Identification of lean blowout precursors in swirling spray flames via CH* chemiluminescence imaging and a deep learning framework

Lean blowout warning is critical for the safe operation and control optimization of aero-engine combustors. However, achieving robust real-time warning under conditions of high-frequency sampling and rapidly evolving flame states remains challenging. To address these challenges, this study proposes a lean blowout precursor identification method for rapid warning analysis that integrates flame CH* chemiluminescence imaging processing with a deep learning framework. First, a convolutional autoencoder (CAE) is employed for preliminary screening of flame images, achieving an F1 score of 97.12%. Bayesian optimization (BO) is subsequently used to adaptively search for the optimal region of interest (ROI), where optimal ROIs at different scales are concentrated within a common spatial neighborhood. Finally, a deep belief network (DBN) is established to model and identify critical precursor patterns. The cumulative-distribution analysis of reconstruction error demonstrates that the ROI size of 80 × 144 pixels yields the best discrimination performance, exhibiting the largest separation between pre-blowout and non-blowout samples. While maintaining the warning effectiveness, the proposed method significantly improves computational efficiency. At this ROI size, processing 5000 images takes only 307 ms. Although the current end-to-end framework is still limited by the computational cost of CAE-based screening, the results demonstrate the potential of the proposed approach for rapid warning analysis and provide guidance for future lightweight online implementations.