风能是一种绿色无污染的可再生能源，在建立绿色低碳循环发展的经济体系和清洁低碳安全高效的能源体系中扮演着重要的角色。随着近些年风电行业的大力发展，风力发电机组的安全稳定运行受到了重点关注，叶片作为风力机最重要的部件之一，是故障的多发区和易发区。风力机叶片复杂严酷的工作环境容易导致表面产生裂纹，表皮脱落等缺陷，如不及时处理维修，容易引发风力机结构性破坏，造成巨大的经济损失，因此风力机叶片表面缺陷的检测极为重要。针对风力机叶片运行维护和检修中的困难，本文分析总结现有叶片故障检测方法的特点，结合计算机视觉和深度学习技术的快速发展和应用，提出了基于深度学习的风力机叶片表面缺陷检测方法，对叶片多种缺陷实现了精确地定位和识别，本文的研究内容如下：基于传统数字图像处理技术设计算法流程实现了缺陷的定位。首先建立图像数据集，对叶片表面缺陷图片进行灰度化处理，减少计算量。之后通过滤波去除图像噪声，使用图像分割算法提取缺陷区域。经形态学处理后，提取缺陷边框参数并在原图中标注，实现对缺陷部位的定位。通过研究表明，该框架可以定位三种特征明显的缺陷，但是鲁棒性不足，面对复杂背景无法精准定位也无法进行缺陷的分类。基于卷积神经网络对风力机叶片表面缺陷实现了分类。首先构建图像数据集，对多种卷积神经网络模型进行简单介绍，使用迁移学习的方法建立模型，并对比各模型的精度。为降低模型参数量，提高模型精度，采用NIN的方式构建了一个新的卷积网络结构。研究结果表明，新模型的参数量远小于常见模型结构，在精确率，召回率和F1值等指标中表现良好，实现了对风力机叶片表面缺陷的分类需求。基于YOLOv5算法对风力机叶片表面缺陷同时实现了定位与分类。首先对SSD，Faster-RCNN和YOLO目标检测算法进行分析，对YOLOv5模型进行原理介绍，对数据量较小的数据集使用数据增强方法。设置训练参数之后使用YOLOv5算法进行训练，针对研究结果中裂纹检测效果差的特点，进一步扩充了裂纹的数据量，并对比了多个目标检测模型的检测结果。研究表明，扩充数据量的方法有效提高了模型精度。使用特征图可视化和Grad-CAM的方法提高了卷积神经网络的可解释性。为了提高模型精度，同时减少计算开销，提出了MHAP-YOLO目标检测模型并于数据集上训练，研究表明，本模型在一定程度上提高了精度，减少了计算复杂度。
Wind energy as a green and non-polluting renewable energy source, plays an important role in establishing a green, low-carbon, circular development economic system and a clean, low-carbon, safe and efficient energy system. With the development of the wind industry in recent years, the safe and stable operation of wind turbines has caused extensive concern. The blade, as one of the most important structures of wind power, is a fault-prone area. In the complex and harsh working environment, wind turbine blades are prone to cracks and skin debonding. If not timely processing and repair, it is liable to the structural damage and resulting in huge economic losses, so the detection of wind turbine blade surface defects is extremely important.In view of the difficulties in wind turbine blade operation and maintenance, this paper analyzes and summarizes the characteristics of existing blade fault detection methods, combines the rapid development and application of computer vision technology, and proposes a computer vision-based wind turbine blade surface defect detection method, which achieves accurate positioning and identification of multiple defects on the blade. The research for this paper is as follows:The dataset used in this study is firstly established and the algorithm flow is designed based on traditional digital image processing techniques. The blade surface defect images are grayed out to reduce the computational effort. After that, the image noise is reduced by filtering and the defect area is extracted using image segmentation algorithm. The parameters of the defect border are extracted by morphological processing and labeled in the original image to realize the localization of the defect area. Through experiments, it is proved that the framework can locate the defects with three obvious features, but the robustness is not enough, and it cannot locate precisely nor classify the defects in the face of complex backgrounds.To classify the surface defects of wind turbine blades, a convolutional neural network approach is used. Firstly, we construct the image dataset and briefly introduce the convolutional neural network models such as ResNet, VGG Net, DesNet and MobileNet. The models are built using the transfer learning method and the accuracy of each model is compared. In order to reduce the number of model parameters and improve the model accuracy, a new convolutional network structure is constructed by NIN. Experiments show that the number of parameters of the new model is much smaller than that of the common model structure, and it performs very well in the accuracy, recall and F1 score, and basically achieves the demand for classification of wind turbine blade surface defects.In this section, it is decided to use YOLOv5 algorithm for defect detection on wind turbine blade surface. A brief analysis of the SSD, Faster-RCNN and YOLO target detection algorithms and an introduction to the principles of the YOLOv5 model are presented. Because of the insufficient date, the data enhancement method was used. After setting the experimental parameters, experiments were conducted using the YOLOv5 algorithm, and the amount of crack data was further expanded to address the poor crack detection in the experimental results. The detection results of several target detection models are compared, and the experiments show that the method of expanding the amount of data effectively improves the model accuracy. The interpretability of the convolutional neural network was improved by using the feature map visualization and Grad-CAM methods. To improve the model accuracy and reduce the computational cost at the same time, the MHAP-YOLO target detection model is proposed and trained on the dataset. Experiments show that this model improves the accuracy and reduces the computational complexity.
