High Spatial Resolution Implementation Method for OFDR System Based on Convolution Neural Network

In this study, a high spatial resolution implementation method for optical frequency-domain reflectometry (OFDR) system based on convolution neural network (CNN) is proposed and experimentally demonstrated. In the data processing flow, the wavelength shift information obtained by cross-correlation operations between the measurement signal and the reference signal for each fiber segment is arranged as a function of the sensing distance into 2-D images, and when the spatial resolution is increased, the cluttered regions present in the wavelength shift information can be equated to the image noise in the 2-D images. CNN model is trained by using low-noise real data measured under multiple sets of strains as well as simulated noise-free data, which was used to suppress image noise from real data at high spatial resolution. With no modification on the OFDR hardware system, strain gradient information is accurately recovered at an effective sensing distance of 75 m with spatial resolution up to 2 mm by denoising with CNN. The mean value of mean absolute error (MAE) for the strain information after denoising with CNN has been reduced to $8.2751 \mu \varepsilon $ compared with the raw data without any denoising method applied of $190.6653 \mu \varepsilon $ , which is better than $13.1792 \mu \varepsilon $ of the traditional Gaussian filter (GF). The mean value of root-mean-square error (RMSE) has been reduced to $10.4029 \mu \varepsilon $ , which is better than $16.5762 \mu \varepsilon $ of the GF. The mean standard deviations (MSDs) of the measured strain gradient along the sensing fiber length for the proposed method is $8.9848 \mu \varepsilon $ , which is reduced by 43.43% compared to the MSDs when the traditional GF method is used, showing better smoothness of the recovered strain information. The experimental results show that the proposed method provides a potential solution for long-range strain sensing applications with high spatial resolution.