| 姜永祺,单慧琳.基于ECSMNet的风力发电机表面缺陷检测研究[J].电子测量与仪器学报,2025,39(5):166-176 |
| 基于ECSMNet的风力发电机表面缺陷检测研究 |
| Research on surface defect detection of wind turbine based on ECSMNet |
| |
| DOI： |
| 中文关键词: 风力发电机 深度学习 缺陷检测 注意力机制 残差网络 |
| 英文关键词:wind turbine deep learning defect detection attention mechanism residual network |
| 基金项目:国家自然科学基金（62071240，62106111）、江苏省研究生科研与实践创新计划项目（SJCX24_0448）、无锡市 “太湖之光”科技攻关(基础研究)项目（K20241047）资助 |
|
|
|
|
| 摘要点击次数: 908 |
| 全文下载次数: 929 |
| 中文摘要: |
| 针对风力发电机表面缺陷检测中存在的背景环境复杂、小目标缺陷检测精度不够等问题，提出了一种高效的风力发电机表面缺陷检测方法。首先构建具有特征提取与融合能力的主干网络，并在残差部分引入改进的通道注意力，帮助网络更好地提取特征信息，其次，使用新一代卷积变形模块进行输出，使模型能够更好地捕捉输入数据中空间和时间的相关性，简化模型的同时提高检测速度。最后在模型下采样部分引入高效空间-深度信息转换模块，将输入特征图中的空间维度降维至通道维度，保留显著特征的同时减少细粒度信息丢失，进一步提高模型检测小目标的能力。实验结果表明，相较于YOLOv7网络，所提网络在图像环境较为复杂的数据集1上准确率提升了3.5%，召回率提升了2.3%，交并比（IoU）为0.5时平均精度提升3.1%，在图片质量较好的数据集2上准确率达到96%，召回率达到94%，IoU为0.5时平均精度达到96.7%。所提模型在解决误检漏检问题方面有明显的优势，并且具有较快的检测速度，更适合在实际检测环境中应用，有良好的工程应用前景。 |
| 英文摘要: |
| Aiming at the problems of complex background environment and insufficient detection accuracy of small target defects in surface defect detection of wind turbines, an efficient surface defect detection method for wind turbines is proposed. Firstly, a backbone network with feature extraction and fusion capabilities is constructed, and an improved channel attention is introduced in the residual part to help the network better extract feature information. Secondly, a new generation of convolution deformation module is used for output, so that the model can better capture the correlation between space and time in the input data, simplify the model and improve the detection speed. Finally, an efficient spatial-depth information conversion module is introduced in the down-sampling part of the model to reduce the spatial dimension in the input feature map to the channel dimension, retain the salient features while reducing the loss of fine-grained information, and further improve the ability of the model to detect small targets.The experimental results show that compared with the YOLOv7 network, the accuracy of the proposed network is improved by 3.5%, the recall rate is improved by 2.3%, and the average accuracy is improved by 3.1% when the intersection over union is 0.5.In the data set 2 with better image quality, the accuracy rate reaches 96%, the recall rate reaches 94%, and the average accuracy reaches 96.7% when IoU is 0.5. The proposed model has obvious advantages in solving the problem of false detection and missed detection, and has faster detection speed. It is more suitable for application in the actual detection environment and has good engineering application prospects. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
