Accurate Strain Sensing With Fiber-Optic Fabry-Perot Sensors Based on CNN-LSTM Model

In this work, we proposed a new method of demodulating the spectrum of fiber-optic Fabry-Perot (FP) strain sensors by employing convolutional neural networks (CNNs) adding long short-term memory (LSTM) networks, hereinafter referred to as CNN-LSTM. The model was built to retrieve the spectra of different strains since the peak tracking algorithm may difficult to accurately find the peaks of spectra and some traditional algorithms have their own limitations, which may cause large errors in demodulation results. The spectra of strain will be normalized first and then sent to the model for training, the strain information can be directly extracted without any peak wavelength or cavity length as an intermediary quantity, which simplifies the demodulation process. Experimental results show that the accuracy represented by the coefficient of determination (R-2) better than 99.99%, and the root mean square error (RMSE) is about 0.524 mu epsilon. Additionally, the model has comprehensive advantages over other models in terms of training cost and accuracy. The proposed model demonstrates good performance based on low-resolution spectra, showing its great potential in developing low-cost sensing systems.