UAV swarm formations rely on coordinated multi-UAV interactions to execute complex missions, but faults in individual nodes may influence formation coordination and degrade system safety. Therefore, timely fault detection is essential. However, labeled swarm fault data are scarce, and multivariate flight observations lack a unified representation. Moreover, fault may initially be weak and remain undetected until their effects accumulate, increasing detection delay. To address these challenges, a two-stage method for early fault detection in UAV swarms is proposed. In the first stage, the distance and velocity deviations are integrated to represent consistency state. Then, the two variables, together with acceleration, attitude angles, and velocity commands, are used as node features, while horizontal distance between the two UAVS is used as the edge feature to construct the swarm interaction graph. In the second stage, an attention-based cellular automata model integrates temporal encoding,neighbor-interaction learning, and cellular state evolution to predict multi-step consistency states using only normal flight data. Fault events are then identified by comparing the smoothed prediction residual with a statistical threshold derived from normal data. Finally, Experiments on UAV swarm formation simulation platform with 20 UAVs have been conducted to validate the effectiveness and timeliness of this method. Results show that the method successfully detected all the injected fault events. Compared with baseline methods, the proposed method achieved the highest accuracy of 96.85%, the highest recall of 95.56%, and the shortest mean detection delay of 0.89 s.

UAV swarms,fault detection,attention mechanism,cellular automata