武宇平，刘海旭，吴劲芳，贾洪岩，刁嘉，朱董军

摘要：为了解决YOLOv3-Tiny对无人机采集的风机叶片图像损伤检测精度不高的问题，提出一种基于深度学习的风机叶片图像损伤检测方法。首先提出一种跨越式特征联合网络结构，由卷积层和拼接层构成，将不同深度的特征信息进行融合再学习，提取目标多层级特征信息;其次引入Inception模块结构，其中4个平行通道的多个卷积核对输入的特征图进行组合和压缩，在减少网络的学习参数的同时更好地表征图像特征信息，提高小目标的检测精度。实验表明，改进后算法的检测精度提高了2.69%，在自制的数据集中mAP可以达到88.58%，并且模型的参数缩小了4倍。因此，改进的方法比传统的YOLOv3-Tiny网络具有更好的检测效果。研究结果可为基于图像的损伤检测和风机叶片损伤智能识别提供参考。

关键词：计算机神经网络;缺陷检测;深度学习;风机叶片;YOLOv3-Tiny

中图分类号：TP391文献标识码：ADOI： 10.7535/hbgykj.2021yx05008

Application of improved YOLOv3-Tiny network

in fan blade damage detection

WU Yuping，LIU Haixu，WU Jinfang，JIA Hongyan，DIAO Jia，ZHU Dongjun

（Zhangjiakou Wind and Solar Power Energy Demonstration Station Company Limited.，State Grid Xin Yuan Company，Equipment Maintenance Center，Zhangjiakou，Hebei 075000，China）

Abstract：In order to solve the low precision of damage detection using YOLOv3-Tiny network for UAV collected fan blade images，a deep learning-based damage detection algorithm was proposed.Firstly，a leaping feature joint network structure was proposed，which was composed of a convolutional layer and a splicing layer.The feature information of different depths was fused and learned to extract the information of the target multi-level feature.Secondly，the inception module structure was introduced，and the input feature maps were compressed by multiple convolution kernels with 4 parallel channels，which reduced the learning parameters of the network and better characterized the image feature information.Then the detection accuracy of small targets was improved.Experiments show that the detection accuracy of the improved algorithm is increased by 2.69%，the mAP reaches 88.58% in the self-made dataset，and the parameters of the model are reduced by 4 times.The detect precision of this method is better than that of YOLOv3-Tiny network in the task of wind turbine blade damage detection.The results provide reference for image-based damage detection and intelligent identification work of fan blade damage.

Keywords：computer neural network;defect detection;deep learning;fan blade;YOLOv3-Tiny

隨着风能等清洁能源的发展，风力发电设备越来越普遍，但是由于风机设备所处的环境复杂、变化多样，以及风机自身的应力耦合作用，极易发生故障和损坏[1]。为了保障风机叶片的正常运行，提高经济效益，风机叶片缺陷检测的质量越来越受到重视。随着人工智能的发展，智能化检测更加快速、便捷。利用无人机采集风机叶片的影像信息，并对采集到的信息进行算法检测，可以快速有效地识别叶片的损伤部位，达到节省人力、物力、财力的效果。

要想精准地检测到风机叶片损伤部位，首先必须了解风机叶片的缺陷成因以及损伤种类。……