Robust open-path methane sensing under light-intensity attenuation using a harmonic-coupled dual-branch neural network
Abstract
Reliable open-path methane (CH 4 ) sensing requires accurate retrieval under variable optical transmission and environmental interference, which amplify errors in conventional 2 f /1 f normalization. We developed a near-infrared open-path CH 4 sensor based on tunable diode laser absorption spectroscopy with wavelength modulation spectroscopy (TDLAS-WMS) and proposed a harmonic-coupled time-frequency dual-domain neural network (HC-TFNet) for robust concentration retrieval. HC-TFNet jointly processes contaminated 2 f and synchronous 1 f signals. A patch-embedded transformer (PET) branch extracts local and global features from the 2 f waveform, while a bidirectional gated recurrent unit models temporal variations in the 1 f background. Cross-attention fuses both feature streams to reconstruct the clean 2 f waveform and predict CH 4 concentration. Online random contamination combining controlled attenuation, measured optical interference, and electrical dark noise improves robustness. Across the test samples, HC-TFNet achieved a root mean square error (RMSE) of 0.66 ppm and a mean absolute error (MAE) of 0.42 ppm, while the concentration-averaged response over 2–10 ppm exhibited an R 2 of 0.9997. Across attenuation factors of 0.1–0.9, its RMSE remained between 0.64 and 0.97 ppm, outperforming conventional filtering-based 2 f /1 f methods, particularly at low transmittance. For 6 h of valid measurements at 3 ppm, HC-TFNet reduced the minimum Allan deviation from 13.4 to 7.33 ppb compared with the 2 f /1 f method. In the 19.4-h outdoor experiment, HC-TFNet achieved an RMSE of 0.15 ppm and an MAE of 0.11 ppm relative to the LI-7700, approximately 72.6% and 70.3% lower than those of the 2 f /1 f method. These results demonstrate improved CH 4 retrieval under weak-light, complex-noise, and outdoor conditions.
// Source
Authors: Ye Zhang, Ligang Shao, Shiling Feng, Kun Liu, Guishi Wang, Jiajin Chen, Jiaoxu Mei, Xiaoming Gao
Institutions: Chinese Academy of Sciences, University of Science and Technology of China, Hefei Institutes of Physical Science, Anhui Institute of Optics and Fine Mechanics