Leakage Detection and Identification of Power Plant Pipelines Based on Improved Faster RCNN

Abstract

For steam and oil leak detection problems in power plant pipelines, the technology and system implementation based on CBAM-Faster RCNN are proposed. First, VGG16 is replaced by the ResNet101 structure in order to obtain rich semantic information of feature graph, which is the feature extraction network of CBAM-Faster RCNN. A lightweight attention mechanism module CBAM is added to the feature extraction network and the full connection layer is replaced by one-dimensional convolution. These are able to solve the problem that the feature information of small leakage objects is tend to lose. Next, the new anchor scale in RPN network is obtained based on Kernel K-means clustering method, which is suitable for corresponding to the scale of the leakage object. Then, the ROI Pooling structure in CBAM-Faster RCNN is replaced by ROI Align to improve the accuracy of leakage detection algorithm. The mAP value of CBAM-Faster RCNN network reaches 96.5 percent, which proves accurate detection of steam and oil leakage is realized. Finally, the pipeline leakage detection system based on CBAM-Faster RCNN is embedded into the platform of intelligent inspection robot in key areas of power plant and the practicability of this method is proved.