Enhanced Raman Distributed Temperature Sensor Based on Self-Constructed Fully Connected Neural Network

The performance of Raman Distributed Temperature Sensing (RDTS) is greatly limited by the effect of optical and other random noise on the backscattered Anti-Stokes signal. Therefore, it is necessary to study the signal denoising method of the system. Existing methods to improve system performance mainly include improving the system structure and using denoising methods like cumulative averaging, wavelet transform, etc., which still have room for improvement. In this paper, a Fully-Connected-RDTS model based on fully connected neural network is designed to reduce temperature errors and improve spatial positioning accuracy. Compared with traditional denoising algorithms, the maximum temperature error is reduced to 0.53 °C (±0.25 °C) using only 500 averages of the temperature data, which is better than 2.61 °C of 10,000 cumulative averages, and 2.25 °C of wavelet transforms. The predictions for the boundary of the temperature curve are highly consistent with the original data, and the error of spatial positioning accuracy is less than 1 m without blurring the spatial information like cumulative averaging and wavelet transform algorithms. Compared with the results of 10,000 cumulative averages, the detection response time is reduced from 8 s to 1-s while the noise is greatly filtered. The measurement performance of the RDTS system is greatly improved by the Fully-Connected-RDTS model without sacrificing spatial accuracy. And once the model is successfully trained, there is no need to change the parameters of the model, so it could be used for large-scale real-time detection.