High-Performance Raman Distributed Temperature Sensing Powered by Deep Learning

Abstract

Raman-based distributed temperature sensing (RDTS) can achieve temperature measurement of tens of kilometers and is widely used in temperature monitoring of crucial facilities. The accuracy of RDTS is highly determined by the signal-to-noise ratio (SNR) of the acquired spontaneous Raman scattering (SpRS) signals, which is about 60 dB weaker than the pump pulse. Therefore, the performance of single-mode fiber (SMF) based conventional RDTS is poor. To improve its performance, many methods have been proposed, including the use of special optical fibers, pulse coding techniques, and denoising algorithms. However, these methods have their limitations. Here, we propose and experimentally demonstrate a deep 1-D denoising convolutional neural network (1DDCNN) to enhance the performance of RDTS. A simplified RDTS model is built to train and optimize the 1DDCNN. To verify the performance of the 1DDCNN on actual data, we experimentally measure the SpRS data of a 10-km SMF with an average time of 1 s and a spatial resolution of 3 m. The well-trained 1DDCNN reduces the temperature uncertainty from 6.4 °C to 0.7 °C with high fidelity. As a comparison, the commonly used wavelet denoising algorithm can only reduce it to 3.7 °C. Besides, the 1DDCNN does not require manual parameter adjustment and is not affected by the sampling rate, giving it significant advantages in practical applications compared with conventional denoising algorithms. Moreover, we believe that the proposed 1DDCNN can be applied to other distributed optical fiber sensing systems by fine-tuning the network with appropriate training data.