Combination of off-axis integrated cavity absorption spectroscopy with a neural network model for precise multi-component gas detection

A multi-gas detection sensor that combined off-axis integrated cavity output spectroscopy (OA-ICOS) with an attention-enhanced dual-path convolutional neural network (DP-CNN) was developed and evaluated. Three DFB lasers operating at wavelengths of 1653.74 nm, 1578 nm, and 1567 nm were used to measure the concentration of CH₄, H₂S and CO simultaneously by a time-division multiplexing method. To reduce residual cavity mode noise in OA-ICOS, Radio frequency (RF) white noise was injected into the laser current. A comparison showed that the DP-CNN outperformed a conventional CNN in detection performance. The linear correlation coefficients (R²) between the predicted concentrations and the standard concentrations of CH₄, H₂S, and CO had increased from 99.30%, 98.10%, and 99.06% (conventional CNN) to 99.70%, 99.49%, and 99.79% (DP-CNN), respectively. To further evaluate the long-term performance of the DP-CNN, a 67-minute continuous measurement was conducted on gas mixtures containing 6.26 ppm CH₄, 104.35 ppm H₂S, and 347.83 ppm CO. When the detection time was 208 s, the limits of detection for the three gases reached 38.59 ppb, 0.58 ppm, and 2.02 ppm, corresponding to improvements of 3.24, 2.84, and 4.15 times, respectively, when compared with those of the gas detection method based on absorption peak fitting. It could be concluded that the DP-CNN model has great potential in the real-time processing of multi-component gas absorption spectroscopy data.