常规的煤炭鉴别方法需进行繁琐的制样过程, 且需结合多种化学参数指标进行综合判定, 以得到较为准确的分析结果。 提出一种基于500～2 350 nm的可见-近红外全谱段光谱分析技术与多层感知器(multilayer perceptron, MLP)分类方法相结合的块状商品煤鉴别方法。 该方法具有非接触、 无前期制样、 无化学分析的优势, 可快速高效的获取煤炭的分类信息。 采用地物光谱仪采集煤炭原始光谱数据, 对噪声过大、 影响后续处理的谱段进行删除, 剩余部分采用小波阈值去噪法进行噪声去除。 将去噪后的数据分成三个数据集: 可见-近红外光谱(500～900 nm)数据集、 短波红外光谱(1 000～2 350 nm)数据集、 全谱段光谱(500～2 350 nm)数据集。 对以上三个数据集进行主成分分析, 将提取出的25个主成分输入多层感知器分类模型。 多层感知器模型由输入层、 隐藏层(两层)、 softmax分类器构成。 对三个数据集进行分类精度的对比, 并采用随机森林(random forest, RF)与支持向量机(support vector machine, SVM)两种分类算法进行进一步的验证分析。 结果表明: 对块状商品煤分类, 全谱段光谱分析技术由于数据信息量丰富, 能够得到更优的分类效果, 在训练样本数为132时, 采用MLP分类器的分类精度最高, 为9803%; 随机森林与SVM的分类结果验证了全谱段数据集的优越性与普适性。 该研究为煤炭的在线分析、 便携式煤炭检测仪器的研发提供了可靠的技术支持。

To obtain the precise result, complex chemical analysis or complicated sample preparation is needed in universal coal analysis methods. In the paper, a new method to distinguish the type of bulk commercial coal using full spectroscopy which combined visible and near-infrared reflectance spectroscopy (Vis-NIRS) and short-wave infrared reflectance spectroscopy (SWIR) analytical technique and Multilayer Perceptron (MLP) classification method was advanced. The method was non-contact with no sample preparation and no chemical analysis. Besides, the classification information of coal can be quickly and efficiently obtained by this method. In the paper, the band range of original spectral data whose noise was excessive was deleted. The noise of remaining part was denoised by wavelet threshold denoising method. The spectral data pretreated was divided into three data sets: Vis-NIRS data set (500~900 nm), SWIR data set (1 000~2 350 nm) and full-spectrum data set (500~2 350 nm). Principal component analysis (PCA) was adopted in three datasets. The extracted principal components were entered in the MLP classification model. Multilayer perceptron was consist of input layer, hidden layers (two layers), softmax classifier. The contrastive study of classification accuracywas made among the three datasets. Random forest and Support Vector Machine (SVM) was used to verification analysis. The research showed: in the classification research of bulk commercial coal, because of the abundant data information of full-spectrum data, a better classification result can be obtained. When the number of training sample was 132, using the MLP classifier can achieve the highest classification accuracy which was 9803%. The classification results of random forest and SVM verified the superiority and universality of the full spectrum dataset. The method provides reliable technical support for on-line analysis of coal and development of portable coal detecting instrument.