2024-DTU Risø Wind Turbine Blade Inpsction Video Dataset Description

AI-based blade identification in operational wind turbines through similarity analysis aided drone inspection

Shohreh Sheiati a, *, Xiaodong Jia a, Malcolm McGugan a, Kim Branner a, Xiao Chen a, *
a Technical University of Denmark, Department of Wind and Energy Systems, Roskilde 4000, Denmark.
Abstract
Developing an automated method to monitor the structural health of each individual wind turbine blade over its lifespan is imperative for timely warning of potential damages. This study proposes an AI-based image recognition model to identify individual blades based on their unique surface features. To eliminate the influence of the image background on the identification task, a deep learning segmentation method was used to isolate the blade image from the complex background in optical images captured by drone. Subsequently, we utilized and repurposed similarity-based Siamese networks to establish a blade search capability to identify images of the same blade. This model automatically retrieves the corresponding blade images in response to a single query image. The results show that our trained model distinguishes similar images from non-similar images and identifies the blades from unseen images successfully. Our developed similarity-based identification method ensures continuous tracking and monitoring of the structural health of each individual blade over time.

Dataset description

We captured 29 videos from a wind turbine situated at DTU Risø campus, Roskilde municipality, Denmark. These videos were recorded under diverse circumstances, accounting for varying imaging conditions and different segments of the blades, using a DJI Mavic 2 drone.
We organized the dataset into a folder named “Dataset”, which comprises 29 subfolders, each dedicated to one of the original videos along with extracted image frames and their corresponding masks.
The initial set of 24 videos in the collection was used for training and validation processes. The remaining 5 videos, not utilized during the training process, were used for testing.
Masks were obtained from two techniques: bi-level thresholding and the GrabCut foreground-background segmentation method. The bi-level thresholding technique was applied to images with clear contrast between the foreground (blade) and background (all videos except 6, 13, 15, 25). For images with very complex backgrounds containing divers objects with different colours and textures (videos: 6, 13, 15, 25), the GrabCut method was employed.
Contact: [e-mail]

基于无人机巡检辅助相似性分析的运行中风力涡轮机叶片AI识别方法

Shohreh Sheiati a,*, Xiaodong Jia a, Malcolm McGugan a, Kim Branner a, Xiao Chen a,*
a 丹麦技术大学（Technical University of Denmark）风能与能源系统系，罗斯基勒4000，丹麦。

摘要
开发一种可在全生命周期内对每一片风力涡轮机叶片开展结构健康监测的自动化方法，对于及时预警潜在损伤至关重要。本研究提出一种基于人工智能（AI）的图像识别模型，可通过叶片独特的表面特征实现单叶片识别。为消除图像背景对识别任务的干扰，本研究采用深度学习分割方法，从无人机拍摄的光学图像中将叶片图像从复杂背景中分离出来。随后，本研究利用并改造了基于相似性的孪生神经网络（Siamese networks），构建叶片检索能力以识别同一片叶片的图像。该模型可通过单张查询图像自动检索到对应的叶片图像。实验结果表明，经训练的模型可有效区分相似与非相似图像，并成功从未见过的图像中识别出目标叶片。本研究提出的基于相似性的识别方法，可实现对每一片叶片结构健康状态的长期持续跟踪与监测。

数据集说明
本研究在丹麦罗斯基勒市丹麦技术大学（DTU）里斯奥校区的一台风力涡轮机上采集了29段视频。本研究使用大疆御2（DJI Mavic 2）无人机，在多样化场景下完成拍摄，涵盖不同成像条件以及叶片的不同分段。
本数据集以名为"Dataset"的文件夹进行组织，内含29个子文件夹，每个子文件夹对应一段原始视频，包含提取的图像帧及其对应的掩码（mask）文件。
数据集中的前24段视频用于模型的训练与验证流程，剩余5段未参与训练的视频则用于测试环节。
掩码文件通过两种技术生成：双阈值法（bi-level thresholding）与GrabCut前景背景分割方法。双阈值法适用于前景（叶片）与背景对比度清晰的图像，对应除编号6、13、15、25之外的所有视频所采集的图像。对于背景极为复杂、包含多种不同色彩与纹理对象的图像（对应编号6、13、15、25的视频），则采用GrabCut方法生成掩码。
联系方式：[e-mail]