Thermocouple signal conditioning using artificial neural networks.

Abstract

Thermocouples are probably the most widely used temperature sensing devices in industrial applications. This is due to their relatively high accuracy. Thermocouples sense temperature using thermoelectric voltages arising due to temperature differences between the hot and cold junctions of the thermocouple. The generated thermoelectric voltage is nonlinear in form. Linear approximations in the conversion of thermoelectric voltages into temperature readings compromise the accuracy of the derived temperature values: requiring further processing of the thermocouple voltage for improved temperature measurements. Moreover, undetected variations in the cold junction temperature could further worsen the accuracy of the temperature sensor. The current study researched the enhancement of the accuracy of thermocouple temperature measurement subjected to both random variations in the reference junction temperature and nonlinearities, with a validation of the design process using T, R, E, and J, thermocouples. To this end, the ITS-90 thermocouple tables based on a fixed 0°C reference junction temperature were not adequate for use in the study, so the thermocouple polynomial equations for the T, R, E, and J thermocouples were simulated in MATLAB, with randomly generated cold-junction temperature values, to produce augmented ITS-90 tables for the four thermocouples studied. Results show that the augmented thermocouple tables accurately compared with the ITS-90 tables when the reference junction temperature was set to 0°C. Data samples were generated from each of the augmented thermocouple tables for neural network studies. Half of the data samples for each of the thermocouples was used to train ‘table-lookup’ Multilayer Perceptron (MLP) neural networks in MATLAB. Each neural network used the cold-junction temperatures and thermoelectric voltages as inputs, while the corresponding hot-junction temperatures were used as the target outputs. The validation process for the augmented ITS-90 thermocouple tables showed that the E, T, R, and J thermocouples could all reproduce the hot junction temperature within 0.01% of the results found on the ITS-90 tables. The performance results for the neural networks showed that the E-type thermocouple neural network has a worst-case error within 0.2% in reproducing the hot junction temperature. The J-type thermocouple neural network showed a worst-case error within 0.1%, while the T and R-type thermocouple neural network produced worst error case within 0.04% of the results generated by the augmented ITS-90 tables. For the practical validation of the development presented in this thesis, the structure of each of the trained MLP neural networks was coded as a subroutine within an Arduino Uno microprocessor. The hot junction of the thermocouple was placed in a TTM-004 controller or oven. The cold junction of the thermocouple was located in the ambient of the used laboratory and monitored by an LM 35 temperature sensor connected to one of the inputs of the microcontroller. The experimental results showed that temperature of the TTM-004 controller or oven was evaluated to within 2%, 4% and 3% by the signal conditioning unit using T-type, J-type and E-type thermocouple respectively.