless sensor node. Comparative results from a number of machine learning techniques and training parameters demonstrate that the defect classification accuracy can be optimised to 82% based upon the training data used.