Combat networks can accelerate the closure of the kill chain in a combat system-of-system and thus multiply combat effectiveness, but they also face the threat of focused destruction. In order to effectively and quickly recover and even improve the robustness of combat networks, the cascade failure modeling and robustness recovery methods of combat networks are studied. Firstly, to address the bias in combat network modeling, a two-layer heterogeneous group-dependent combat network model is constructed from real world. Then the conditional group-dependent failure, non-connected failure and critical overload cascading failure processes are analyzed and designed. Further, a network robustness index with operational significance is proposed. Considering the limitation of timelines and recovery resources, a prior recovery based on capacity and importance (PRCI) of boundary nodes is proposed, taking into account the attribute features of the combat network. Finally, the effectiveness and feasibility of the proposed method are validated through simulation experiments, which involve comparing and adjusting model parameters using different approaches. The results show that the recovery performance of the PRCI method is significantly better than other benchmark methods with fast onset and fewer iterations, enabling rapid and effective restoration of combat network under same conditions. Additionally, it is also found that the recovery performance of the method is proportional to the tolerance, capacity parameters, overload bearing coefficients and recovery ratio, while inversely proportional to the load parameters. These findings further provide insights for the structural optimization of combat networks.