基于SW-YOLO模型的航空发动机叶片损伤实时检测

Real time detection of aircraft engine blade damage based on SW-YOLO model

关键词:  航空发动机  叶片损伤  深度学习  孔探检测  目标检测  实时检测

Key words:  Aircraft engine  Blade damage  Deep learning  Borescope detection  Object detection  Real-time detection

摘要:

孔探检测技术是航空发动机叶片损伤检测的主要手段，但目前依赖人工操作，耗时耗力。本文提出了一个孔探视频检测的SW-YOLO模型，该模型包括输入端、主干网络、颈部网络、头部网络4个模块。首先，在主干网络加入了空间通道注意力模块（Spatial Channel-Convolutional Block Attention Module，SC-CBAM），有效避免位置信息丢失，提高目标边界回归能力，相较于YOLOv5，其平均精度均值PˉA@0.5提高了5.4%。其次，在颈部网络对特征金字塔网络（Feature Pyramid Network，FPN）进行了改进，通过融合低层特征，扩大了模型感受野，有利于较小损伤区域的检测，如烧蚀损伤，平均精度提高了8.1%。最后，通过与YOLOv5，Faster R-CNN，SSD模型的对比实验，结果表明SW-YOLO模型的平均精度均值分别提高了7%，6.2%，6.3%，检测速度满足实时检测需求，有利于提高航空发动机孔探检测的自动化和智能化水平。

Abstract:

Borescope detection technology is one of the main means for detecting damage of aero-engine blades， but currently it mainly relies on manual operation and is time-consuming and labor-intensive. This paper proposes a SW-YOLO model for aero-engine blade damage borescope video detection. The model includes 4 modules： input terminal， backbone network， neck network and head network. Firstly， by adding a space channel attention module Spatial Channel-Convolutional Block Attention Module （SC-CBAM） to the backbone network to alleviate the loss of location information and improve the ability of target boundary regression， and its average accuracy PˉA@0.5 increases by 5.4% compared with YOLOv5. Secondly， the structure of Feature Pyramid Network （FPN） is improved in the neck network， and the low-level features are fused to expand the receptive field of the model， which has a better detection effect for the smaller damage area， such as ablation， and the average accuracy is improved by 8.1%. At last， compared with YOLOv5， Faster R-CNN and SSD models， the experimental results show that the average precision mean of the SW-YOLO model has been improved about 7%， 6.2%， 6.3%， respectively， and the detection speed meets the real-time detection requirements， which is conducive to improving the automation and intelligence level of aero-engine blade damage borescope detection.