This paper proposes an VDAM-MASAC algorithm for UAV target search scenarios with unknown prior information. The algorithm models the UAV cluster tracking and searching problem as a distributed partially observable Markov decision process, and designs a reward function to guide autonomous UAV learning. It achieves balanced optimization among target searching, tracking, path planning, and collaborative collision avoidance. To address the dimension explosion issue in traditional multi-intelligence reinforcement learning, the joint action value function is decomposed into individual action value combinations, reducing network complexity. Additionally, an attention mechanism dynamically adjusts action value weights based on real-time environment states, enhancing adaptability to dynamic scenes. Simulation experiments with three UAVs and three targets in the OpenAI multi-intelligence particle environment show high average tracking success rates, better average reward convergence values compared to MASAC and QPLEX, and maintained safe distances between UAVs, demonstrating the algorithm’s efficient collaborative ability in unknown environments.