The collaborative patrol of AUV swarms is of great significance for maintaining the security of sensitive sea areas. Conducting research on dispatching strategies and achieving sustainable patrols is of great importance. This paper establishes a collaborative patrol model for AUV swarms based on Double Deep Q-networks (DDQN). Firstly, by defining the absolute importance of each area and building a non-homogeneous task environment map, the patrol problem is transformed into a Non-homogeneous Patrolling Problem (NHPP). The Gaussian distribution function is used to initialize the original map to obtain the importance map. Secondly, elements such as states, actions, and rewards in the non-homogeneous patrol model were established. A patrol model was constructed based on the DDQN learning algorithm to achieve regular patrols of sensitive and important areas. Finally, the effectiveness of the non-homogeneous patrol method was verified through simulation experiments.