Monitoring aero-engine flame states—such as stable combustion, ignition, and extinction, is critical for propulsion system safety and performance optimization, with image analysis serving as a key detection technique. However, the high dynamic range and resolution of flame images challenge conventional deep learning models due to limited generalization and convergence difficulties. Convolutional Neural Networks (CNNs) excel at local feature extraction but are constrained by small receptive fields, hindering effective global context modeling. On the other hand, Vision Mamba captures long-range dependencies through state space modeling but struggles to preserve fine-grained visual details in high-resolution inputs. To address these limitations, we propose a lightweight yet effective hybrid architecture that integrates ResNet-18 for robust local feature encoding with an enhanced Vision Mamba module for efficient global dependency learning. The proposed model is evaluated on an industrial high-speed flame image dataset and achieves 97.8% classification accuracy, outperforming several state-of-the-art methods. Our results suggest that this architecture offers a promising solution for aero-engine combustion monitoring, balancing model accuracy, inference speed, and computational efficiency for practical deployment in aerospace scenarios.