根据岩爆的影响因素、特点及成因,选取围岩最大切应力、单轴抗压强度、单轴抗拉强度、应 力系数、脆性系数和弹性能量指数构成岩爆预测指标体系。 搜集国内外 46 组典型岩爆案例数据, 考虑到概率神经网络(PNN)中高斯函数要求各指标变量互不相关,采用主成分分析法(PCA)对原 始数据预处理,消除指标间相关性并降维,得到线性无关的 3 个主成分即岩爆综合预测指标 RCI1 , RCI2 ,RCI3 ,构成概率神经网络的输入向量。 将岩爆烈度分级预测视为共有 4 种类别的模式分类问 题,在满足均匀分布的前提下,选取[0.02,1.00] 内的 50 个 Spread 值,观察模型预测正确率随 Spread 值的变化。 经测试,Spread 值为 0.36 时,预测结果首次同时达到最优,故创建平滑因子为 0.36 的概率神经网络。 岩爆案例数据由主成分分析法处理后分为训练样本和测试样本,对训练后 的 PNN 网络进行性能测试,两组数据预测正确率分别为 100% ,90% 。 将该结果与随机森林(RF) 模型、支持向量机(SVM)模型、人工神经网络(ANN)模型进行比较,可知 PCA-PNN 模型的预测结 果稍好于 SVM 模型、ANN 模型,误判率与 RF 模型的训练样本平均误判率、测试样本平均误判率一 致。 此外 PNN 网络收敛速度快,通常在数秒内即可完成,表明基于 PCA-PNN 的岩爆烈度预测模 型是合理可行的。

According to the influ-encing factors,the characteristics and causes of rockburst,the maximum tangential stress of surrounding rock,the uniaxial compressive and uniaxial tensile strength of rock,the stress coefficient,brittle-ness coefficient and elastic energy index of rock are chosen to form the rockburst prediction indexes system. In this study,about 46 groups of typical rockburst engineering samples are gathered internationally. Considering that the Gaussian function of the probabilistic neural network ( PNN) requires that any index variable is not correlated with others,initial rockburst engineering samples are preprocessed by using the principal component analysis (PCA) in or-der to eliminate the correlation between different indexes and realize a dimension reduction of data to get three linearly independent principal components that can be defined as rockburst comprehensive prediction indexes RCI1 ,RCI2 and RCI3 ,which make up the input vectors of the PNN. The prediction of rockburst intensity is regarded as the problem of classification with four different patterns. About 50 Spread values, which obey the uniform distribution in [ 0． 02,1.00],are selected to observe the change of accuracy of the prediction model. After testing,when the Spread value is 0． 36,the prediction results are simultaneously optimal for the first time,therefore,the probabilistic neural network with a smoothing parameter of 0． 36 is created. The rockburst engineering samples processed by the principal component a-nalysis are divided into training samples and testing samples,which are input into PNN separately for testing the per-formance of the PNN. The correct rates of the two prediction results are 100% and 90% respectively. Compared the prediction results with the results of Random Forest (RF) model,Support Vector Machines (SVM) model and Artifi-cial Neural Network (ANN) model,the results of this study is slightly better than that of SVM model and ANN model, and two misjudgment rates are consistent with the average misjudgment rates of the training samples and the testing samples of the RF model. In addition,the calculation results of PNN model converges fast,which are usually completed in a few seconds,indicating that the PCA-PNN model is reasonable.