Application Of Deep Learning In Detection Of Surface Defects Of Submarine Cables

In recent years,with the rapid development of submarine cables,submarine cable construction has been recognized by governments all over the world as a complex and arduous major engineering project,during which,the quality of submarine cables is of the top priority.For example,during production,submarine cable defects not detected results in long-term external corrosion of submarine cable,which will cause certain material failure problems,even endanger life safety.Therefore,how to detect the surface defects of submarine cables quickly and accurately lays a foundation for the quality and material maintenance of cables,which has become a very important research topic.Under the above background,it is a vital task for cable testing equipment to detect surface defects and diagnose them with high accuracy and efficiency.Nowadays,the detection method of submarine cable is mainly manual identification,which is more likely to make mistakes.With the continuous development of industrialization,the improvement of detection technology is more urgent.In view of the above problems,this thesis studied a defect detection program on the submarine cable surface based on deep learning,which integrates the five main situations of submarine cable surface defects.The program can basically accurately and quickly realize the classification and defect location of submarine cable surface defects,thus provide scientific and technological equipment support for construction personnel.The main research contents of this thesis are as follows:(1)Image pre-processing.Most of the submarine cable pictures collected in factory are with problems like incorrect orientation,blurred pixels,shades of color,unclear outline and so on.Halcon software is what we used here to pre-process all the collected pictures,and relevant formulas are selected accordingly for different problems to process the images,so as to ensure the stability and reliability of the detection process.(2)Submarine cable surface defect classification and feature extraction.In this thesis,the resnet50 model in Imagenet was used for migration learning.The pre-training classification model would initialize the parameters of the pre-convolution network layer and train the parameters of the RPN network,so as to realize the function of extracting images of the model.(3)Submarine cable defects identification and improvement.FPN pyramid model was used to identify small target defects.Through the analysis of submarine cable fault characteristics,this thesis proposed a method for fast RCNN model improvement,introducing the focal loss loss function for data imbalance.In addition,replacing the original ROI pooling with ROI align can reduce the quantitative error and improve the positioning accuracy of small target defects.(4)Improved fast RCNN comparison with other network models.In this thesis,SSD and yolov3 were selected as comparing factor.Although the improved accuracy rate was up to 98.01%,and the accuracy rate of all cable classification exceeded SSD and yolov3,the operation speed was instead reduced by 45 ms,which indicates that the speed needs further improving.