【作者】 李浩；

【导师】 许文波；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2019， 硕士

【摘要】 航空发动机叶片是航空发动机中重要的组成部分,它的正常运转可以为发动机提供持续不断的飞行动力。它工作于高温高压的环境,并且服役时间往往较长,这样的环境容易使其产生疲劳裂纹,这些发动机叶片上的裂纹对航空发动机的正常运行构成了潜在的威胁。事实上,只要发动机叶片上存在裂纹,无论其大小如何,都会危及人员并对机器构成严重的威胁,重则机毁人亡,造成不可挽回的损失。因此现在迫切需要有一种智能化和高效化的方式来对航空发动机叶片裂纹进行检测,这具有重要的安全意义。论文主要工作如下:(1)将R-FCN(Object Detection via Region-based Fully Convolutional Networks)算法应用到航空发动机叶片裂纹的检测任务,之后利用R-FCN算法的位置敏感得分图和感兴趣区域池化操作来确保此网络可以准确的区分出待检测目标的类别以及确定其位置。实验的改进部分主要是结合航空发动机叶片裂纹自身的细而长特点,对R-FCN算法中的区域推荐网络(Region Proposal Network,RPN)中的anchor部分进行了改进,提高了检测精度。同时对所有的航空发动机叶片的测试集图像添加噪声后进行检测,由实验结果分析可知R-FCN算法对含有噪声的叶片裂纹图像具有较好的鲁棒性。(2)将YOLOv3(You Only Look Once)算法应用于航空发动机叶片裂纹的目标识别任务。YOLO系列的算法是基于one-stage方法的典型代表算法,其检测速度非常快。本文借鉴FPN(Feature Pyramid Networks)特征金字塔的多尺度、多层级检测结构,对初始的YOLOv3中的特征金字塔结构做出改进,构建了一个特征尺度更丰富,层级更多的金字塔结构,将高层次高语义的特征图与低层次高分辨率的特征图进行充分融合,得到四个不同尺度的预测层,有助于获取更多的关于小型目标物体的特征和位置信息。通过实验对比分析,改进后的算法相比于改进前具有更好的检测精度,可达到实时检测。同时对所有的航空发动机叶片的测试集图像添加噪声后进行检测,由实验结果分析可知YOLO算法对含有噪声的叶片裂纹图像的鲁棒性较差。更多还原

【Abstract】 Aero-engine blades are important components of aero-engine and their normal operation can provide continuous flight power for the engine.They work in high temperature and high pressure environment and the working time is often very long.Such a poor environment tends to cause fatigue cracks in the engine blades,and these cracks on these engine blades will pose a potential threat to the normal operation of the aircraft engine.In fact,as long as there are cracks on the engine blades,regardless of their size,they will endanger the personnel and pose a serious threat to the machine and even cause irreparable damage.Therefore,there is an urgent need for an intelligent and efficient way to detect aero-engine blade cracks,which has important safety significance.The main work of the thesis is as follows:(1)The R-FCN(Object Detection via Region-based Fully Convolutional Networks)algorithm is applied to the detection task of aero-engine blade cracks,and then the position-sensitive score maps and the position-sensitive ROI pooling of R-FCN algorithm is used to ensure that this neural network can distinguish the category of the object to be detected and determine its location accurately.The improvement of this algorithm mainly considers the fine and long characteristics of the aero-engine blade crack itself,and this thesis improves the anchor part of the Region Proposal Network(RPN)in the R-FCN algorithm to improve detection accuracy.In addition,after adding noise to the image test sets of the aero-engine blades,the experimental results show that the improved R-FCN algorithm is robust to blade crack images containing noises.(2)The YOLOv3(You Only Look Once)algorithm is applied to the detection task of aero-engine blade cracks.The YOLO series of algorithms is a typical representative algorithm based on the one-stage method,and its detection speed is very fast.This thesis draws on the multi-scale and multi-level detection structure of the FPN(Feature Pyramid Networks)and improves this pyramid structure in the initial YOLOv3.It constructs a pyramid structure with more feature scales and more levels,and then it fully merges high-level high-semantic feature maps with low-level high-resolution feature maps.Finally,the fused feature map is used as the prediction layer to predict the object and achieve multi-scale prediction to obtain more information about the characteristics and location of small objects.Through experimental comparison and analysis,the improved algorithm has better detection accuracy than before,and it can achieve real-time detection.In addition,after adding noise to the image test sets of the aero-engine blades,the experimental results show that the improved YOLO algorithm is less robust to blade crack images containing noises.更多还原