Accurate and Efficient Insulator Maintenance: A DETR Algorithm for Drone Imagery

Authors: Yanfeng Tian, Rodina Binti Ahmad, Nor Aniza Binti Abdullah
Affiliation: Faculty of Computer Science and Information Technology, Universiti Malaya, Kuala Lumpur, Malaysia
Published: February 25, 2025
DOI: 10.1371/journal.pone.0318225

Abstract

As electricity demand rises, ensuring the safety and stability of power grids becomes increasingly critical, necessitating effective maintenance and inspection strategies. Insulators serve as essential protective devices on outdoor high-altitude conductors within power grids. However, drone-based inspections often yield suboptimal image quality due to adverse weather conditions such as rain, snow, and fog, as well as challenges related to sunlight, rapid movement, and long-distance imaging.

To tackle these issues, this manuscript presents an insulator defect detection algorithm designed specifically for low-quality images captured by drones. By utilizing a patch diffusion model, we generate high-quality images that enhance the precision of insulator defect detection methods. Additionally, we introduce an optimized DETR (Detection Transformer) method that integrates a Spatial Information Interaction Module to better capture the characteristics of minor defects. A specialized convergence network further boosts the detection capabilities of the DETR.

Experimental results reveal that our proposed insulator detection technique achieves an impressive detection accuracy of 95.8%, significantly outpacing existing defect detection methods in complex environments. This advancement addresses the limitations of traditional methods through sophisticated computational models, ultimately leading to more efficient, cost-effective, and secure power grid maintenance and inspections.

Introduction

The growing demand for electricity drives the need for efficient and safe energy transmission systems, making power grid maintenance and inspection vital. Insulators are integral components of these systems, serving to support and secure conductors while ensuring electrical insulation. Their failure can lead to catastrophic consequences for power grids, highlighting the necessity for timely and accurate detection of insulator faults.

Insulators are often exposed to harsh environments, which can lead to various types of damage. These include the accumulation of dirt and salts in dusty conditions that, when moist, can compromise their insulating performance. Severe weather conditions—such as hail, storms, and extreme temperatures—can physically damage insulators or diminish their insulating capabilities. Furthermore, materials like porcelain or glass can degrade over time, reducing their mechanical strength and insulation effectiveness. Thus, developing efficient methods for rapid and accurate insulator condition detection is crucial to maintaining stable power grid operations.

Traditional inspection methods, such as helicopters, climbing robots, and manual checks, are often costly and inefficient. With advancements in drone technology, drones have emerged as the preferred tool for power line inspections. However, drones face significant challenges due to the low-quality images they capture, which affect detection efficiency.

Several factors contribute to this issue. Drones must operate from a distance, which introduces background noise into their raw data. Air currents can cause vibrations, resulting in blurry images. Adverse weather conditions can also distort aerial images. As a result, a large volume of low-quality images reduces the overall efficiency of drone inspections. Therefore, improving image quality through noise reduction is key to enhancing detection efficiency.

Given the complex structure and widespread distribution of power grid components, insulator detection via drone imagery can significantly enhance efficiency, even in challenging weather conditions. To address this, we propose a dual-part solution. First, we employ a patch diffusion model to generate high-quality images that mitigate the impact of adverse weather conditions. Second, we utilize an optimized DETR model to improve small target identification accuracy, enabling all-weather inspections and ensuring the stability of high-voltage transmission lines.

Related Works

The methodology proposed in this study for detecting insulator faults in low-quality images is divided into two main parts: image denoising and object detection models.

Image Denoising

Image denoising algorithms are widely used to reduce random noise in images. However, traditional denoising methods can blur fine details, which is detrimental to detecting small defects. While image enhancement techniques can improve overall image quality, they typically do not focus on noise reduction.

Deep learning models, particularly Generative Adversarial Networks (GANs), have shown promise in image denoising but often require extensive computational resources. Under challenging weather conditions, images from drones may contain various noise types. Therefore, rapid and accurate detection of insulator defects necessitates effective image denoising techniques.

Recent developments in diffusion models have gained attention for their ability to generate high-quality images through iterative optimization. Conditional diffusion models can incorporate additional information during the generation process, allowing for more controlled and effective image restoration.

Object Detection Models

Convolutional Neural Networks (CNNs) have demonstrated significant potential in remote sensing image object detection. Despite their effectiveness, detecting small objects remains a challenge, particularly when they occupy only a few pixels in an image.

DETR, based on the Transformer architecture, offers a novel approach by transforming object detection into a direct set prediction problem. However, DETR struggles with small object detection and requires extensive training to achieve optimal performance.

To enhance small object detection, researchers are exploring various strategies, including model architecture improvements and training optimizations. Our study focuses on improving the DETR architecture and introducing an optimized model for better small object detection performance.

Solution for Low-Quality UAV Image Detection

Overview

The IDD-DETR model proposed in this study addresses the challenge of identifying insulator defects in low-quality images. The model is structured to generate high-quality images through a patch diffusion model and enhance detection accuracy with a Spatial Information Interaction Module.

Patch-Based Diffusion Image Restoration

Diffusion models have gained traction for their ability to progressively restore images. Conditional diffusion models allow for more precise control over image generation by incorporating degraded images into the restoration process.

The patch-based diffusion model focuses on decomposing images into smaller patches, allowing for better local feature capture and improved restoration quality. Each patch is processed independently, enhancing the clarity of the overall image while maintaining detail.

Insulator Defect Detection Transformer

The IDD-DETR model employs a Spatial Information Interaction Module (SIIM) to enhance the interaction of spatial information during encoding. This module captures pixel relationships, improving model performance in image detection tasks.

A Feature Convergence Module (FCM) integrates global dependencies into CNN features, enhancing the model’s representation and detection capabilities.

Experiment

Data Preparation

The study utilized several publicly available datasets, including the Chinese Power Line Insulator Dataset (CPLID), Unifying Public Datasets for Insulator (UPID), and Synthetic Foggy Insulator Dataset (SFID). These datasets provide diverse images of insulators, including variations under different environmental conditions.

Training Settings

The training environment was built using a high-performance computing node, and the model was trained over 50,000 iterations using an Adam optimizer. The training process focused on the denoising capabilities of the model, with evaluations conducted at regular intervals.

Comparison with Other Methods

To assess the effectiveness of the IDD-DETR model, comparisons were made with other models, including Weather Diffusion and DDRM. The IDD-DETR model demonstrated superior performance across various metrics, especially in detecting small defects.

Conclusion

This study addresses the challenges of insulator fault detection under adverse weather conditions, particularly in foggy environments. The proposed IDD-DETR model effectively combines an optimized DETR with a patch diffusion model, achieving significant improvements in both image restoration and defect detection.

Future research will focus on enhancing model capabilities, expanding datasets, and exploring broader applications in fields requiring small object detection. This comprehensive approach aims to refine the IDD-DETR model to predict and prevent insulator failures, thereby improving the reliability of power transmission systems.