Research On Positioning And Fault Detection Of Thermal Pipeline Supports

Thermal power generation is the main form of power generation in China,and water vapor is the intermediate medium for energy conversion in the process of power generation.Ensuring the safe and stable operation of the thermal pipeline is a necessary way to avoid safety accidents.The support components of the thermal pipeline play a role in fixing and supporting the pipeline,and the fault detection of the support components can alert the occurrence of pipeline bending and deformation faults.In view of the limitations of manual inspection,this article uses a robotic dog instead of manual inspection.Based on the video images captured by the ro botic dog as data,and based on a deep neural network,a two-stage algorithm of first positioning the support and then segmenting and detecting cracks is proposed.The specific work done in this article is as follows:(1)Using YOLOv5 as the basic algorithm for support positioning,combined with application background,the output branches of small target detection are removed,simplifying the network structure and reducing the amount of parameters to be optimized.Aiming at the defect of insufficient information concentration in the YOLOv5 backbone network,the CBAM attention mechanism was added to CSPDarket53 to improve the network’s information extraction ability.Deleting the classification loss part of the loss function is beneficial to improving the convergence speed of the model during the training process.Comparing the improved YOLOv5 with the YOLOv5 in the same dataset,the detection speed and accuracy are both improved.(2)For the support crack segmentation task,this article improves on the TransUnet network and proposes a CoTNet-TransUnet hybrid model.By using CoTNet to replace the CNN module in TransUnet,it overcomes the defect that TransUnet’s feature extraction network cannot pay too much attention to distinctive information.Comparative tests were conducted on CoTNet-TransUnet,TransUnet,Unet,and FCN on multiple datasets such as DeepCrack.The experimental results confirmed the advantages of the proposed method in the accuracy of crack segmentation.