摘要
针对发电站制冷管射线图像的焊缝区域对比度较低、特征不明显,传统方法难以实现精确搜索的问题,提出一种基于深度学习的发电站制冷水管焊缝区域搜索方法。利用限制对比度的自适应直方图均衡化限制图像统计直方图的幅度,抑制噪声放大,得到直方图的累积分布函数,以校正图像的低对比度;利用深度神经网络的24个卷积层提取输入图像的特征、2个全连接层预测图像位置和类别概率,实现水冷壁管焊缝区域的检测,以克服传统模板匹配精度低、时间复杂度高的问题。对100张制冷管射线图片按4∶1∶5分为训练集、验证集和测试集,利用训练集和验证集对深度神经网络进行训练,将图像送进训练好的模型,预测制冷管焊缝区域的位置。试验结果表明,基于深度学习的焊缝区域搜索方法可以实现焊缝的精确搜索,准确率达到96%,搜索效率及准确度高。
Aiming at the problem that the weld area of the power station cooling pipe ray image has low contrast,the features are not obvious,and the traditional method is difficult to achieve an accurate search,a method for searching the power station cooling water pipe weld seam area based on deep learning is proposed. Contrast-limited adaptive histogram equalization( CLAHE) is used to limit the amplitude of the statistical histogram of the image and suppress the amplification of noise to obtain the cumulative distribution function( CDF) of the histogram to correct the low contrast of the image. The 24 convolutional layers of the deep neural network are used to extract the features of the input image,and the 2 fully connected layers predict the image position and class probability to achieve the detection of the welded seam area of the water-cooled wall tube and overcome the problems of the traditional template with low accuracy and high time complexity. The 100 refrigeration tube ray pictures were divided into training set,validation set and test set according to 4 ∶1 ∶5. The training set and validation set are used to train the deep neural network,and the position of the welding pipe area of the refrigeration pipe is predicted using the trained model. The experimental results show that the method of searching seam area based on deep learning can realize the precise search of the weld seam with an accuracy rate of 96% and high search efficiency and accuracy.
作者
王立辉
秦成帅
杨贤彪
沈秋成
WANG Lihui;QIN Chengshuai;YANG Xianbiao;SHEN Qiucheng(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China;Jiangsu Frontier Electric Technology Co.,Ltd.,Nanjing 211102,China)
出处
《电力工程技术》
2020年第5期191-196,共6页
Electric Power Engineering Technology
基金
国家自然科学基金资助项目(61773113)。
