Bayesian Regularizers of Artificial Neural Networks applied to the reliability forecast of internal combustion machines in the short-term

Abstract

Predictive as well as preventive maintenance are tools of maintenance programs that aim to increase or maintain the life expectancy of an equipment through computational techniques and tools. Bearing in mind that the power generation industry has a high maintenance rate with machines and / or electric generators stopped, this research aims to develop a computational model for predicting the Reliability Key Performance Indicator (KPI) to identify how available the equipment will be in a time span of 22 days, for this the methodology to be used will be based on analyzes and tests of artificial neural network (ANN) architectures using the Bayesian Regularizers training algorithm, alternating the transfer functions in the layers hidden to find the best state of convergence and the minimum Root Mean Square Error (RMSE) value calculated between the real and simulated outputs. According to the results obtained by the training, validation and test steps, the algorithm presented a RMSE rate of 0.0000104202 and a 99.9% correlation between the real and simulated values, thus the model is able to identify which machine will have the greatest efficiency and less efficiency within the defined time span.