摘要
To address the limitations of traditional image feature extraction methods in flame image classification—such as limited accuracy, low automation, and poor adaptability to non-stationan data—this study constructed a flame state recognition model integrating Deep Residual Network (ResNet) and Vision Transformer (ViT). Validation on a real combustor flame image dataset containing three states—ignition, stable combustion, and lean blow-out—demonstrated that the proposed hybrid model converged faster, and achieved a classification accuracy of 94.6%, representing an improvement of 9 percentage points over a classic ViT model. Furthermore, to tackle the challenge of significant intra-class variation, high inter-class similarity, and ambiguous transitional states in swirling spray flame state recognition, an adaptive focal loss optimization was applied to the standard loss function. This further increased the testing accuracy of the hybrid model to 99.5%, with improvements of 4.9 percentage points and 1.3 percentage points compared to the standard cross-entropy loss and the focal loss, respectively.
| 投稿的翻译标题 | Combustor flame state identification based on deep visual feature mining |
|---|---|
| 源语言 | 繁体中文 |
| 页(从-至) | 67-75 |
| 页数 | 9 |
| 期刊 | Tuijin Jishu/Journal of Propulsion Technology |
| 卷 | 47 |
| 期 | 6 |
| DOI | |
| 出版状态 | 已出版 - 1 6月 2026 |
关键词
- Combustor
- Computer vision
- Flame classification
- Loss function
- Vision Transformer
学术指纹
探究 '基于深度视觉特征挖掘的燃烧室火焰状态识别' 的科研主题。它们共同构成独一无二的学术指纹。引用此
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver