The baseline drift of the Fiber Bragg Grating (FBG) spectral signal is usually one of the main problems, caused by the complex external environment. A spectral baseline correction method based on the improved Long Short Term Memory (LSTM) model is proposed in this paper.

Compared with LSTM model, the improved LSTM model extracts feature information of FBG spectral signal by the Convolutional Neural Network (CNN). The improved LSTM model is composed of CNN, full connection, and LSTM network. In this paper, the improved LSTM model is trained by artificial datasets and measured datasets. The artificial datasets are made up of feature noise, baseline, and FBG spectroscopy. Five methods including wavelet soft threshold method, penalty least square method, Recurrent Neural Network (RNN), LSTM, and the improved LSTM model are used as baseline correction. Identification signal probability and Root Mean Square Error (RMSE) are used to evaluate correction results by the five methods.

The artificial datasets of FBG signal are corrected by the improved LSTM model, and the identification signal probability is increased by 60.8%. The improved LSTM model with training by artificial datasets and measured datasets shows better correction results, compared with training by measured datasets. The mean of the RMSE for FBG spectrum decreases by 10.95%. The standard deviation of RMSE decreases by 4%. The measured datasets of FBG signal are corrected by the improved LSTM model, and the identification signal probability is increased by 50.5%. Compared with wavelet soft threshold method, penalty least square method, RNN and LSTM, the improved LSTM model shows best correction results. The mean values of RMSE and the standard deviation of RMSE are 0.012 2 and 0.002 4, respectively. The RMSE value of the demodulated central wavelength is 0.036 pm. And the baseline correction process takes only 9.68 ms.

The improved LSTM model is an effective method to achieve baseline correction, and has wide range of application prospects in complex external environment.