Reducing Strain Measurements in Brillouin Optical Correlation-Domain Sensing Using Deep Learning for Safety Assessment Applications

Abstract

Distributed Brillouin sensors have emerged as efficient tools for monitoring strain and temperature distributions within large structures and materials. Among various types of distributed Brillouin sensors, Brillouin optical correlation domain analysis (BOCDA) is inherently a point sensor, providing accessibility to arbitrary positions. While BOCDA systems offer unique advantages such as high spatial resolution and random accessibility, acquiring a full distribution map typically requires a long measurement time, as each measurement by the system corresponds to a single sensing point. In this paper, we propose, for the first time to our knowledge, a deep learning-based signal analysis to reduce the number of measurements in the BOCDA system to one-fifth while maintaining the same number of sensing positions and ensuring an accuracy of at least 94.8%. We present a multiple-scale, multiple-output 2D convolutional neural network that can simultaneously estimate stains of five distinct positions using just a single BOCDA signal. Training artificial neural networks for high-multiplicity classification is challenging, particularly when working with uniformly distributed data and a limited amount of ground truth data, such as only a hundred samples. Moreover, training deep neural networks from scratch with such limited ground truth data is infeasible. To overcome these issues, we employ transfer learning to train the proposed neural network using a synthetic dataset generated through a BOCDA measurement simulation program and ground truth data. From only a single measurement, the 2D CNN precisely estimates strains or Brillouin frequency shifts at five different locations, achieving accuracies of 96.52%, 97.55%, 98.02%, 94.79%, and 96.14%. IEEE