True physics informed hybrid neural network based linearization for high accuracy thermocouple temperature measurements

Thermocouples (TC) are used in process industries for measurement of temperature due to their wide range and fast response time. However, TCs have a nonlinearity of 13.34%, junction-temperature dependency, and a low accuracy of ± 4.4 °C. They are mainly used in rugged environments. This work introduces a novel true physics informed hybrid neural network (PIHNN) for improving the characteristics of T type thermocouple in −270 °C to 400 °C range, which is integrated to an op-amp based resistive amplifying and compensating signal conditioning circuit (ORACSCC) and microcontroller. The simple design of ORACSCC, a physics-based model, and the Bayesian regularization (BR) algorithm, help achieve balanced real-time results. The experimentation resulted in reduction of nonlinearity from 13.3% to 0.15% in full-scale output (FSO). The ORACSCC-PIHNN system attained an accuracy of 0.024 °C (FSO), a sensitivity of 0.999, an inference time of 0.08 ms, and a latency time of 0.20 ms. No signs of overfitting and underfitting was detected. For this real-time research prototype, these results are excellent compared to interpolation and polynomial methods. The best performance is produced by BR algorithm in 1–35-25–15-1 configuration, with a mean square error (MSE) of 0.055 °C. Training time used is 49.46 sec for 671 data points. Consistent results with a maximum drift of ± 0.04 °C is obtained in real time testing with ESP32 microcontroller. The true PIHNN’s based regularization compensated noise, stray and ambient effects without much performance losses. The ORACSCC-PIHNN system will provide accurate and precise measurements to instrumentation systems in the oil and gas industry.