基于光谱特征学习的煤质快速分析方法与关键技术研究

Quality analysis is crucial for the clean and efficient utilization of coal. However, current near-infrared spectroscopy-based analytical methods fail to perform with accuracy and stability. Therefore, this project aims to develop a spectral learning method and key technologies for the efficient analysis of coal quality. In order to address the concerns including sample abnormality, high-level noise and high redundancy in the raw spectra, a novel preprocessing method based on constraint learning is proposed to collect more effective information about mineral characteristics as well as to provide higher-quality data for the subsequent analysis. In order to model the complex non-linear relationship between the spectral features and the coal index accurately, we will develop an analytical model based on the adaptive depth belief network. The analytical model, which is conspired by the laboratory research on model-construction-theory, can be transplanted into portable devices based on transfer learning. In summary, a rapid analytical system of coal-quality will be developed based on the artificial intelligence theory mentioned above. Research results can achieve a more efficient analysis about the whole coal index as well as more-advanced technical supports to the online coal quality evaluation in the process of coal blending.

煤质分析是煤炭清洁高效利用的重要环节。现有的基于近红外光谱技术的煤质分析方法存在检测结果准确性低、稳定性差等问题。本项目拟开展基于光谱特征学习的煤质快速分析方法与关键技术研究。围绕原始数据集中的问题样本、高噪声和高维度，提出了基于约束学习的煤样数据预处理，以获取煤中物质组分特征表达的有效信息，为煤质分析模型提供高质量的建模数据。构建基于自适应深度置信网络的煤质光谱分析模型，以精确描述特征光谱与煤质指标间复杂的非线性关系，用于煤炭各指标的快速定量预测。利用迁移学习策略，研究基于相似匹配的深度置信网络模型传递，将模型封装移植到便携式设备上。基于上述智能理论方法，开发煤质快速分析系统。研究成果可实现煤质全元素的快速分析，为配煤过程中煤炭品质的在线检测提供技术支持。

煤质分析是实现煤炭高效生产利用和畅通国际煤炭贸易的关键环节。传统基于物理化学试验的分析方法耗时耗力且受主观因素影响大，不能满足煤炭生产和消费等环节中快速、实时的要求。鉴于此，面向煤质的定性、定量分析问题，本研究融合近红外光谱分析和机器学习技术，探索了煤炭定性、定量分析模型的构建。主要研究成果包括：(1)针对非目标因素影响原始煤样数据质量的问题，研究建模数据的清洗、光谱信息的恢复与提取以及采集参数的优化，以提高样本表观光谱的可靠性。(2)针对近红外光谱数据集样本量小、存在异质性且输入变量(吸光度)与输出变量(性质标签)关系复杂等问题，研究了基于宽度学习的煤炭定性分析（包括产地鉴别和煤矸识别）模型构建。(3)针对煤质工业参数之间存在的关联性，本研究引入多任务深度学习技术，提出基于注意力U型网络的回归分析方法，实现了水分、灰分、挥发分和发热量的协同预测。研究成果可实现煤质参数的在线分析，为配煤过程中煤炭品质的在线检测提供技术支持，具有广阔的应用前景。