Improving the Spatial Resolution of a BOTDA Sensor Using Physics-Enhanced Deep Neural Network

A high spatial resolution Brillouin Gain Spectrum (BGS) reconstruction method combining a neural network and a physical model is designed and proposed for Brillouin optical time domain analysis (BOTDA) systems. This approach achieves sub-meter spatial resolution directly from a single long pulse experimental data through self-supervised learning. To the best of our knowledge, this is the first time a physical model has been used to enhance neural network models in the field of optical fiber sensing. Simulation results show that the BGS distribution reconstructed by this technique has a mean squared error (MSE) of 0.02 compared to the ideal BGSs. And the standard deviation of the extracted BFS is 0.65 MHz. Experimental results indicate that, for a 40 ns pulse BGS distribution, the proposed approach accurately reconstructs a BGS distribution with 0.5 m spatial resolution. The proposed technique demonstrates a significant advantage over results from the deconvolution algorithm and supervised learning methods.