为快速、准确地预测矿区采空塌陷的危险性,针对矿区采空塌陷预测的复杂非线性特点,在统计分析实测资料的基础上,选取7项指标作为初始特征指标,30组塌陷样本作为原始样本,其中,前17组为原始训练样本,后13组为测试样本;运用粗糙集(RS)理论,对原始训练样本进行对象约简和属性约简。将属性约简后的3项指标作为支持向量机(SVM)的输入向量,建立矿区采空塌陷危险性预测的RS-SVM模型。将对象约简后的7组样本作为训练样本,用于模型训练。采用回代估计法进行回检,误判率为0。利用训练好的模型对13组待评样本进行预测,并与贝叶斯、BP神经网络(BPNN)方法进行比较。结果表明,RS理论与SVM算法相结合,能降低属性维数,去除冗余样本,简化模型,该模型所得预测结果准确率为100%。 In order to quickly and accurately predict the danger of mined-mined collapse in mining areas, aiming at the complex nonlinear characteristics of mined-mined collapse prediction in mining area, seven indexes are selected as initial characteristic indexes based on the statistical analysis of measured data. Thirty groups of collapsed samples are used as primitive Samples, of which the first 17 groups for the original training samples, the latter 13 groups for the test sample; the use of rough sets (RS) theory, the original training samples for object reduction and attribute reduction. The RS-SVM model of risk prediction of goaf collapse is established by using the three indexes after attribute reduction as input vectors of Support Vector Machine (SVM). 7 groups of samples after object reduction were used as training samples for model training. Using the back-estimation method to conduct a retest, the false positive rate is 0. Thirteen groups of samples to be evaluated were predicted using the trained model and compared with Bayesian and BP neural network (BPNN) methods. The results show that the combination of RS theory and SVM algorithm can reduce the number of attributes, remove redundant samples and simplify the model. The accuracy of the prediction results obtained by this model is 100%.