Cost benefit of using a committee of parallel neural networks for bushing diagnostics

This paper presents a cost benefit analysis of applying an ensemble of parallel artificial neural networks (ANN) compared to an entirely human decision process. The comparison is based on a committee of ANN that was successfully able to diagnose the condition of bushings using IEEE C57.104 criteria taking fourteen variables of dissolved gas analysis (DGA) data for each oil impregnated paper bushing. The works compares the speed, stability and accuracy of a human to that of the collective parallel artificial neural networks (ANN) made of radial basis function (RBF), support vector machines (SVM), multiple layer perceptron (MLP) and Bayesian (BNN) networks. The analysis on 1255 bushings concludes that collective network is a more cost effective solution than the human alone, in deciding whether to remove or leave a bushing in service. The accuracy of the human was 60% in 16 hour to diagnose 1255 bushings. This was slightly less than that of the committee of ANN which produced an accuracy of 99% in 35 minutes with a 99% reliability, dependability of 99%, and availability of 80%. Giving an overall performance of 78% for the ensemble diagnosing 1255 bushings. The realisable return is when using the technology is a modified internal rate of return (MIRR) of 19.27% and a profitability index (PI) 3.1, a net present value in 2004 of R910 946 and a discounted payback period of 2.0 years.