An artificial neural network to linearize a G (tungsten vs. tungsten 26% rhenium) thermocouple characteristic in the range of zero to 2000°C

An alternative way to correct sensors in order to exhibit good linearity is proposed. This paper describes the design and behavior of an artificial neural network (ANN) with two hidden layers of twelve neurons each for linearizing a static characteristic of a G thermocouple ranged from zero to 2000°C. The use of interpolation methods to perform such corrections is also discussed. In the training phase of the ANN, backpropagation and random optimization algorithms have jointly been used to adjust the weights of the network for a desired final error. After training the ANN, it is then used as a neural linearizer to generate the primary variable (temperature) from the thermocouple´s output voltage. The accuracy of this method is compared with those obtained by interpolation