基于大数据迁移学习的灰岩地区排水孔淤堵自动识别技术

Automatic Identification Technique for Siltation and Blockage Conditions in Drainage Pipes in Limestone Areas Based on Big Data and Transfer Learning

隧道排水管淤堵或者失效,会危及边坡的稳定和公路的交通安全。目前缺乏对排水孔淤堵分类的研究。排水孔淤堵分类对排水管道的后期养护具有重大意义。为探索排水孔淤堵智能化检测方法,文章依托广连高速高峰隧道等粤北实体工程,研究了一种采用模型迁移的方法,将排水孔图像数据输入预训练卷积神经网络进行训练,以对新图像进行分类。在采集的排水孔图像数据集上进行试验,对比了三种不同网络模型对该数据集的准确率,结果表明,使用ResNet-18进行排水孔淤堵分类识别,准确率可达到93%,实现了对排水孔淤堵状态的有效分类,并且随着日后数据集的扩大,识别准确率将会有进一步的提高。

Siltation,blockage or failure of tunnel drainage pipes could endanger the stability of slopes and the traf⁃fic safety of the roads.At present,there is a lack of research on the classification of siltation and blockage conditions in drainage pipes.To explore the intelligent detection method of siltation and blockage conditions in drainage pipes,this paper studies a transfer learning based convolutional neural network classification algorithm for siltation and blockage targets in drainage pipes,taking the Gaofeng tunnel on Guangzhou-Lianzhou Expressway and other tunnel projects in northern Guangdong as the background.By using a model transfer method,the drainage pipe image data are input into a pre-trained convolutional neural network for training in order to classify the new images.The collect⁃ed image dataset of drainage pipes is tested to compare the identification accuracy of three different network models for the dataset.The results show that the classification and identification of siltation and blockage conditions in drainage pipes by using ResNet-18 could reach a 93%accuracy and achieve the effective classification of siltation and blockage conditions in drainage pipes.Also,the identification accuracy will be further improved with the expan⁃sion of the dataset in the future.