Attention-Enhanced CNN–Transformer for Direct and Interpretable Nitrous Oxide Retrieval from Noisy TDLAS Spectra

Noise-induced spectral fluctuations can substantially degrade gas concentration retrieval in tunable diode laser absorption spectroscopy (TDLAS), particularly when the absorption features are weak. A coordinate-attention-enhanced convolutional neural network and Transformer (CA-CNN-Transformer) is developed for direct retrieval of N₂O concentration from noisy mid-infrared absorption spectra. The model combines convolutional feature extraction, channel-wise feature recalibration using the CA module, and Transformer-based modeling of spectral dependencies, allowing for direct extraction of concentration-related features from the measured spectra without an explicit denoising step. After transfer learning using experimentally acquired spectra, the model achieved an R² of 0.99998 across the concentration range of 100–1000 ppb. In one-hour continuous measurements at approximately 400 ppb, the standard deviation of the retrieved concentration decreased from 2.1 ppb for a conventional CNN to 1.0 ppb for the proposed model. The minimum Allan deviation reached 150 ppt at an averaging time of 180 s. SHapley Additive exPlanations (SHAP) analysis identified the spectral region associated with the dominant N₂O absorption feature and showed that the model uses information from multiple spectral regions rather than depending on individual spectral points. The proposed CA-CNN-Transformer thus provides accurate and stable N₂O concentration retrieval from noisy TDLAS spectra while offering an interpretable basis for understanding the spectral information used by the model.