基于MSSA-XGBoost小样本核爆地震事件分类  

Classification of Small Sample Nuclear Explosion Seismic Events based on MSSA-XGBoost

作　　者：李鸿儒 李夕海 谭笑枫 张云 刘天佑 刘继昊 牛超 Hongru Li;Xihai Li;Xiaofeng Tan;Tianyou Liu;Yun Zhang;Jihao Liu;Chao Niu(School of Nuclear Engineering,Rocket Force University of Engineering,Xi’an 710025,China)

机构地区：[1]火箭军工程大学核工程学院,陕西西安710025

出　　处：《Applied Geophysics》2024年第1期108-118,205,共12页应用地球物理（英文版）

基　　金：supported by Natural Science Basic Research Program of Shaanxi (2023-JC-YB-244)。

摘　　要：核爆地震与天然地震事件的分类鉴别是全面禁止核试验条约中的一项重要任务。受限于核爆数据数量缺乏,论文研究了XGBoost模型在小样本场景下两类事件的分类问题,并利用SSA算法对模型关键超参数进行自主寻优。同时针对SSA算法的不足,采用高斯混沌映射方法、提出种群比例动态调整策略和引入步长调整因子进行改进,构建了MSSA-XGBoost分类模型。模型解决了初始种群分布不均匀,导致种群多样性减少,影响算法收敛速度的问题;解决了麻雀种群比例固定,容易陷入局部最优解的问题;以及解决了发现者位置更新步长固定,从而限制算法全局搜索能力与寻优效率的问题,实现了避免人工特征提取和XGBoost迭代次数、最大树深以及学习率三个重要超参数的自主寻优,在小样本地震事件分类中取得了优异的效果。实验结果表明,MSSA-XGBoost模型分类准确率达到了96.37%,优于原模型93.47%,也优于支持向量机与卷积神经网络,同时较原始模型计算效率提升了近30%。The classification and distinction between nuclear explosions and natural earthquake events are essential to the Comprehensive Nuclear Test Ban Treaty.Nuclear explosion data are lacking;thus,classification problems must be studied in small sample scenarios.The classification problem of the eXtreme Gradient Boosting(XGBoost)model in one small sample scenario is examined using the sparrow search algorithm(SSA)algorithm to optimize the key hyperparameters of the model automatically.The shortcomings of SSA are addressed by using a Gaussian chaotic mapping method,introducing a population proportion dynamic adjustment strategy,and proposing a step-size adjustment factor for modification.The problem of the uneven initial population distribution is addressed by constructing the(modified SSA)MSSA–XGBoost classification model,thereby reducing population diversity and affecting the convergence speed of the algorithm.The fixed proportion problem of the sparrow population,which easily falls into the local optimal solution,is solved using the aforementioned approach.The fixed update step position of the discoverer is also resolved,thus limiting the global search capability and optimization efficiency of the algorithm and realizing the independent optimization of three important hyperparameters.Furthermore,artificial feature extraction can be avoided using this approach,and the number of iterations,maximum tree depth,and learning rate can be automatically optimized,achieving excellent results in small sample seismic event classification.Experimental results reveal that the classification accuracy of the MSSA–XGBoost model is 96.37%,demonstrating its superiority to the original model(93.47%)as well as to the support vector machine and convolutional neural network.Meanwhile,a nearly 30%improvement is observed in computational efficiency.

关 键 词：核爆地震分类 XGBoost 麻雀算法 小样本 特征提取 

分 类 号：P631.4[天文地球—地质矿产勘探]