In this paper, a deep neural network named Mask-MRNet is proposed to detect wind turbine (WT) blade fault based on images taken by unmanned aerial vehicles. Two datasets of the blade image are built for training and optimizing. Based on the proposed network, the blade images can intuitively express the mask, bounding box, and type of fault. In the detection, the network is stacked with Mask R-CNN-512 and MRNet. Optimized Mask R-CNN, Mask R-CNN-512, can significantly reduce inference time when performing large object detection such as WT blade fault. MRNet is proposed to correct the fault mask angle for cropping the low noise fault image from the original image and classify the fault type. Compared with more than 20 classification models based on indices including training and testing accuracy, the f1-score, and detection efficiency, DenseNet-121 was chosen as the classification model for Mask-MRNet. In addition, it is better to choose the classifier according to specific application demands in practical environments. A computational study was performed to further demonstrate that Mask-MRNet can not only achieve the multifunctional WT blade fault detection but also dynamic monitoring during the running of the WT.

中文翻译：

---

Mask-MRNet：用于风力涡轮机叶片故障检测的深度神经网络

在本文中，提出了一种名为 Mask-MRNet 的深度神经网络，用于基于无人机拍摄的图像检测风力涡轮机 (WT) 叶片故障。为训练和优化构建了刀片图像的两个数据集。基于提出的网络，叶片图像可以直观地表达掩码、边界框和故障类型。在检测中，网络堆叠了 Mask R-CNN-512 和 MRNet。优化的Mask R-CNN，Mask R-CNN-512，在执行WT叶片故障等大对象检测时可以显着减少推理时间。提出了MRNet来校正故障掩膜角度以从原始图像中裁剪低噪声故障图像并对故障类型进行分类。与20多种基于训练和测试准确率、f1-score、检测效率等指标的分类模型相比，DenseNet-121 被选为 Mask-MRNet 的分类模型。另外，最好根据实际环境中的具体应用需求来选择分类器。进行了计算研究，以进一步证明 Mask-MRNet 不仅可以实现多功能 WT 叶片故障检测，还可以实现 WT 运行期间的动态监控。