基于ResNet和Transformer深度联合网络的TDLAS气体检测信号降噪算法研究

Research on Denoising Algorithm for TDLAS Gas Detection Signals Based on ResNet and Transformer Deep Joint Network

摘要: TDLAS技术的核心在于利用吸收谱线的峰形、峰位和线宽等特征来反演气体浓度等关键信息。然而，实际探测中不可避免地会受到光源不稳定、热噪声及干涉条纹等复杂噪声的干扰。这些噪声会淹没信号细节、扭曲峰形，严重影响测量精度，尤其是在低浓度检测场景下。传统的去噪方法往往难以在有效抑制噪声的同时保持谱线关键形状特征的完整性。为解决这一难题，本文提出了一种结合ResNet-1D与Transformer优势的混合神经网络框架(RTDNet)。该模型利用ResNet-1D强大的局部特征提取能力作为编码器，捕捉谱线的多尺度局部结构。同时，在瓶颈层引入集成了卷积前馈网络的Transformer模块，借助其自注意力机制来建模全局长程依赖关系，从而更好地识别和剔除复杂背景噪声。随后通过解码器和跳跃连接恢复信号细节。此外，针对处理长光谱数据的工程需求，模型还引入了窗口化滑动推理与重叠相加策略以满足计算约束。实验验证表明，RTDNet在不同信噪比条件下均能显著提升信号质量，有效降低误差，并在强噪声背景下稳定保持谱线的峰形与峰位特征。

Abstract: The core of TDLAS technology lies in utilizing characteristics such as the peak shape, peak position, and linewidth of absorption spectra to invert critical information like gas concentration. However, practical detection inevitably suffers from interference caused by complex noise sources such as light source fluctuation, detector thermal noise, and interference fringes. These noises obscure signal details, distort peak shapes, and severely degrade measurement accuracy, particularly in low-concentration detection scenarios. Traditional denoising methods often struggle to balance effective noise suppression with preserving the integrity of critical spectral shape features. To address this challenge, this paper proposes RTDNet, a hybrid neural network framework that combines the strengths of ResNet-1D and Transformer. The model leverages the powerful local feature extraction capability of ResNet-1D as an encoder to capture multi-scale local structures of the spectra. Meanwhile, a Transformer module integrated with a convolutional feed-forward network is introduced at the bottleneck layer, using its self-attention mechanism to model global long-range dependencies to better identify and eliminate complex background noise. Subsequently, signal details are recovered through a decoder and skip connections. Furthermore, to address engineering requirements for processing long spectral data, the model incorporates windowed sliding inference and overlap-add strategies to meet computational constraints. Experimental validation demonstrates that RTDNet significantly improves signal quality and effectively reduces errors across varying signal-to-noise ratios (SNRs), while stably preserving peak shape and position features under strong noise conditions.