Prediction of Dynamic Temperature Rise of Thermocouple Sensors Based on Genetic Algorithm-Back Propagation Neural Network

Abstract

Laser-based dynamic calibration of thermocouple sensors is limited by the improper setting of laser parameters that lead to temperature overshoot damage. To address these problems, a model for predicting the temperature difference between steady-state and initial temperatures (dynamic temperature rise) of a K-type thermocouple was developed using a back propagation (BP) neural network optimized using a genetic algorithm (GA). In this study, a semiconductor laser was used as the temperature excitation source, and the effects of laser power, repetition frequency, and duty cycle parameters on the dynamic temperature rise of the thermocouple were studied. The experimental results were randomly selected for training and testing datasets for the neural network, and the performance of the GA-BP was compared with that of the standard BP neural network. The results showed that the laser power and duty cycle were positively correlated with the thermocouple temperature increase, and laser repetition frequency showed a negative correlation with the thermocouple temperature increase. The root-mean-square error (RMSE) and mean absolute percentage error (MAPE) were selected to evaluate the prediction models. Compared to the BP neural network model, the GA-BP network reduced the RMSE and MAPE by approximately 43% and 5.7%, respectively, and made predictions with an average accuracy of 99%. In the future, this technology can be applied to real engineering problems.