Estimação de Séries Temporais via Rede NARX em Aplicações Industriais

Abstract

The widespread use of automation in industry has led industrial processes to be controlled from information coming from sensors, in the form of time series. On the other hand, the high cost of sensors and the high maintenance requirement, due to the need for frequent calibration, motivated studies in the field of forecasting time series. A method that has gained prominence in this field of study is the prediction of time series through the nonlinear autoregressive neural network with exogenous inputs. This neural network integrates the use of time delay and recurrence to capture short-term memory, which allows the extraction of sequential information from its predictors. In this work, this network is applied in forecasting time series of interest to the industry. The first problem addressed corresponds to the application of the neural network for thermal monitoring at critical points of a permanent magnet synchronous motor, considering that the usual techniques are costly and/or difficult to implement. The second case refers to the validation of oil refinery reactors' temperature sensors. The results are evaluated by calculating the mean square error. Performance regarding robustness, filtering, and spillover was also evaluated. The performance evaluation also includes a comparison between the time series prediction methods obtained through artificial intelligence and kernel regression. Finally, satisfactory results were found, demonstrating the success of the application of the NARX network in the proposed problems.