摘要
实际选煤过程中一些关键质量参数往往依靠人工化验,使得感知数据稀疏,即标签数据较少.建立数据驱动技术的选煤产品质量软测量模型需要采用半监督学习(SSL)方法.然而,SSL对提高模型质量比较有限,容易导致建立的数据驱动模型存在较大误差.本文充分分析样本信息度、样本代表性与过程非线性对建模精度的重要性,将随机向量函数链接网络(RVFLN)与基于多样性评价指标的主动学习(AL)策略相结合,提出一种主动半监督随机向量函数链接网络(ASS-RVFLN)的选煤灰分软测量建模方法.所提方法首先通过回归问题进行试验研究,验证了在获得较好模型性能时能够有效减少标注负担;然后应用于煤炭工业重介质选煤过程中的灰分估计,表明其有效性与工业应用的潜力.
Some key quality parameters often rely on manual testing in coal preparation process, which leads to sparse perception data, namely less labeled data. Semi-surpervised learning(SSL) method is needed to establish the soft-sensing model of coal preparation product quality based on data-driven technology. However, SSL is limited in improving model quality, which causes large error in the constructed data-driven model. This paper proposes an active semi-supervised RVFLN(ASS-RVFLN) soft-sensing modeling method for ash estimation in coal preparation process, which combines random vector functional link network(RVFLN) and active learning(AL) strategy based on multiple evaluation indexes, by analyzing the importance of process nonlinearity, information and representativeness on data to modeling accuracy sufficiently. Firstly, experimental studies are conducted through regression issues, which verifies the effectively for reducing the labeling burden when achieving better promising performance, then applied to the coal dense medium preparation process for estimating the ash, which further verifies the effectiveness and potential of industrial application.
作者
胡金成
HU Jincheng(China Coal Technology and Engineering Group Changzhou Research Institute Co.,Ltd.,Changzhou,Jiangsu 213015,China;Tiandi(Changzhou)Automation Co.,Ltd.,Changzhou,Jiangsu 213015,China)
出处
《中国矿业大学学报》
EI
CAS
CSCD
北大核心
2022年第6期1232-1240,共9页
Journal of China University of Mining & Technology
基金
江苏省高等学校自然科学研究面上基金项目(20KJB440001)
中煤科工集团常州研究院科研项目(2018ZX001-06,2021GY2010)
天地科技股份有限公司科技创新创业资金专项项目(2020-TD-ZD010)。
关键词
选煤过程
半监督学习
随机向量函数链接网络
主动学习
软测量建模
coal preparation process
semi-supervised learning
random vector functional link network
active learning
soft-sensing modeling
