Feedforward neural-network conditioning of type-B thermocouple with variable reference-junction temperature

Thermocouple data come in standard tables and must be interpolated for any readings not directly contained in such tables. Also, variations in the temperature of the reference junction of the thermocouple affect the repeatability of the thermocouple. This paper presents two feedforward neural networks for conditioning the mV output of the type-B thermocouple: one, a two-layer network for structural identification and the second, a radial basis network for repeatability enhancement. The networks were trained in MATLAB. Results show that complete thermocouple data could be reproduced using the logistic network. The radial basis function network was verified to recover true junction temperatures for all simulated variations in the reference junction temperature.