Stable combustion is desirable for efficient operation of scramjet engines at high flight Mach numbers, and being able to reconstruct the flow-field wave patterns in stable combustion facilitates proactive evaluation of engine operating conditions. Proposed here is a multi-level connected shifted-window transformer (MCSwinT) model for reconstructing the flow-field wave patterns of stable combustion in a supersonic combustor. A combustion feature conversion block is used to convert high-dimensional and low-dimensional combustion features; a deep pressure feature extraction block is used to extract the flow-field wave patterns, and MCSwinT blocks enable multi-level fusion, thereby extracting the high-dimensional combustion features of the flow-field wave system. A dynamic loss function unifies spatial content loss and feature space loss, leading to enhanced reconstruction results. Separately, data on the stable combustion process of a hydrogen-fueled scramjet engine were collected in a direct-connect supersonic pulse combustion wind tunnel, and these data are used to validate the robustness and generalization capability of MCSwinT. The experimental results show that the flow-field wave patterns of stable combustion are reconstructed successfully using MCSwinT of different scales. Compared to other models, MCSwinT exhibits lower model complexity while achieving performance improvements of 7% and 17% in peak signal-to-noise ratio and structural similarity index, respectively. Additionally, the high generalization ability of the proposed model is validated in a sparsity experiment. This model effectively reconstructs the flow-field wave patterns of stable combustion, providing a crucial foundation for further research on scramjet engines.