Research On Defect Detection Of Generator Set Impeller Based On CenterNet Network Model

Pump turbine is the core of a pumped storage power station.In order to ensure safe operation and give full play to the efficiency of pump turbine,it is necessary to do a good job in the maintenance of pump turbine.At present,the detection of impeller defects is mainly in the process of production and shutdown maintenance.According to the traditional way of maintenance,it is necessary to build a large detection experimental platform,which is not conducive to high-frequency detection.Based on this,this paper analyzes the common fault types of impeller,and puts forward an impeller defect detection and processing method based on CenterNet by combining computer vision and target detection technology.In this paper,the application of target detection algorithm to impeller defect recognition and location is proposed for the first time.Firstly,the basic defect types and common detection methods at home and abroad are introduced,and then an impeller image acquisition platform based on Jetson TX2 is built.603 impeller defect samples of a pumped storage power station are collected,and then the collected samples are compressed through adaptive cutting,bilinear interpolation The four methods of image denoising and edge detection preprocess the image and enhance the key features of the image,so as to achieve the purpose of sample expansion of the data set.Secondly,the principles of four target detection algorithms are introduced,in which the backbone extraction network in CenterNet algorithm is improved and optimized.After comparative analysis,the processed impeller defect image data are predicted through four target detection algorithms:CenterNet network,Faster R-CNN network,SSD target detection network and EfficientDet network,and the recognition accuracy of different algorithms for impeller defect image is obtained.Finally,the accuracy,recall,detection speed and other performance indicators are further compared,and the CenterNet target detection algorithm is selected as the optimal algorithm in this study.The average recognition accuracy is 91.18%,and the single operation time is 0.016s,The subsequent recognition test experiments using defect sample images instead of impeller entities also verify the effectiveness of CenterNet network.This paper presents an impeller defect detection algorithm based on CenterNet network model,which can detect and locate impeller defects efficiently.This method can quickly identify the defects and positions on the impeller surface,which not only provides a new idea for the accurate identification of impeller defects,but also enriches the application of target detection algorithm.Figure[53]Table[11]Reference[89]...