Second-harmonic signal processing is critical for tunable diode laser absorption spectroscopy (TDLAS), where accurate signal extraction and noise suppression directly affect gas concentration inversion accuracy. Traditional filtering methods, such as empirical mode decomposition and Savitzky-Golay filtering, often struggle to balance noise suppression and feature preservation, especially under low signal-to-noise ratio (SNR) conditions. To address these challenges, we propose a multi-level feature fusion dual-channel convolutional neural network (PCA-GASF-DCResNet) for second-harmonic signal denoising. The method integrates feature dimensionality reduction, time-series feature encoding, and deep learning-based noise suppression to improve signal representation and denoising robustness.

The proposed method integrates Principal Component Analysis (PCA) and Gramian Angular Summation Field (GASF) transformation to enhance feature representation, where PCA removes redundant information and reduces data dimensionality while GASF encodes one-dimensional second-harmonic signals into two-dimensional matrices to enhance temporal dependencies for deep learning models. To extract both global and local features, a dual-channel convolutional neural network (CNN) is designed, where one channel processes PCA-reduced one-dimensional signals with large convolution kernels to capture global features, while the other channel applies small convolution kernels to the GASF-transformed two-dimensional signals for fine-grained local feature extraction. Residual learning (ResNet) is incorporated to optimize deep network performance by mitigating gradient vanishing issues, accelerating convergence, and maintaining model stability. The proposed PCA-GASF-DCResNet is systematically evaluated under various SNR conditions using simulated noisy second-harmonic signals, and its performance is compared with conventional denoising methods, including EEMD-WT, VMD-WTFD, VMD-SG, and CNN, to validate its effectiveness in low-SNR environments. Ablation studies are conducted to assess the contributions of each channel and feature fusion. Furthermore, to validate the robustness of the proposed method, it is tested on signals with varying gas concentrations. The proposed method is further applied to experimentally measured TDLAS signals to demonstrate its practical utility in improving gas concentration inversion accuracy.

The proposed PCA-GASF-DCResNet model is systematically evaluated in both simulations and experiments. Compared to traditional denoising approaches, PCA-GASF-DCResNet achieves a substantial SNR improvement, significantly enhancing second-harmonic signal recovery (Table 1). The impact of each channel is assessed by introducing PCA-ResNet and GASF-ResNet as single-channel models. The results show that single-channel processing achieves only moderate SNR improvements, while the dual-channel fusion in PCA-GASF-DCResNet leads to superior noise suppression and feature retention, validating the necessity of multi-level feature fusion. The robustness verification indicates that the denoised second-harmonic signal maintains a high level of consistency across different concentration levels. The linear correlations of the peak and peak-trough values with gas volatilization time reach 0.99989 and 0.99987, respectively (Fig. 8). Furthermore, experimental results based on the absorption line of CO₂ at the center wavelength ν₀=1578.2 nm confirm that the proposed method effectively suppresses noise and baseline fluctuations and the full width remains stable over time. Additionally, the peak fitting correlation reaches 0.99138, further demonstrating the accuracy and reliability of the algorithm for gas concentration inversion (Fig. 10).

In this paper, we propose PCA-GASF-DCResNet, a dual-channel convolutional neural network for second-harmonic signal denoising in TDLAS. The method integrates PCA for dimensionality reduction and GASF transformation to enhance feature extraction and improve denoising performance. Using the 1578.22 nm CO₂ absorption line, noisy second-harmonic signals were simulated and validated through simulations and experiments. The results demonstrate that PCA-GASF-DCResNet significantly improves SNR and produces denoised signals closest to the ideal second-harmonic waveform. Ablation studies confirm that dual-channel fusion outperforms single-channel processing, further enhancing SNR. Additionally, experimental results show that the denoised signal’s peak values exhibit a strong linear correlation with gas volatilization time, while baseline fluctuations are significantly reduced, maintaining stability even under low SNR conditions. These findings highlight the superior denoising performance and practical applicability of the proposed method.