Physical Model-Assisted Hybrid Neural Network Enabling High Spatial Resolution Distributed High-Temperature Sensing With Single-Crystal Fiber

Raman distributed high temperature sensing offers unique advantages, particularly in addressing crosstalk issues of the traditional point sensors, as well as the challenges associated with multi-point deployment of sensors and complex wiring in temperature field measurements. However, due to limitations of the system hardware, achieving ultra-high spatial resolution in distributed high-temperature sensing systems remains challenging, thereby hindering the acquisition of dense temperature data required for comprehensive temperature field measurements. To break through these intrinsic hardware limitation and challenges, we proposed a hybrid neural network model combined with convolution neural network (CNN) and bidirectional long short-term memory (Bi-LSTM) for the distributed high temperature sensing system with the specialty yttrium aluminum garnet single crystal fiber. Based on experimental data acquired from room temperature to high temperature of 1400°C, the spatial resolution was shown to improve significantly, and further validation through point heat source experiments demonstrated a maximum enhancement in spatial resolution of 4.43 cm. The hybrid model of CNN and Bi-LSTM effectively achieves both denoising and spatial resolution enhancement. In addition, the optimized hybrid model can reconstruct high-fidelity signal traces that align closely with the experimental system. Furthermore, the system is capable of processing and reconstructing frame-by-frame signals, enabling high-speed temperature response capabilities.