针对航空发动机燃烧室点火预测中工况特征复杂，训练样本量小且数据不平衡，一般深度学习预测模型难以满足预测任务需求的问题，提出了融合残差连接与自注意力机制的网络模型——注意力残差视觉几何组网络（attention residual visual geometry group,Att ResVGG）。该模型通过多头自注意力机制捕捉工况参数间的依赖关系，建立工况参数与点火状态之间的映射关系。针对数据量不足和类别不平衡问题，设计了基于物理约束的数据增广策略，通过维持燃油空气比、温度压强比等关键物理参数关系合成新工况样本。此外，设计基于贝叶斯优化自动机器学习算法优化模型超参数，进一步提升模型预测能力。为验证模型效果，在两个数据集上的实验表明，AttResVGG模型的准确率分别达97.67%和90.48%,Kappa系数分别达0.950 5和0.834 6，优于所对比的模型。
To address the challenges posed by complex operating conditions, limited training samples, data imbalance, and the inability of conventional deep learning models to meet the requirements of ignition prediction for aero-engine combustion chambers, a network model integrating residual connections and a self-attention mechanism, called the attention residual visual geometry group network(AttResVGG), was proposed. The model used a multi-head self-attention mechanism to capture dependencies among operating condition parameters and establish a mapping between these parameters and ignition status. To address insufficient data size and class imbalance, a physically constrained data augmentation strategy was designed to synthesize new operating condition samples while maintaining key physical parameter relationships, such as the fuel-air ratio and temperature-pressure ratio.In addition, an automated machine learning algorithm based on Bayesian optimization was designed to optimize model hyperparameters, further enhancing the model's predictive performance. To validate the effectiveness of this model, experiments on two datasets show that the accuracy of the AttResVGG model reaches 97.67% and 90.48%, and the Kappa coefficients reach 0.950 5 and 0.834 6, respectively, which are better than those of the compared models.
[1] Poinsot T. Prediction and control of combustion instabilities in real engines [J]. Proceedings of the Combustion Institute, 2017, 36(1): 1-28.
[2] Brunton S, Noack B, Koumoutsakos P, et al. Machine learning for fluid mechanics [J]. Computing Research Repository, 2020, 52(1): 477-508.
[3] Mo D, Lin Y Z, Liu Y X, et al. A review of recent advances in the application of machine learning algorithms for gas turbine combustion [J]. Propulsion and Energy, 2025, 1(20): 1-28.
[4] üstün E C, Freitas R D, Okafor E C, et al. Machine learning applications for predicting fuel ignition and flame properties: current status and future perspectives [J]. Energy and Fuels, 2025, 39: 13281-13314.
[5] Can O, Baklacioglu T, Ozturk E, et al. Artificial neural networks modeling of combustion parameters for a diesel engine fueled with biodiesel fuel [J]. Energy (Oxford, England), 2022, 247: 123473.
[6] Yang M, Zhou D Z. Ignition delay prediction for fuels with diverse molecular structures using transfer learning-based neural networks [J]. Energy and AI, 2025, 19: 100467.
[7] Zhu W, Zhou C, Zhang Y, et al. Review on the application of deep learning in reconstruction of turbulent Flame [J]. International Journal of Aeronautical and Space Sciences, 2026, 27: 568-580.
[8] Lu L, Jin P Z, Pang G F, et al. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators [J]. Nature Machine Intelligence, 2021, 3(3): 218- 229.
[9] Solji C, Ikhwan J, Haeun K, et al. Physics-informed deep learning for data-driven solutions of computational fluid dynamics [J]. Korean Journal of Chemical Engineering, 2022, 39(1): 515- 528.
[10] Cheng C, Zhagn G. Deep learning method based on physics informed neural network with resnet block for solving fluid flow problems [J]. Water, 2021, 13(4): 1-17.
[11] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need [C]//Conference and Workshop on Neural Information Processing Systems, 2017: 5998-6008.
[12] Bode M, Gauding M, Goeb D, et al. Applying physics-informed enhanced super-resolution generative adversarial networks to turbulent premixed combustion and engine-like flame kernel direct numerical simulation data [J]. Proceedings of the Combustion Institute, 2023, 39(4): 5289-5298.
[13] Imeraj G, Iyatomi H. Waveshift 2.0: an improved physics-driven data augmentation strategy in fine-grained image classification [J]. Electronics, 2025, 14(9): 1735.
[14] Kamalloo E, Rezagholizadeh M, Passban P, et al. Not far away, not so close: sample efficient nearest neighbour data augmentation via minimax [C]//Annual Meeting of the Association for Computational Linguistics, Stroudsburg, 2021: 3522-3533.
[15] Jacqueline B, Dominik H. Gaussian noise up-sampling is better suited than SMOTE and ADASYN for clinical decision making [J]. BioData Mining, 2021, 14(1): 1-11.
[16] Chen T Q, Guestrin C. XGBoost: a scalable tree boosting system [C]//ACM International Conference on Knowledge Discovery and Data Mining, 2016: 785-794.
[17] Christian S, Liu W, Jia Y Q, et al. Going deeper with convolutions [C]//IEEE Conference on Computer Vision and Pattern Recognition, 2015: 1-9.
[18] Liu Z, Mao H Z, Wu C Y, et al. A ConvNet for the 2020S [C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 11966-11976.
[19] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition [C]//International Conference on Learning Representations, 2015: 1-12.
[20] Huang G, Liu Z, Weinberger K, et al. Densely connected convolutional networks [C]//IEEE Computer Vision and Pattern Recognition, 2017: 2261-2269.
