COMBINED NEURAL NETWORK METHOD FOR DETERMINING THE BRILLOUIN SPECTRUM MAXIMUM IN DISTRIBUTED FIBER-OPTIC SENSORS

Abstract

Distributed fiber-optic sensors are becoming an increasingly common solution for monitoring and diagnosing extended information transmission lines, industrial devices, buildings and devices. One of the directions – Brillouin reflectometry, which allows diagnosing a fiber line for changes in ambient temperature or mechanical deformation, is becoming increasingly popular for engineers and researchers. However, modern standards impose increasingly strict requirements on diagnostic systems in terms of the accuracy of determined parameters. For Brillouin reflectometry, where the value of the environmental parameters is determined by the position of the maximum of the Brillouin gain spectrum, the task of more accurately determining this maximum becomes the main one. The paper considers modern computer-computational methods for detecting the maximum of the Brillouin gain spectrum in an optical fiber. The authors note that imperfections in the shape of the optical spectrum, such as the signal-to-noise ratio, as well as possible digital defects that occur during digitization, can significantly impair the accuracy of the system. The authors consider three approaches to detecting the maximum of the spectrum: the classical method of Lorentz curve fitting, the method of cross-correlation with the ideal Lorentzian function, as well as the method of inverse correlation developed by the authors earlier. To combine the results of the work of the three methods, a neural network model was developed that accepts the input data of each method, together with the parameters of noise and distortion of the spectrum. The model is presented in the form of a four-layer perceptron with two hidden layers. As a result, the authors achieved an increase in the accuracy of determining the position of the maximum of the spectrum by 10% on the model data.