基于遗传算法优化的XGBoost煤质LIBS定量分析算法研究

doi:10.3964/j.issn.1000-0593(2025)11-3048-09

摘要
参考文献
相关文章 (15)

全文: PDF (7389 KB)
输出: BibTeX | EndNote (RIS)

摘要：针对燃煤铜冶炼厂对煤质成分在线检测的高精度需求，提出了一种基于遗传算法（GA）优化的极端梯度提升决策树（XGBoost）集成模型GA-XGBoost，提升了激光诱导击穿光谱（LIBS）技术在复杂煤质分析中的工业适用性。针对传统XGBoost模型因超参数敏感性导致的过拟合及泛化能力不足问题，GA-XGBoost采用遗传算法全局搜索策略，自适应优化关键超参数（学习率、树深度、正则化参数等），并结合二进制编码实现特征子集的动态筛选，从而有效降低高维LIBS光谱数据中的噪声干扰。为验证模型的实用性与普适性，选取山西大同（烟煤）及内蒙古鄂尔多斯（褐煤）两大典型产区的59个标准煤样作为实验对象。通过自适应迭代基线校正（airPLS）与Savitzky-Golay滤波等预处理方法，将原始光谱特征维度从14 328维压缩至500维，并利用SHAP值筛选关键特征。在此基础上，分别采用GA-XGBoost、XGBoost、随机森林（RF）、支持向量机（SVM）、多元线性回归（MLR）与偏最小二乘（PLS）进行灰分与热值的预测对比实验。实验结果表明，相较于传统XGBoost、随机森林等模型，GA-XGBoost在煤质灰分和热值预测任务中表现更优，灰分预测的决定系数（R²）提升0.053，均方根误差（RMSE）降低0.964，平均绝对误差（MAE）降低0.324，RSD降低了1.494%；热值预测的R²提升0.003，RMSE降低0.021，MAE降低0.07，RSD降低了0.871%。此外，为验证模型在复杂工业环境下的鲁棒性，该研究选取了20组未参与建模的煤炭现场数据进行外部验证，结果表明GA-XGBoost在灰分和热值预测中的误差均保持在合理范围内，验证了模型对不同煤种的适应能力。该研究将GA优化与LIBS煤质光谱分析相结合，有效应对光谱干扰和模型泛化问题，为实时、多参数煤质分析提供了统一框架。现场测试结果表明，该模型可集成至燃煤铜冶炼厂的LIBS在线检测系统，灰分预测误差<1%、热值误差<0.5 MJ·kg^-1。该研究将遗传算法与LIBS光谱分析深度融合，提出了一个兼顾精度和鲁棒性的煤质多参数快速定量分析框架，为煤炭清洁与高效利用提供了新的技术途径。

关键词：LIBS；煤质；GA-XGBoost；超参数优化；工业鲁棒性

Abstract：To address the high-precision demand for online coal quality detection in coal-fired copper smelters, this study proposes a genetic algorithm (GA)-optimized extreme gradient boosting decision tree (XGBoost) ensemble model (GA-XGBoost), enhancing the industrial applicability of laser-induced breakdown spectroscopy (LIBS) for complex coal analysis. To overcome overfitting and limited generalization in traditional XGBoost caused by hyperparameter sensitivity, GA-XGBoost implements a GA-based global search strategy to adaptively optimize key hyperparameters (e. g., learning rate, tree depth, regularization parameters) coupled with binary-encoded chromosome representation for dynamic spectral feature selection, effectively suppressing noise in 14 328-dimensional LIBS data. Validation employed 59 standard coal samples from Datong, Shanxi (bituminous coal) and Ordos, Inner Mongolia (lignite). Preprocessing via adaptive iteratively reweighted penalized least squares (airPLS) and Savitzky-Golay filtering reduced spectral dimensions to 500, with dominant features identified by SHapley Additive exPlanations (SHAP) values. Comparative experiments demonstrated GA-XGBoost's superiority over XGBoost, random forest (RF), support vector machine (SVM), multiple linear regression (MLR), and partial least squares (PLS) in predicting ash content and calorific value. For ash content, GA-XGBoost achieved a 0.053 increase in R², 0.964% reduction in RMSE, 0.324% decrease in MAE, and 1.494-percentage-point lower RSD. For calorific value, it yielded a 0.003 R² improvement, 0.021 MJ·kg^-1 RMSE reduction, 0.07 MJ·kg^-1 MAE decrease, and 0.871-percentage-point RSD reduction. External validation using 20 unmodeled on-site coal datasets confirmed robustness in industrial environments, with errors constrained within 1% (ash) and 0.5 MJ·kg^-1 (calorific value). The GA-LIBS integration addresses spectral interference and generalization challenges, establishing a unified framework for real-time multi-parameter coal analysis. Field deployment in an operational copper smelter demonstrated seamless integration into LIBS online systems, providing a technically viable pathway toward clean and efficient coal utilization.

Key words：LIBS; Coal quality; GA-XGBoost; Hyperparameter optimization; Industrial robustness

收稿日期: 2025-04-17 修订日期: 2025-07-25

中图分类号:

O433.4

基金资助: 国家重点研发计划项目(2024YFC3909605)资助

通讯作者: 潘从元 E-mail: pancy@goldstar-china.com

作者简介: 朱婷婷，女，1999年生，安徽理工大学数学与大数据学院硕士研究生 e-mail: 15655446560@163.com

引用本文:

朱婷婷，潘从元，赵前进，薛骅骎，申远，张兵. 基于遗传算法优化的XGBoost煤质LIBS定量分析算法研究[J]. 光谱学与光谱分析, 2025, 45(11): 3048-3056.
ZHU Ting-ting, PAN Cong-yuan, ZHAO Qian-jin, XUE Hua-qin, SHEN Yuan, ZHANG Bing. Research on a Genetic Algorithm-Optimized XGBoost Algorithm for Quantitative Analysis of Coal Quality Using Laser-Induced Breakdown Spectroscopy. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(11): 3048-3056.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2025)11-3048-09 或 https://www.gpxygpfx.com/CN/Y2025/V45/I11/3048

[1] Zuo Z, Feng Y, Dong X et al. Fuel Processing Technology, 2022, 235: 107361.
[2] YANG Guang-cai, KONG Xiang-tong(杨光彩, 孔祥同). China Nonferrous Metallurgy(中国有色冶金), 2020, 49(6): 60.
[3] Zhang Q, Yi H A, Yu Z H, et al. Applied Thermal Engineering, 2018, 144: 614.
[4] Chen T C, Iliyasu A M, Alizadeh S M, et al. Applied Artificial Intelligence, 2023, 37(1): 2166225.
[5] Li J T, Jia W B, Hei D Q, et al. Nuclear Engineering and Technology, 2022, 54(6): 2221.
[6] Nakano H, Sonoda S, Matsuyama T, et al. X-Ray Spectrometry, 2021, 50(3): 224.
[7] XU Shui-xiu, YU Zi-yu, QIN Huai-qing, et al(徐水秀, 喻子彧, 覃淮青, 等). Chinese Journal of Quantum Electronics(量子电子学报), 2021, 38(6): 727.
[8] Wang Z, Afgan M S, Gu W, et al. Trends in Analytical Chemistry, 2021, 143: 116385.
[9] Gornushkin I B, Smith B W, Potts G E, et al. Analytical Chemistry, 1999, 71(23): 5447.
[10] Lu P, Zhuo Z, Zhang W, et al. Applied Optics, 2020, 59: 6443.
[11] Li W, Lu J, Dong M, et al. Energy & Fuels, 2018, 32(1): 24.
[12] Zhang W, ZhuoZ, Lu P, et al. Journal of Analytical Methods in Chemistry, 2020, 35(8): 1621.
[13] Zhang Y, Xiong Z, Ma Y, et al. Analytical Methods, 2020, 12(27): 3530.
[14] WANG An, CUI Jia-cheng, SONG Wei-ran, et al(王安, 崔佳诚, 宋惟然, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2024, 44(7): 1940.
[15] WANG Ming-xuan, WANG Qiao-yun, PIAN Fei-fei, et al(王铭萱, 王巧云, 骈斐斐, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析) , 2022, 42(6): 1721.
[16] Razavi-Termeh S V, Sadeghi-Niaraki A, Abba S I, et al. Applied Water Science, 2024, 14(11): 244.
[17] LI Ke, HUANG Dong-chen, TAO Zi-bin, et al(李科, 黄东晨, 陶子彬, 等). Electric Power Engineering Technology(电力工程技术) , 2021, 40(6): 95.
[18] Zhao J, Li Y Y, Jin H, et al. Innovation in Aging, 2024, 8: 429.
[19] Goldberg D E. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley, 1989: 135.
[20] LIU Shu-lin, WANG Meng, LI An, et al(刘树林, 王猛, 李安, 等). Optics and Precision Engineering(光学精密工程), 2024, 32(10): 1470.