Abstract:A novel hierarchy support vector machines(H-SVMs) model is presented to recognize the headstreams of water inrush in coal mine. Firstly,an analytical model is deduced to analyze the generalization power of H-SVMs. According to the results,a feasible approach is put forward to improve the performance of H-SVMs to guarantee the performances of each SVM node,whose position is located at a high level. Secondly,a novel method is presented to build H-SVMs,i.e. MMH-SVMs(maximal margin hierarchical SVMs),taking the separating margins of each SVM node as indices for classification and clustering,using TopDown and BottomUp routes from top to bottom to classify the input samples at each SVM node by maximal separating margin and from bottom to top clustering the input samples by minimal separating margin. Experimental results show MMH-SVMs have a simple structure,and a good generalization performance. It can predict the headstreams of water inrush correctly;and its tree structure can also denote the hierarchy of headstreams. Moreover,the normal vector parameter W in each SVM decision function can describe the weights of discrimination indices of the headstreams of water inrush,in which a novel scientific method is introduced to predict the headstream of water inrush in coal mine.