Over-expansion flow and separation phenomena exist inside the nozzle during the startup and shutdown of the rocket engine, and the transition between the separation modes is prone to cause significant side loads and jeopardize the safety of the engine. Accurate, comprehensive and prompt nozzle flow field state perception is of great significance to the stable operation of the engine. Here, a generative adversarial network is proposed to reconstruct the internal Mach number and temperature fields for different nozzle pressure ratios and flow separation modes based on the discrete pressure at the nozzle wall. The proposed generative adversarial network employs Wasserstein distance and gradient penalty to improve the problems of gradient vanishing and pattern collapse during discriminator network training. The large expansion ratio nozzle selected was the VOLVO S1 nozzle and numerical simulations were carried out under different nozzle pressure ratio conditions to obtain data for the training of data-driven model. The results indicate that generative adversarial networks can accurately reconstruct restricted shock separation and free shock separation modes under different nozzle pressure ratios based on discrete pressure and temperature measurement points on the nozzle wall, with a relative error of less than 1.2%, and accurately identify important flow field characteristics such as Mach disk, supersonic jet flow, and multiple recirculation regions. The advantages of the generative model in recognizing the strong normal shock Mach disk are compared with those of the supervised learning model. This study lays the foundation for the subsequent study of nozzle side load and structural fatigue life based on three-dimensional flow field.