Development of optical fiber strain sensor system based on machine learning and polarization

Abstract

Based on the principle that the polarization state of light propagating in a single-mode fiber changes with external strains, an optical fiber sensor system based on machine learning and polarization for multi-point strain measurement is proposed. To address the influence of the front sensor on the rear sensor and to minimize interference from unrelated inputs, we have employed a data processing method that constructs an individual neural network model for each sensor. This approach uses the polarization state of the reflected light of the sensors as the neural networks’ input and the sensors’ rotation angles as the output, training the designed neural networks for learning. The trained neural networks produce predicted outputs that demonstrate high consistency with the experimental data, achieving an average prediction accuracy of 99% on test data. These results validate the effectiveness of our sensor system and data processing method.