ANN Based Multi-Stage Binary Classification Technique for Incipient Fault Diagnosis of Oil Immersed Transformer

This paper presents an intelligent binary, multi-stage fault classification technique for power transformers based on the selectively hybridised conventional dissolved gas analysis (DGA) methods. These methods are Roger's ratio method, IEC 60599 method and Doernenburg's method. Binary output ANN models are constructed for each ratio method independently and for each stage of fault classification. The learning of these ANN counterparts is based on the conventional DGA methods' prescribed rule base. The published data from the faulty equipment inspected in service is used to assess the performance of the developed ANN models. Based on the per-stage comparative assessment of these methods, a hybrid incipient fault diagnostic model is developed in which the most accurate of the DGA-based ANN binary classifiers are elected for each particular stage of classification. The stages of classification start from ascertaining the presence or absence of fault to identifying the type and severity of the fault at later stages. The test results show that the ANN based binary classification technique and the hybridization approach result in an excellent diagnosis in terms of accuracy, consistency and complexity for the incipient fault diagnosis of power transformers.