| Insulators,as one of the essential devices in the power transmission system,play an indispensable role in ensuring the stable operation of the power system.Therefore,conducting insulator inspection is of great significance for the stable operation of the power system.Traditional insulator inspection models mostly rely on CPU and GPU,require image data exchange through networks,and have poor real-time performance and need to run on a PC platform.This paper takes the transmission line as the research background,focuses on the research status and problems of insulator inspection,and designs a deep learning object detection network model on the ARM + FPGA heterogeneous platform ZYNQ for insulator image inspection.Firstly,based on the analysis of insulator self-explosion defect features,an insulator image dataset is created.Then,experiments are conducted on the insulator image dataset using the YOLOv3(You Only Look Once version 3)and SSD(Single Shot Multi Box Detector)object detection networks.Comparing the detection results,the SSD object detection network model performs better than YOLOv3 in detecting small-scale insulator defects.Therefore,this project is based on the SSD object detection network as the basic network model.Secondly,the computational complexity of the feature extraction network of the SSD object detection model is high,and it has certain requirements for device memory size.Thus,the Mobile Netv3 is used to improve the feature extraction network of the original SSD object detection model.To improve the detection ability of small target defects in the insulator dataset,the activation function and attention module of the original network model are optimized.The improved SSD object detection network model is trained on the insulator dataset,achieving an m AP(mean Average Precision)value of 92.47%,better than the original SSD object detection network model.Finally,based on the improved SSD object detection model,the network model is deployed on the ZYNQ hardware platform.This article analyzes the ZYNQ hardware platform technology,parallelism of the improved SSD object detection model,collaboration between the PS(Processing System)and PL(Programmable Logic)ends,and storage of feature maps and weight parameters,providing a basis for implementing the improved SSD object detection network on the ZYNQ hardware platform.By designing the software and hardware of the ZYNQ platform,combining the PS and PL ends,and quantizing the network parameters of the improved SSD object detection model,the improved SSD object detection model is deployed on the ZYNQ hardware platform to achieve insulator image inspection.By means of collaborative debugging of software and hardware,and designing comparative experiments,this paper compares the implementation of insulator image inspection based on the ZYNQ platform with the implementation based on the PC platform.The final experimental results show that the ZYNQ platform has an average detection time of only 334 ms and a power consumption of only 2.642 W.Through actual testing,the inspection system has successfully achieved the function of insulator image inspection and possesses the advantage of low power consumption. |
