Abstract

This article proposes a novel method of demodulating optical Vernier effect spectra based on neural networks. To achieve the sensitivity enhancement based on optical Vernier effect, two Sagnac interferometers with similar free spectral ranges (FSRs) are cascaded using six-hole suspended-core fiber (SH-SCF), whose pressure sensitivity can be enlarged by 5.53 times compared to a single interferometer. Instead of tracing the spectrum manually, we used three neural networks: long short-term memory (LSTM) network, convolutional neural network (CNN), and CNN+LSTM to demodulate the full spectra with Vernier effect enhancement. The results show that the root mean squared error (RMSE) of unknown spectra based on the well-trained CNN model is only 1.17 × 10⁻² MPa, which is better than that (7.49 × 10⁻² MPa) using LSTM model. Moreover, such a result could be further improved to 3.68 × 10⁻³ MPa after adding LSTM layers to CNN model. The average training and testing time of CNN and CNN+LSTM models are only 11 ms and 15 ms, respectively, showing a good real-time ability of spectral demodulation. To demonstrate the universality of the proposed model, a similar sensing structure is constructed using a polarization-maintaining photonic crystal fiber (PM-PCF), and comparable results are achieved although the sensing fiber is different. The proposed method shows high potential in applications using pressure sensing systems with high sensitivity improved by the Vernier effect and fast spectral demodulation of neural networks.