High precision temperature measurement for cryogenic temperature sensors based on deep learning technology

Nonlinear error compensation is a significant factor that affects the measurement accuracy of temperature sensors. Cryogenic temperature sensors require a precise calibration technique to achieve accurate temperature measurements. BP (Back Propagation) neural networks are not suitable for sensor temperature prediction due to slow convergence and poor learning ability. In this study, a Chebyshev polynomial-based BiLSTM (Bi-directional Long Short-Term Memory) algorithm (C-BiLSTM) is proposed to improve the accuracy of measurement. Firstly, we demonstrate vacuum packaged temperature sensors based on zirconium oxynitride (ZrO_xN_y) thin films with ultra-high sensitivity at cryogenic temperatures. Secondly, a dataset consisting of sensor resistance and temperature values was obtained from five above-mentioned sensors, which contains 29 calibration points in the temperature range of 16 K-300 K measured by the Calibration System. Then, the dataset was divided into two temperature ranges (16 K-54.358 K and 40 K-300 K). In the range 16–54.358 K, 14 set-points are selected as training set and 3 set-points as testing set. In the range 40–300 K, 12 set-points are selected as training set and 2 set-points as testing set. Thirdly, a neural network model was built and trained using the TensorFlow framework. By comparing C-BiLSTM we proposed with the BP neural network and BiLSTM, the results show that the C-BiLSTM model converges faster and greatly improves the prediction accuracy after adding Chebyshev polynomial features. The fitting error and prediction error are less than 1 mK in the temperature range of 16 K-40 K. They can also keep less than 10 mK even at the wide temperature range of 40 K-300 K, which is a significant improvement respect to improving the accuracy of temperature measurement.