Gramian angular field encoding enabled ultra-wide-range cryogenic temperature sensing by deep learning.
TL;DR
A deep learning-assisted Gramian Angular Field (GAF) spectral encoding scheme for achieving precise demodulation of wide dynamic range sensing signals is proposed, making the interferometric cryogenic temperature sensing system a competitive candidate for practical deployment in extreme environments.
Abstract
An interferometric cryogenic temperature sensor, featuring anti-electromagnetic interference and chemical corrosion resistance, provides increasing opportunities for precise monitoring in spacecraft, biomedicine, and cryobiology. However, in existing sensing systems operating over a wide dynamic range, spectral overlap arising from the free spectral range (FSR) limitation makes it difficult for conventional strategies to demodulate the sensing signal. Here, we propose a deep learning-assisted Gramian Angular Field (GAF) spectral encoding scheme for achieving precise demodulation of wide dynamic range sensing signals. As a proof-of-concept, a Sagnac interferometer (SI) based on a panda polarization-maintaining fiber (PPMF) is employed as the temperature sensing element. The acquired signals are converted into images using GAF, which are fed in parallel into a dual-channel two-dimensional convolutional neural network (DC-2DCNN) for feature extraction. Notably, the proposed strategy achieves a coefficient of determination (R2) of 0.9986 in the ultra-wide cryogenic temperature range of 325-35 K, and the mean absolute percentage error (MAPE) is 1.7310%. These outstanding performances fully demonstrate the effectiveness of the proposed strategy, making the interferometric cryogenic temperature sensing system a competitive candidate for practical deployment in extreme environments.