Hybrid CNN-LSTM model for concentration retrieval in active fiber loop ring-down spectroscopy under EDFA gain saturation

In active fiber loop ring-down spectroscopy (FLRDS) systems, gain saturation in the erbium-doped fiber amplifier (EDFA) causes inter-pulse gain fluctuations, resulting in significant deviations from ideal exponential decay. These deviations compromise measurement accuracy because traditional methods depend on precise extraction of the ring-down time (τ). To address this issue, we propose a hybrid deep learning framework that combines a convolutional neural network (CNN) with a long short-term memory (LSTM) network for FLRDS-based gas sensing. This architecture captures both local temporal features and long-range dependencies, enabling nonlinear compensation that maps distorted ring-down signals directly to gas concentrations—bypassing the error-prone step of τ estimation. Experimental results show that the CNN-LSTM model substantially improves concentration retrieval accuracy, outperforming not only standalone CNN and LSTM models, but also other common machine and deep learning methods in key metrics such as RMSE, MAE, MAPE, and R². Over a concentration range of 40–1600 ppm, the model effectively mitigates EDFA gain saturation effects, achieving relative errors below 2.5 % and offering improved robustness for practical gas sensing. An uncertainty budget analysis further confirms that the proposed method yields lower combined standard uncertainty and narrower confidence intervals than exponential fitting, underscoring its reliability for high-precision gas detection.