Improvement of the strain measurement accuracy via a denoising convolutional neural network in the OFDR system

We propose an improved cross-correlation strain demodulation method of the optical frequency domain reflectometry (OFDR) to further enhance the measurement accuracy in dual-requirement scenarios of high spatial resolution and large dynamic strain measurement range. Rather than a conventional signal processing method based on one-dimensional (1D), we transform the distributed global spectrum shifts obtained through cross-correlation calculation along the sensing fiber as a two-dimensional image matrix and employ a denoising convolutional neural network to smoothen the systemic artifacts manifested as cross-correlation fake peaks due to the local similarity degradation between the obtained reference (Ref.) and measured (Mea.) spectra, thereby enabling the reconstruction of accurate strain gradient profiles. Experimental results reveal that this performance enhancement achieves a fivefold and twofold improvement, from 19.63% and 49.53% to 98.27%, respectively, in demodulation accuracy over the 1D moving average smoothing and 1D convolutional neural network methods under an employed strain of 300 µɛ and without hardware modifications. Meanwhile, the strain profiles across both zero-strain and stretched regions at a consistent spatial resolution of 16 mm are measured accurately and clearly. The measured strain linearly changes with the applied strain, with a slope and R² of 1.03 and 0.99, respectively, when the applied strain is increased from 100 to 900 µɛ, which confirms the efficacy of our proposed scheme in addressing traditional OFDR demodulation challenges.