Abstract

This article presents a machine learning solution to identify the peak wavelengths of fiber Bragg grating (FBG) sensors multiplexed in a network with high spectral overlapping. Machine learning solutions generally require high-quality, extensive datasets, which can be difficult to obtain. In contrast, the proposed model corresponds to a convolutional neural network (CNN) autoencoder that can be trained in an unsupervised manner using only FBG spectrum data, without the need for information about the spectral positions of the sensors. The model is specifically designed to encode the spectral positions of overlapping FBG sensors as Dirac deltas, facilitating the straightforward extraction of spectral positions by computing their center of mass. The effectiveness and precision of the model are validated using both simulation and experimental data of a two-sensor serial array. In simulations, the model demonstrates promising adaptation capability, outperforming methods reported in the literature by over an order of magnitude in terms of mean absolute error (MAE). Meanwhile, when evaluated with experimental sensor data, the proposed autoencoder matches the performance of one of the most effective existing methods, but employing a much more efficient computing approach, thereby offering the potential for real-time inference of the spectral position of highly overlapping FBG sensors.