Neural networks on transformer fault detection evaluating the relevance of the input space parameters

Following a number of studies that have employed different forms of neural network models to perform dissolved gas-in-oil analysis (DGA) of transformer bushings, this manuscript focuses on evaluating the relevance of the parameters that form part of the model input space. Using a multilayer neural network initially populated with all the 10 input parameters (10V-Model), a matrix containing causal information about the possible relevance of each input parameter is obtained. The information from this matrix is proven to be valid through the construction and testing of another two, separate, multilayer networks. One network´s input space is populated with the 5 most relevant parameters (MRV-Model), while the other is populated with the 5 least relevant parameters (LRV-Model). The obtained classification accuracy values are as follows: 100% for the 10V-Model, 98.5% for the MRV-Model, and 53.0% for the LRV-Model.