Single Event Upsets Fault Tolerance of Convolutional Neural Networks Based on Adaptive Boosting

Single-event upsets in the space radiation environment pose a serious threat to the reliability of satellite-borne intelligent systems. Traditional fault-tolerance methods such as triple modular redundancy and periodic scrubbing face issues like high resource overhead and power consumption. This paper proposes a lightweight fault-tolerance method based on an adaptive boosting algorithm (AB-FTM), which constructs a heterogeneous ensemble architecture of ResNet20/32/44 weak models. While reducing the parameter scale by 18.2% compared to the original ResNet110, it improves classification accuracy and robustness through a dynamic weight adjustment mechanism. Experimental validation on datasets including CIFAR-10, MNIST, and EuroSAT shows that when 0.0004% of parameters experience single-event upsets, the proposed method improves accuracy by 20.39%, 26.25%, and 21.02% respectively compared to the ResNet110 baseline model, significantly outperforming existing fault-tolerance solutions. This method provides a new solution for future space science satellites using satellite-borne intelligent systems that balances reliability, lightweight design, and computational efficiency.