为提高复合材料加筋圆柱壳后屈曲分析和优化效率，提出了一种融合样本分簇和改进K折交叉验证的增广径向基函数（ARBF）快速近似建模方法。采用K-means聚类算法确定了样本最优分簇，基于样本局部密度确定了样本的基准形状参数，各分簇均引入缩放系数进行自适应调整形状参数，有效兼顾了优化形状参数的效率和精度。进而建立了基于偏差-方差分解的ARBF近似模型泛化性能评估准则，解决了传统K折交叉验证样本信息利用不足的难题；基于分块矩阵求逆技术推导了ARBF辅助近似模型的高阶系数矩阵快速求逆方法，据此提出了基于改进K折交叉验证的缩放系数优化方法，大幅降低了确定最优形状参数的计算复杂度，提升了ARBF近似建模效率和精度。数值和工程算例表明，样本最优分簇和快速交叉验证对近似建模效率和精度有显著增益，降低了建模效率对样本规模和问题维度的敏感性，且相同训练样本数量下，本文方法建模精度显著优于其他典型方法，验证了该方法的有效性和先进性，具有一定的工程应用价值。

In order to improve the analysis and optimization efficiency of stiffened composite cylindrical shells, a novel fast modeling method for the augmented radial basis function (ARBF) approximate model is proposed by fusing sam-ple clustering and improved K-fold crross validation.The K-means clustering algorithm is used to determine the optimal sample clustering, and the reference shape parameters of the samples are determined based on the local density of the samples. The scaling coefficient is introduced into each cluster to adjust the shape parameters adap-tively, which effectively taking into account the efficiency and accuracy of optimizing the shape parameters.Then, the evaluation criteria of ARBF approximation model generalization performance based on bias-variance decom-position is established, which solves the problem of insufficient utilization of traditional K-fold cross-validation sample information. Based on the block matrix inversion technique, a fast inversion method of the high-order coef-ficient matrix of the ARBF auxiliary approximation model is derived, and the scaling coefficient optimization method based on the improved K-fold cross-validation is proposed, which greatly reduces the computational complexity of determining the optimal shape parameters and improves the efficiency and accuracy of ARBF approximation mod-eling.Numerical and engineering examples show that the optimal sample clustering and fast cross-validation have significant gains on the approximate modeling efficiency and accuracy, and reduce the sensitivity of modeling effi-ciency to sample size and problem dimension. Moreover, under the same number of training samples, the model-ing accuracy of the proposed method is significantly superior than that of other typical methods, which verifies the effectiveness and advancement of the method and has certain engineering application value.