Volatile matter is a key indicator for evaluating coal quality and combustion characteristics, as its content directly affects combustion efficiency, reaction activity, and process adaptability. Near-infrared spectroscopy is a rapid and non-destructive analytical technique that has been widely applied to the quantitative prediction of coal volatile matter. However, in cross-batch applications, systematic spectral drift often occurs. When new data is used to rebuild the model, the predictive capability for previous samples diminishes, exacerbating the issue of catastrophic forgetting during model updates. To address this challenge, this study introduces a continual-learning modeling method for coal volatile-matter prediction. During the feature extraction stage, a network combining dense connections and self-attention mechanisms is developed to effectively capture both local spectral details and global dependencies. In the model updating stage, a continual learning strategy integrating statistical feature replay and knowledge distillation is implemented to achieve knowledge retention and task adaptability without accessing original historical samples. Experiments conducted with three batches of coal samples demonstrate that the proposed method effectively mitigates catastrophic forgetting and maintains stable prediction performance under distributional shifts. Its overall accuracy approaches that of independent single-task models, thereby providing a feasible technical solution for long-term and stable volatile matter detection in complex industrial environments.
密集连接块由多个密集连接层组成,每一层均包含卷积、整流线性单元(rectified linear unit,ReLU)激活函数和批标准化处理。与传统串行结构不同,密集连接块通过逐层拼接不同卷积层的输出特征,使浅层与深层的特征能够在整个网络中实现累积与复用。输入特征经逐层卷积后,在通道维度上依次拼接,最终形成的特征输出,从而在保持序列长度不变的同时,显著增强了特征表达的丰富性和训练效率。密集连接通过显式特征拼接,在网络层间保留了更多细粒度光谱信息,从而降低了信息在传递过程中的弱化风险。在谱峰混叠与弱信问题普遍存在的煤炭近红外光谱数据中,这种累积式特征融合机制尤其有利于在建模过程中整合不同层次的光谱信息[15]。
为使合成特征在可分性和标签一致性上逼近历史分布,SRD在回放模块中引入投影条件判别器,并结合WGAN-GP(Wasserstein GAN with gradient penalty)进行优化。WGAN-GP在Wasserstein距离的基础上,通过在判别器中引入梯度惩罚函数,有效缓解了生成对抗训练的不稳定性,并提升了特征分布的逼近质量;生成器的优化目标则为最大化判别器对伪样本的评分。生成器的WGAN-GP损失()计算式为
LINBoqiang, SHIFengyuan. Coal price, economic growth and electricity consumption in China under the background of energy transition[J]. Energy Policy, 2024, 195: 114400.DOI:10.1016/j.enpol.2024.114400 .
ZHANGYuanbo, ZHANGYutao, LIYaqing, et al. Determination of ignition temperature and kinetics and thermodynamics analysis of high-volatile coal based on differential derivative thermogravimetry[J]. Energy, 2022, 240: 122493.DOI:10.1016/j.energy.2021. 122493 .
SONGWeiran, HOUZongyu, GUWeilun, et al. Industrial at-line analysis of coal properties using laser-induced breakdown spectroscopy combined with machine learning[J]. Fuel, 2021, 306: 121667.DOI:10.1016/j.fuel.2021.121667 .
[7]
LIUKe, HEChao, ZHUChenwei, et al. A review of laser-induced breakdown spectroscopy for coal analysis[J]. TrAC Trends in Analytical Chemistry, 2021, 143: 116357.DOI:10.1016/j.trac.2021.116357 .
[8]
TIANZhihui, LIJiaxuan, WANGShuqing, et al. Development and industrial application of LIBS-XRF coal quality analyzer by combining PCA and PLS regression methods[J]. Journal of Analytical Atomic Spectrometry, 2023, 38(7): 1421-1430.
[9]
GELichao, LIUXiaoyan, FENGHongcui, et al. The interaction between microwave and coal: A discussion on the state-of-the-art[J]. Fuel, 2022, 314: 123140.DOI:10.1016/j.fuel.2022.123140 .
[10]
TIANJun, LIMing, TANZhiyi, et al. Intelligent non-destructive measurement of coal moisture via microwave spectroscopy and chemometrics[J]. Chemometrics and Intelligent Laboratory Systems, 2024, 252: 105175.DOI:10.1016/j.chemolab.2024.105175 .
[11]
TSUCHIKAWAS, MATe, INAGAKIT. Application of near-infrared spectroscopy to agriculture and forestry[J]. Analytical Sciences, 2022, 38(4): 635-642.
[12]
QIAOFengkang, LIUTingting, LIZhipeng, et al. Exploration of an apple grading model based on near-infrared spectroscopy[J]. Academic Journal of engineering and technology science, 2024, 7(3): 84-88.
[13]
JUNAEDIE C, LESTARIK, MUCHTARIDIM. Infrared spectroscopy technique for quantification of compounds in plant-based medicine and supplement[J]. Journal of Advanced Pharmaceutical Technology and Research, 2021, 12(1): 1-7.
[14]
TAOWei, FAROKHZADO C. Theranostic nanomedicine in the NIR-Ⅱ window: classification, fabrication, and biomedical applications[J]. Chemical reviews, 2022, 122(6): 5405-5407.
[15]
XUShifan, XUZhibin, ZHENGJiannan, et al. Where does the crude oil originate: The role of near-infrared spectroscopy in accurate source detection[J]. Memetic Computing, 2024, 16(3): 429-443.
[16]
SANTOSF D, VIANNAS G T, CUNHAP H P, et al. Characterization of crude oils with a portable NIR spectrometer[J]. Microchemical Journal, 2022, 181: 107696.DOI:10.1016/j.microc.2022.107696 .
JIANGDaiyu, HUGang, QIGuanqiu, et al. A fully convolutional neural network-based regression approach for effective chemical composition analysis using near-infrared spectroscopy in cloud [J]. Journal of Artificial Intelligence and Technology, 2021, 1(1): 74-82.
[19]
ZOULiang, QIAOJiahui, YUXinhui, et al. Intelligent proximate analysis of coal based on near-infrared spectroscopy and multioutput deep learning[J]. IEEE Transactions on Artificial Intelligence, 2023, 5(3): 1398-1410.
[20]
WANGLiyuan, ZHANGXingxing, SUHang, et al. A comprehensive survey of continual learning: Theory, method and application[J]. IEEE transactions on pattern analysis and machine intelligence, 2024, 46(8): 5362-5383.
BUZZEGAP, BOSCHINIM, PORRELLOA, et al. Dark experience for general continual learning: A strong, simple baseline[J]. Advances in neural information processing systems, 2020, 33: 15920-15930.
[23]
MAGISTRIS, TRINCIT, SOUTIF-CORMERAISA, et al. Elastic feature consolidation for cold start exemplar-free incremental learning[J]. arXiv preprint arXiv:2024.
KIRKPATRICKJ, PASCANUR, RABINOWITZN, et al. Overcoming catastrophic forgetting in neural networks[J]. Proceedings of the national academy of sciences, 2017, 114(13): 3521-3526.
[27]
ALJUNDIR, BABILONIF, ELHOSEINYM, et al. Memory aware synapses: Learning what (not) to forget[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 139-154.