针对机器学习技术的固有不确定输出特性给航空器适航安全性定量评估造成的挑战，在SAE ARP4761标准框架下，基于泛化理论提出一种面向卷积神经网络（CNN）分类功能的系统安全性评估方法。首先，基于PAC-Bayes理论，通过条件高斯分布改进训练方法，优化泛化界，获取CNN模型不确定性量化表示；其次，实现软件不确定性与硬件可靠性融合，支持整机/系统的定量安全性评估。最后，以基于CNN的GNSS干扰信号识别模块装机为案例，表明该方法对适航安全性评估的有效支撑作用，同时也实验验证基于条件高斯的泛化界比普通PAC-Bayes及VC维都具有更紧的计算边界。

Aiming at the airworthiness safety challenges caused by the inherent uncertainty ouputs of machine learning technology in airborne systems, a method based on generalization theory is proposed for CNN classification model under the framework of SAE ARP4761 standards. First, based on PAC-Bayes theory, the training method is im-proved through conditional gaussian process to optimize the generalization bound and obtain a quantified representation of the uncertainty of the CNN model. Second, the uncertainty and the hardware reliability is integrated, which supports quantitative safety assessment of the system. Finally, taking airborne GNSS interference signal recognition module based on CNN as a case, it shows the effective of the proposed method on safety assessment, and also experimentally verifies that the generalization boundary based on conditional gaussian process has a tighter computational boundary than that of ordinary PAC-Bayes and VC dimensions.