The rapid development of deep reinforcement learning (DRL) has offered a promising paradigm for active flow control research. Most current research is based on numerical simulation, allowing for the arbitrary placement of sensors. However, for practical engineering applications, determining the placement of sensors and using data from fewer sensors to reflect the state of the flow field is crucial. In this study, a rotating flap is implemented on the NACA0012 airfoil to represent the wing of unmanned aerial vehicles (UAVs) situated in the wake of a cylinder to simulate its working environment, and the objective is to minimize lift fluctuations by adjusting the angle of the rotating flap. Specifically, computational fluid dynamics (CFD) results are used as the environment, adopting the attention branch neural network (ABN) based deep deterministic policy gradient (DDPG) algorithm to derive the control strategy and obtain the attention weight distribution. 7 sensors with the highest attention weight (HAW) and 7 with the lowest attention weight (LAW) were selected, respectively, to replace all 66 sensors for the agent. Training based on 7 HAW sensors converges faster and yields a better flow control effect than the 7 LAW sensors. Moreover, the control effect and generalization ability based on 7 HAW sensors are not significantly different from using all 66 sensors, both reducing the standard deviation of the lift coefficient by about 78%. This demonstrates that sensor placement based on attention mechanism is effective, offering a reliable method for practical engineering applications.