Aiming at the scenario of high-speed photographic data analysis and automatic diagnosis in ground combustion test benches, this study presents a neural network based on multi-channel variational Auto-Encoder (MVAE) fo...
详细信息
Aiming at the scenario of high-speed photographic data analysis and automatic diagnosis in ground combustion test benches, this study presents a neural network based on multi-channel variational Auto-Encoder (MVAE) for clustering and identifying temporal patterns in flame morphological properties. While variationalautoencoder architectures have been previously developed for clustering tasks, their effectiveness in capturing the complex, nonlinear features of high-speed flame visualization data has not been thoroughly explored. The method was applied to several datasets obtained from high-speed photography of a scramjet combustor. The network extracts temporal sequence features through shared convolutional and hidden layers, utilizing reconstruction residuals as clustering criteria while optimizing feature extraction and clustering simultaneously. When applied to high-speed photography datasets from a strut-based scramjet combustor under varying equivalence ratios, the MVAE significantly outperforms traditional clustering methods, reducing clustering discrepancies by approximately 96% compared to K-means approaches. Application of this algorithm to two experimental cases with different equivalence ratios reveals distinct flame organization modes, with clear correlations to specific frequency characteristics in time-frequency spectrograms. The demonstrated capability of this method to automatically identify complex flame patterns without potential anthropogenic bias offers researchers a powerful tool for data-driven analysis of combustion dynamics and flame stabilization mechanisms in hypersonic propulsion systems.
暂无评论