针对现有基于神经网络的气动参数辨识方法只能获得气动系数总量而难以分离其内部组成分量的问题，提出了一种基于气动增量神经网络模型（Aero-INN）的辨识框架，可在先验气动模型结构未知的前提下分离气动系数的各组成分量，从而实现具有物理可解释性的神经网络辨识。该框架基于气动增量建模思想，不直接对气动系数总量进行端到端建模，而是以神经网络替代传统气动导数模型中的气动导数参数，通过“直连相乘”的网络结构设计嵌入增量物理约束，将气动力表示为基本子网络与若干物理意义明确的增量子网络之和，使模型输出符合物理直觉的气动增量表达。在此基础上，提出“先分量提取，后白箱重构”的两步走辨识流程，通过规则网格采样与符号回归技术，将增量子网络白箱化为显式插值表与解析表达式。验证工作分别从三轴激励飞行仿真数据和近失速真实飞行数据两个层面开展。研究结果表明，所提出方法在保证气动系数总量预测精度的同时可有效实现内部气动分量的可解释分离，通过将高维白箱化问题降维为若干低维子空间的白箱建模问题，降低了白箱化重构的难度，为神经网络非线性辨识的可解释性提升以及实际工程中的气动力数据库天地相关性修正提供了新的技术途径。

To address the problem that existing neural-network-based aerodynamic parameter identification methods can only obtain the overall aerodynamic coefficients but fail to separate their internal constituent components, an identification framework based on the Aerodynamic Incremental Neural Network (Aero-INN) is proposed. The framework is capable of separating the constituent components of aerodynamic coefficients without requiring a priori knowledge of the aerodynamic model structure, thereby achieving physically interpretable neural-network-based identification. The framework is developed based on the concept of aerodynamic incremental modeling. Instead of directly performing end-to-end modeling of the total aerodynamic coefficients, neural networks are employed to replace the aerodynamic derivative parameters in conventional models, such that the aerodynamic forces are expressed as the sum of a baseline aerodynamic subnetwork and several increment subnetworks with clear physical meanings. By embedding incremental physical constraints into the network architecture, the model outputs aerodynamic increments that conform to physical intuition. On this basis, a two-step identification strategy of “component extraction followed by white-box reconstruction” is proposed, in which the extracted increment subnetworks are converted into explicit interpolation tables and analytical expressions via regular grid sampling and symbolic regression. Validation is conducted using both three-axis excitation flight simulation data and near-stall real flight data. The results demonstrate that the proposed method ensures prediction accuracy of the total aerodynamic coefficients while enabling interpretable decomposition of internal aerodynamic components. By reducing the high-dimensional white-box modeling problem to white-box modeling in several low-dimensional subspaces, the difficulty of white-box reconstruction is significantly lowered, providing a new technical approach for improving the interpretability of neural-network-based nonlinear identification and for correcting the ground-to-flight correlation of aerodynamic databases in practical engineering applications.