Steam generator liquid level control system (SGLCS) is an important energy exchange equipment in nuclear power plant, and its control performance has an important impact on the safety, economy and efficient operation of nuclear power plant. The performance optimization process of this system has prominent problems such as time consuming, experience dependence and difficulty to achieve the best. In order to improve the performance optimization efficiency of the control system, a SPSA method (CG-SPSA) based on data-driven optimization idea and SPSA method. This method focuses on system performance measurement. Firstly, the dynamic reconstruction mechanism of iterative performance data is designed; Then, the gradient approximation is realized; then, the organic fusion of model gradient approximation and data gradient approximation at the dynamic confidence level is realized, and the optimal estimation of SPSA gradient approximation value is realized. Under the data-driven framework, this method provides a unique idea and implementation mechanism of data-driven and model-driven fusion, which maximizes the use of iterative measurements and effectively improves the efficiency of SPSA method. Simulation experiments show that the traditional SPSA and model-free optimization framework, and can significantly improve the efficiency of the level control performance of steam generator.