【摘要】 液体火箭发动机作为航天运载器的核心动力装置，其地面测试中出现的异常火焰是结构性失效甚至灾难性事故的关键早期征兆。此类故障发展迅速且破坏性大，所以准确、迅速识别故障火焰非常重要。为此提出了一种基于优化YOLOv11n的火焰识别算法。首先，在C3k2模块中引入可变形卷积DCNv4，并添加到YOLOv11n骨干网络中，增强模型对复杂几何形状和尺度变化的感知；其次，引入DySample上采样替代邻近插值上采样，减少上采样过程中的特征信息丢失，从而提升模型对小目标的识别能力；最后，将CIoU Loss替换为Focal-EIoU损失函数，提高收敛速度和回归精度。实验结果表明，优化后算法的检测效果有了明显提升，平均检测精度达到了91.8%，较基准模型YOLOv11n提升2.4百分点，在参数量仅增加25%的代价下，实现了检测精度和模型复杂度的平衡。更多还原

【Abstract】 Abnormal flames observed during ground tests of liquid rocket engines that is core propulsion systems for aerospace launch vehicles suggests critical early indicators of structural failures or catastrophic accidents. Given the rapid progression and severe destructiveness of such faults, accurate and prompt flame detection is essential. To address this issue, an optimized flame detection algorithm based on YOLOv11n is developed. Firstly, DCNv4 is integrated into the C3k2 module within the YOLOv11n backbone network, enhancing the model’s perception of complex geometries and large-scale variations. Secondly, DySample upsampling is introduced to replace nearest-neighbor interpolation to mitigate feature information loss during resolution recovery, improving small-target detection. Thirdly, the CIoU loss function is substituted with Focal-EIoU to accelerate convergence and refine localization precision. Experimental results demonstrate that the optimized algorithm achieves significant improvements, and mean average precision reaches 91. 8% and improves a 2. 4 percentage point compared with the baseline YOLOv11n model. While maintaining a 25% parameter overhead, this method achieves a balance between detection accuracy and computational complexity.更多还原