针对具有未知但有界（ＵＢＢ）误差的线性回归模型辨识问题，提出了一种新的鲁棒结构选择方法．该方法以重复递推椭球外界算法所得椭球轴信息阵的行列式相对值最大作为模型定阶准则．不同于以往对噪声独立性、常方差或鞅差特性的假设，该方法假设噪声是渐近独立的．文中证明了该方法的强相容性 Aiming at the problem of linear regression model identification with unknown but bounded (UBB) errors, a new robust structure selection method is proposed. The method takes the maximum of the relative determinant of the information matrix of the ellipsoid axis obtained from the algorithm of repeating recursive ellipsoid as the model order criterion. Unlike previous assumptions about noise independence, constant variance or martingale difference characteristics, this approach assumes that noise is asymptotically independent. The paper proves the strong compatibility of this method