In a rocket based combined cycle engine, the incoming flow parameters change drastically during the wide range working process. Thus, the isolator is highly susceptible to overflow conditions resulting in engine performance degradation or even non-starting conditions. To broaden the stable working boundary, we design a deep learning-based flow field reconstruction method to obtain the flow field information within an isolator. The two-dimensional velocity field can be reconstructed rapidly by studying the one-dimensional wall pressure of the isolator. This method we proposed can realize the rapid sense of its own abnormal state, which provides the necessary conditions for timely engine regulation and the robustness improvement. To optimize the effectiveness of the method, we compare the reconstruction accuracy of the velocity field for four typical neural network models. The results show that the MPFC-CNN neural network has high reconstruction accuracy. In this way, the starting positions of the separation zone are accurately reconstructed, and the morphology of the center high-speed core flow shock train is obtained. It is indicated that the model can accurately identify the working state of the isolator, and provide accurate feedback for the efficient control of the engine. Therefore, this study has obviously high value in engineering application for the combustion stabilization and wide-range performance improvement of the combined cycle engine.