震灾人员伤亡预测的改进SVM模型及其应用

The improved SVM model and its application under earthquake casualty's prediction

震灾人员伤亡预测是有效确定应急物资筹集量和及时掌握应急物资需求变动的基础性工作。本文将改进的支持向量机(Support Vector Machine,SVM)运用到震灾人员伤亡预测模型构建中,提出鲁棒小波v-SVM的震灾伤亡预测模型。通过设计一种能够有效压制数据大幅值和奇异点的鲁棒损失函数,使其有效处理震灾伤亡预测指标中出现的各类数据。为改变常规核函数缩小偏差的局限性,本文将Morlet和Mexican两类母小波核函数的自变量用满足Mercer平移不变核的小波核函数进行替换,得到用于机器学习的两类小波核函数。数字算例表明:用于预测震灾伤亡的鲁棒小波v-SVM模型具有学习速度快、预测精度高和稳定性强的特点,这为震灾伤亡人口的预测提供了有效方法。

Earthquake casualty＇s prediction is a fundamental work to effectively determine how much emergency supply to gather, and track the real-time demand variation of emergency supply demand. In the earlier stage of earthquake, the information for casualty prediction is lacking, and the relation between predictors is nonlinear. This paper applied the improved Support Vector Machine（SVM） to build an earthquake casualties prediction model, and proposed an earthquake casualties prediction model based on robust wavelet v-SVM. Considering the disadvantage that the single loss function in SVM is not able to suppress the large range and singular point of earthquake casualties index, we combine Gaussian loss function, Laplace loss function, ？-insensitive loss function, and a robust loss function that allow for segmented suppression to handle different indices in casualties prediction. In order to minimize the nonlinear classification error of SVM in higher dimensional space, the independent variables in both Morlet and Mexican parent wavelet kernel functions are replaced with a wavelet kernel function satisfying Mercer translation invariant kernel. Thus, two wavelet kernel functions are obtained for machine learning in order to mitigate the limitation of convenient kernel function, and reduce errors. In the first part, this paper considered the characteristics of nonlinearity, randomness, and singularity under earthquake disaster, and designed a comprehensive robust loss function. The function could suppress data noise through different functions, including Gaussian loss function, Laplace loss function, and insensitive loss function to improve the noise reduction processing capacity of SVM. In the second part, Kernel function of SVM is to solve the high dimension space of inner product operation and it can minimize nonlinear classification errors. At present, the commonly kernel functions included radial basis kernel function, Fourier kernel function, Sigmoid kernel function, polynomial kernel function, and linear kernel function. These kernel functions can effectively reduce the error, but the effect is not ideal. In order to improve SVM classification error in high dimensional space, the paper changed the limitation of single kernel function by reducing errors. By replacing the independent variables of two types of mother wavelet kernel functions of Morlet, we derive satisfying translation invariant wavelet kernel function of Mercer in order to build robust wavelet v-SVM forecasting model of emergency materials. In summary, the numerical examples show that robust wavelet v-SVM model features rapid learning, high-precision prediction, and advanced stability when the model is used to predict earthquake casualties, which provides an effective method to predict earthquake casualties.