【摘要】 为满足有色金属冶炼过程中二氧化硫排放的精准监测需求（排放限值12.25ppm）,本研究针对实际光谱信号中存在的噪声干扰与标定样本稀缺问题,提出了一种融合预处理优选、数据增强与集成学习的分析框架。实验设定SO₂浓度范围10～25ppm,覆盖排放限值以验证方法适用性。该框架首先对比了多种预处理方法,筛选出最优去噪策略以提升信号质量；其次,构建带梯度惩罚的Wasserstein生成对抗网络（WGAN-GP）模型,在有限样本下实现高质量光谱数据生成,并同步从判别器提取深度光谱特征；最后,基于XGBoost、SVM与KNN构建Blending集成回归模型,融合多算法优势以提升预测鲁棒性。实验结果显示,集成模型决定系数达0.954,较最优单一模型提升超过9%,验证了所提框架在小样本条件下对SO₂浓度监测的有效性与先进性。更多还原

【Abstract】 To meet the requirement for accurate monitoring of sulfur dioxide emissions（with an emission limit of 12.25 ppm） in non-ferrous metal smelting processes, this study addresses the challenges of noise interference in actual spectral signals and the scarcity of calibration samples by proposing an integrated analytical framework combining optimal preprocessing, data augmentation, and ensemble learning. The SO₂ concentration range was set at 10–25 ppm in the experiment, covering the emission limit to validate the applicability of the method. The framework first compares multiple preprocessing methods to select the optimal denoising strategy for improving signal quality. It then constructs a Wasserstein Generative Adversarial Network with Gradient Penalty（WGAN-GP） to generate high-quality spectral data under limited sample conditions while simultaneously extracting deep spectral features from the discriminator. Finally, a Blending ensemble regression model based on XGBoost, SVM, and KNN is built to integrate the advantages of multiple algorithms and enhance predictive robustness. Experimental results show that the ensemble model achieves a coefficient of determination（R²） of 0.954, representing an improvement of over 9 percentage points compared to the best single model, confirming the effectiveness and advancement of the proposed framework for SO₂ concentration monitoring under small-sample conditions.更多还原