【作者】 李瑞；

【导师】 李博； 席庆祥；

【作者基本信息】 太原理工大学 ， 机械工程（专业学位）， 2023， 硕士

【摘要】 煤矿智能化是未来煤炭工业高质量发展的方向,煤矸智能化识别技术是实现煤矿智能化亟待解决的关键技术之一。可见-近红外光谱(Visible and near-infrared spectroscopy,VIS-NIR)检测技术具有无损、快速、稳定、无污染、样品无需预处理等特点,符合煤矸识别的要求,但目前该技术在煤矸识别领域的研究以对静止状态样品进行离线识别为主,与实际生产要求不符。因此本文按照样品状态由静止、拟动态到动态,识别方式由离线识别转向在线识别进行递进研究,最终初步实现对运动状态下的煤与矸石的在线识别。主要研究内容和结论如下:(1)在煤矸静态研究中,首先设计并搭建了煤矸静态可见-近红外光谱采集平台,采集了不同煤矿的煤与矸石样品光谱,并对数据进行了预处理、建模分析、煤矿适用性检验等研究。使用K近邻(K-Nearest Neighbor,KNN)、随机森林(Random Forest,RF)、支持向量机(Support Vector Machine,SVM)、极端梯度提升树(e Xtreme Gradient Boosting,XGBoost)机器学习分类算法建立的全波段光谱分类模型以及结合连续投影算法(Successive Projections Algorithm,SPA)、递归特征选择(Recursive feature element,RFE)、采用竞争性自适应重加权法(Competitive Adaptive Reweighted Sampling,CARS)特征选择算法建立的特征波长分类模型均能够较好的分类煤与矸石,其中XGBoost模型准确率达到97.05%,基于RFE选出的特征建立的XGBoost分类模型准确率达到98.03%。在对近红外光谱煤矸识别技术的煤矿适用性研究中,上述模型同样能够实现准确分类,在神木煤矿中准确率为100.00%、100.00%,巴隆图煤矿为96.42%、100.00%,表明可见-近红外光谱识别技术具有良好的煤矿适用性,是一种实现煤矸识别的通用技术。在在线识别研究中,以Lab VIEW为平台,引入MATLAB Script与Python Node模块建立的在线识别系统能够在1.226 s内实现对样品的准确识别。(2)在煤矸拟动态研究中,为解决上述静态研究中的问题,模拟煤矸真实运动条件采集样品在探头下方呈直线排列的多个位置的光谱并建立了基于多种特征融合方法——平均法(Mean)、拼接法(Concat)、伪彩色图像(Pseudo-RGB,p-RGB)及特征层融合(Feature Layer,F-L)与分类算法——机器学习(Machine Learning,ML),二维卷积神经网络(Two-dimensional Convolutional Neural Network,2DCNN),一维卷积神经网络(One-dimensional Convolutional Neural Networks 1DCNN)的多位置光谱分类模型。与单位置分类模型相比,多位置光谱分类模型可以显著提高煤和矸石的识别精度。基于F-L构建的1DCNN在众多模型中分类性能最好,准确率达到97.61%,较分类能力最强的单位置分类模型提升了8.84%,预测速度也达到1.51 ms。另外为探寻多位置光谱分类模型表现更佳的原因,使用注意力可视化算法(Gradient-weighted Class Activation Mapping,Grad-CAM)对特征重要性进行可视化,发现在不同位置的光谱中对预测起积极贡献的特征不同,而特征融合方法使得来自样品多位置的特征参与决策,特别是基于F-L融合方法的模型提取了每个位置光谱中的关键信息。按照拟动态离线研究中的数据处理流程,构建了基于特征融合方法的煤矸在线识别系统,经测试可以实现快速、准确的识别,平均准确率与识别时间分别为93.15%与1.241 s。(3)在动态研究中,引入运动模块、位置感知模块等构建了煤矸动态VIS-NIR采集系统用于运动状态下样品的多位置光谱采集。运动状态下煤与矸石的光谱,与拟动态研究中的特征表现一致,不同位置的光谱特征差异明显,同一位置下的煤与矸石的光谱特征也存在着明显差异,再次验证了多位置光谱采集方法在获取样品光谱信息上的先进性。基于F-L特征融合方法建立的1DCNN模型分类准确率达到95.00%,较单位置模型提升了3.64%-11.76%。基于真实图像建立的2DCNN模型中,Mobile Net模型分类表现最佳,准确率达到95.00%。鉴于不同信息下模型对煤与矸石的分类表现不同,使用多模态融合方法,将基于光谱信息建立的1DCNN模型与基于真实图像信息建立的2DCNN模型进行决策融合,可进一步提升模型分类准确率。最后在在线识别研究中,煤矸动态在线识别系统初步实现了运动状态下煤与矸石的在线识别,为近红外光谱煤矸识别技术产业化提供了参考。更多还原

【Abstract】 The intelligent coal mine represents the future direction of high-quality development of the coal industry,and intelligent identification of coal and gangue is one of the key technologies that need to be solved to achieve intelligent coal mines.Visible and near-infrared spectroscopy(VIS-NIR)detection technology has the characteristics of non-destructive,fast,stable,non-polluting,no pretreatment for samples and so on,which is in line with the requirements of coal gangue identification.However,in most current coal and gangue identification studies based on this technology,the samples are stationary,which deviates from actual production conditions.Therefore,this paper adopts a progressive approach by examining the sample state in static,quasi-dynamic and dynamic modes,as well as exploring offline and online recognition methods.Finally,the online identification of coal and gangue in the moving state is preliminarily realized.The main research contents are as follows:(1)In the study of static coal and gangue,a VIS-NIR spectral acquisition platform for coal and gangue was designed and built,and the spectra were pre-processed,modeled,and analyzed,and the applicability of coal mines was checked.The full band classification models were constructed using K-nearest neighbor(KNN),random forest(RF),support vector machine(SVM),and extreme gradient boosting tree(XGBoost)machine learning algorithms.Characteristic wavelength classification models were built based on the feature wavelengths selected by successive projection algorithm(SPA),recursive feature selection(RFE),and competitive adaptive reweighted sampling(CARS)feature selection algorithm.Among them,the accuracy of XGBoost model reached 97.05 % and the accuracy of XGBoost classification model based on the features selected by RFE reached 98.03 %.In the study on the applicability of NIR spectral coal gangue recognition technology in coal mines,the above models can achieve accurate classification,with accuracy rates of 100.00 % and 100.00 % in the Shenmu coal mine and 96.42 % and 100.00 % in the Balongtu coal mine.The results indicated that the VIS-NIR spectral recognition technology exhibited excellent applicability in coal mines and represented a universal approach to achieving coal gangue recognition.In the online identification study,Lab VIEW served as the platform and the established online identification system,incorporating MATLAB Script and Python Node module,achieved precise sample identification within 1.226 s.However,the system had more misjudgments for new batches of coal and gangue samples with more inclined surfaces.(2)In the quasi-dynamic study of coal gangue,to address the aforementioned static research issues,simulating the actual movement conditions of coal gangue and collected spectra of samples at multiple locations along a straight line below the probe.A variety of feature fusion models were built using Mean,Concat,pseudo-color image(p-RGB),and feature layer fusion(F-L)feature fusion methods,along with machine learning(ML),two-dimensional convolutional neural networks(2DCNN),and one-dimensional convolutional neural networks(1DCNN)classification algorithms.The multi-location spectral classification model significantly improved the recognition accuracy of coal and gangue compared to the single location classification model.The 1DCNN model based on the F-L feature fusion method had the best classification performance among the models,with an accuracy of 97.61 %,an improvement of 8.84 % over the most powerful unitary classification model,and a prediction speed of 1.51 ms.Additionally,the Gradient-weighted Class Activation Mapping algorithm(Grad-CAM)was used to visualize the importance of features to identify the reasons for the better performance of the multi-location spectral classification model.It was found that the bands contributing positively to the prediction varied in different location spectra.In particular,the model based on the F-L fusion method accurately extracted the key information from each location’s spectra.Based on the proposed dynamic offline study,an online identification system was constructed using the feature fusion method.It was tested and achieved fast and accurate identification,with an average accuracy of 93.15 % and an identification time of 1.241 s.(3)In the dynamic study,the motion module and location sensing module were introduced to build a dynamic VIS-NIR acquisition system for the multi-location spectral acquisition of samples in motion.The spectra of coal and gangue in the motion state were consistent with the performance of the features in the proposed dynamic study.There were obvious differences in the spectral features of different positions,as well as significant differences in the spectral features of coal and gangue at the same location.This further validated the effectiveness of the multi-location spectral acquisition method in capturing spectral information from samples.The classification accuracy of the 1DCNN model based on the F-L feature fusion method reached 95.00 %,showing an improvement of 3.64 % to 11.76 %compared to the unit location model.Among the 2DCNN models built based on real images,the Mobile Net model performed the best,achieving an accuracy of 95.00 %.Considering the different performances of the models in classifying coal and gangue with different information,a multimodal fusion approach can be utilized to combine the 1DCNN model based on spectral information with the 2DCNN model based on real image information,thereby further improving the classification accuracy of the models.Lastly,in the online recognition study,the dynamic coal and gangue online recognition system successfully achieved the online recognition of coal and gangue in the motion state.This provides a reference for the industrialization of near-infrared spectral coal gangue recognition technology.更多还原