为合理高效地进行公路隧道施工坍塌风险评估，通过粗糙集（rough set, RS）理论、网格搜索法（grid search, GS）和支持向量机（support vector classification, SVC）研究了公路隧道施工坍塌风险评估模型。首先融合超前地质预报，构建公路隧道施工坍塌风险评价指标体系，同时收集100个隧道坍塌相关案例信息并对指标数据进行离散化处理，其次基于粗糙集条件信息熵进行属性约简，得到约简后的核指标集，而后采用网格搜索法寻找支持向量分类训练集的最优参数，建立基于粗糙集-网格搜索-支持向量（RS-GS-SVC）公路隧道施工坍塌风险评估模型，最后将所建模型用于对测试样本的预测。结果表明：在相同学习样本的条件下，相较于粗糙集-遗传-支持向量模型（RS-GA-SVC）和粗糙集-粒子群-支持向量模型（RS-PSO-SVC），RS-GS-SVC模型具有更高的分类精度；在训练集与测试集比例相同的条件下，RS-GS-SVC模型的预测准确率高于GS-SVC模型，准确率分别为93.33%和90%，且RS-GS-SVC模型的运算时间更短。可见，经粗糙集条件信息熵属性约简，可以有效降低模型复杂度，提高分类精度。

In order to reasonably and efficiently carry out the risk assessment of road tunnel construction collapse, the risk assessment model of road tunnel construction collapse was studied by rough set (RS), grid search method(GS) and support vector classification(SVC). Firstly, the index system of highway tunnel construction collapse risk evaluation was constructed by integrating advanced geological prediction. At the same time, the information of 100 tunnel collapse cases was collected and the index data was discretized. Secondly, attribute reduction was conducted based on the condition information entropy of rough set to obtain the reduced core index set. Then, grid search method was used to find the optimal parameters of the support vector classification training set, the risk assessment model of highway tunnel construction collapse based on rough set-grid search-support vector classification (RS-GS-SVC) was established. Finally, the model was used to predict the test samples. The results show that under the condition of the same learning sample, compared with rough set-genetic algorithm-support vector classification (RS-GA-SVC) model and Rough set-particle swarm optimization-support vector classification (RS-PSO-SVC) model, RS-GS-SVC model has higher classification accuracy; Under the same proportion of training set and test set, the prediction accuracy of RS-GS-SVC model is higher than that of GS-SVC model, with the accuracy rates of 93.33% and 90% respectively, and the operation time of RS-GS-SVC model is shorter. It can clearly be seen that the model complexity is effectively reduced and the classification accuracy is improved through the reduction of rough set conditional information entropy attributes.