In optical communication networks,the perception of communication signals can be divided into single-node perception and multi-node cooperative perception.Amo ng them,multi-node cooperative sensing is widely studied because it can overcom e the effects of multipath fading and hidden terminals.However,there are still m any problems to be solved in the communication overhead and fusion method of mul ti-node cooperative sensing.In this paper,a quantitative and coding method is p roposed ot reduce the communication overhead between the local node and the Fusi on center,and a weighted fusion method based on node feature is proposed to impr ove the judgment accuracy of fusion center,the fusion method can effectively ref lect the actual role of each node in cognitive network.In addition,reinforcement learning is used to improve the sensing efficiency in familiar environment.The experimental results show that the proposed spectrum sensing method can reduced communication overhead,improve decision accuracy and reduce sensing time under t he same conditions as traditional methods.