Super Spatial Resolution Raman Distributed Temperature Sensing via Deep Learning
Hao Wu, Can Zhao, Ming Tang
Abstract
Raman distributed temperature sensing (RDTS) can obtain temperature information along with an optical fiber. Its spatial resolution is defined by the minimum spatial range that can be resolved by the system. The spatial resolution can be improved by reducing the optical pulse width. However, simultaneously, the signal-to-noise ratio is degraded, and higher speed equipment is required. Recently, deconvolution algorithms have been employed to improve spatial resolution without hardware modification. However, conventional deconvolution algorithms have limited effectiveness. Here, we propose and experimentally demonstrate an RDTS super spatial resolution deep convolutional neural network (SSRNet) to improve the spatial resolution with high fidelity. The convolution kernel of the RDTS is estimated from the falling signal at the fiber end rather than using the waveform of pulsed light. Then a modified RDTS model is built to generate a large amount of training data. Through optimizing the structure of SSRNet, the spatial resolution is increased from 4 m to 0.8 m which has reached the limit of the sampling rate of 250 MSa/s. The simulation and experimental results show that the restored results using SSRNet are more accurate with fewer artifacts than the results using a conventional deconvolution algorithm.