为提高焦煤惰质组显微组分分类的准确性，减少对分类器训练的人工干预，提出一种基于改进极限学习机(ELM)的焦煤惰质组显微组分分类方法。首先根据焦煤惰质组各显微组分在光特性及形貌特性上存在的差异及特点，从亮度、纹理等层面分别提取其显微图像中基于灰度统计分布的亮度比、均值、方差、偏度、一致性及峰度等6个亮度相关特征量和基于灰度共生矩阵的能量、熵、惯性矩、局部平稳性及最大概率等5个纹理相关特征量，构建11维初始特征量集，并采用主成分分析法(PCA)对初始特征进行抽取以降低特征空间维数、去除信息冗余;再将奇异值分解引入到极限学习机中，推导利用奇异值分解求解ELM隐含层输出权值矩阵的方法，构建改进的极限学习机。改进后的极限学习机解决了普通的ELM训练中为了求解权值矩阵需要通过大量实验确定参数的问题，有效地提高了学习机的智能化程度。实验结果表明:与支持向量机(SVM)分类方法相比，改进后的ELM方法对分类器训练及样本测试的速度、对焦煤惰质组测试样本分类的准确率均明显提高;与单一的ELM方法相比，改进后的ELM方法分类器的网络训练更加快速便捷，网络隐含层节点数减少近40%，对测试样本分类的准确率进一步提高，可达96.7%。

To improve the classification accuracy of inertinite macerals of coking coal and reduce the manual intervention in the training of classifier,a novel classification approach for the inertinite macerals of coking coal based on improved Extreme Learning Machine (ELM) is proposed. Firstly,according to the characteristics of inertinite macerals and difference between them,a 11-dimensional preliminary feature set about the intensity and texture,including six gray-level statistics based features as contrast,mean,standard deviation,deviation,consistency and kurtosis,and five gray level co-occurrence matrix based features as energy,entropy,moment,local smooth of coal microscopic images and maximum possibility,was built,and extracted with the principal component analysis ( PCA) method to reduce the dimension of feature space and remove redundancy. Secondly,a singular value decomposition ( SVD) method was introduced into the ELM,and the solution to calculate the output weight matrix of ELM was deduced by using SVD,an improved ELM was constructed. After improvement,the problem of parameter training for calculating the output weight matrix,which needs a large number of experiments to determine in conventional ELM,was solved,and the intelligent level of ELM was enhanced. Experimental results show that compared with SVM,the training and testing speed,the classification accuracy for the inertinite testing samples of improved ELM are obviously higher. Compared with the conventional ELM,the network training of classifier is convenient and faster,the number of hidden layer nodes is reduced about 40% , and the classification accuracy for testing samples is further improved, up to 96.7% .