Abstract:UAV aerial images of insulators have problems such as complex backgrounds, small targets, and obscured targets, resulting in low detection accuracy of insulator targets. In order to solve the above problems, a HRD-YOLOX (Hybrid Attention Mechanisms, Regularization, and Depth Separable Convolution for YOLOX) algorithm suitable for recognizing insulators and detecting defects is proposed based on the YOLOX network. First, the hybrid attention module HAM-CSP (Communication Sequential Processes for Hybrid Attention Mechanisms) is incorporated into the feature extraction network to suppress the interference from the complex background of insulator pictures. Second, the PANet (Path Aggregation Network) structure is replaced with the R-BiFPN(Regularization-Bidirectional Feature Pyramid Network) structure for feature fusion, and an improved regularization module is introduced into the convolutional layer of the BiFPN structure to reduce the overfitting problem of the model and the inference speed of the network. Then, in order to solve the computationally intensive and overfitting phenomenon generated by the detection of small targets such as insulator discharge traces, the DHConv (Deep Separable and Hybrid Attention Mechanisms Convolution) convolution is proposed to be applied to the Conv convolutional layer in order to reduce the number of parameters and improve the generalization ability of the model. Finally, the binary cross-entropy function for confidence loss is changed to the Focal Loss function to solve the problem of missed detection and positive and negative sample imbalance due to the image overlap problem The experimental results show that the HRD-YOLOX algorithm improves the average accuracy of insulator detection from 86.23% to 91.34%, and at the same time improves the detection speed from 70.6FPS/S to 82.23FPS/S, which allows real-time detection of insulators, and improves the performance of complex background targets as well as small targets detection.

Experimental identification of Ion Cyclotron Emission on HL-2A using YOLO neural network algorithm

Recent Experiment Results During Electron Cyclotron Heating On Hl-2a

OV / 4-5 1 Overview of Experimental Results on the HL-2 A Tokamak

Identifying L-H transition in HL-2A through deep learning

Identification of core ion cyclotron instabilities on HL-2A tokamak

Development of the Ion Cyclotron Emission Diagnostic on the HL-2A Tokamak

Real-time equilibrium reconstruction by neural network based on HL-3 tokamak

Ion Cyclotron Emission Diagnostic System on the Experimental Advanced Superconducting Tokamak and First Detection of Energetic-Particle-driven Radiation

Real-time equilibrium reconstruction by multi-task learning neural network based on HL-3 tokamak

Optical Design and Synthetic Analysis of the Electron Cyclotron Emission Imaging Diagnostic of HL-2M Tokamak

Determining 1D fast-ion velocity distribution functions from ion cyclotron emission data using deep neural networks

HRD-YOLOX Based Insulator Identification and Defect Detection Method for Transmission Lines

Plasma Real-time Diagnostics Based on Visual Image Recognition

Quantifying fusion born ion populations in magnetically confined plasmas using ion cyclotron emission

Investigation of the long-lived saturated internal mode and its control on the HL-2A tokamak

Application of machine learning for detecting and tracking turbulent structures in plasma fusion devices using ultra fast imaging

A neural network for the analysis of Langmuir-probe characteristics

Neural network identification of the weakly coherent mode in I-mode discharge on EAST

Machine learning aided line intensity ratio method for helium-hydrogen mixed recombining plasmas

Classification of the L-, H-mode, and plasma-free state: Convolutional neural networks and variational autoencoders on the edge reflectometer for KSTAR

Using convolutional neural networks to detect edge localized modes in DIII-D from Doppler backscattering measurements