With the increase of software defined optical network(SDON) scale, multiple controllers are needed to provide services for the whole network. The load balancing problem of controllers affects the service capability and network survivability of the whole network. A self-learning load balancing algorithm for SDON controller based on Bayesian network is proposed. Firstly, considering the controller parameters such as controller load, controller throughput and optical switch migration delay, Bayesian network is used to predict the degree of load congestion; secondly, the optimal action decision is made by combining reinforcement learning algorithm, and the feedback mechanism is used to realize the self-adjustment of parameter weights so as to adjust the control. Load balancing can be achieved by the degree of load congestion. The simulation results show that compared with Q-learning algorithm, the proposed algorithm achieves better load balancing function of the controller and improves load balancing efficiency more quickly.