Abstract

This paper presents a machine learning (ML) solution to detect the peak wavelength of fibre Bragg grating (FBG) sensors multiplexed with overlapped reflection spectra, and using a serial topology. ML solutions generally require high-quality, high-volume datasets, which can be difficult to obtain in some scenarios. In contrast, here the proposed model is a sparse autoencoder convolutional neural network that can be trained using only the joint reflection spectrum containing all multiplexed FBGs, without information on the spectral position of each sensor. The technique is first verified through simulations and then with experimental data using two wavelength-multiplexed FBG sensors in series. Comparing with existing methods, results verify that the proposed model has promising adaptation capability under multiple simulation scenarios, outperforming existing methods, whilst the model matches one of the best existing approaches when using experimental data.