Real-time power system intrusion detection via feature-based edge intelligence and Yolov7 segmentation

Foreign objects intrusion poses a major threat to reliable operation of the power system, and detecting these intrusions quickly and accurately is of paramount importance to prevent confusion. Xinan Wang from IEEE, Di Shi and Fengyu Wang are tackling this challenge with a new framework for real-time detection and intrusion tracking. Their system combines a robust tracking algorithm, combining a fast object localization model with advanced image processing techniques, including a good feature extraction method for distinguishing objects. The resulting pipeline not only achieves high accuracy in a variety of scenarios, but also runs efficiently on low-cost edge computing hardware, paving the way for scalable and practical deployments on critical infrastructures.

Detecting aviation threats using deep learning

This study details a system for detecting foreign objects near high-voltage power lines, improving grid safety and reliability. Using computer vision and deep learning techniques, scientists automatically identify potential hazards from aerial images, such as those captured by drones, and actively reduce the risk of short circuits, outages, and safety incidents. This work covers problem definition, proposed solutions, and implementations and is based on existing work in object detection, computer vision, and power grid safety. The system employs a deep learning-based object detection system to experiment with models such as Yolov7, Yolov8, Mask R-CNN, ConvNext, and more to optimize performance.

The team improved the loss function, increased model accuracy using data augmentation techniques, focused on efficient models with real-time or near-real-time detection, automated inspections and improved safety. The system could significantly improve the reliability of the power grid by proactively identifying and dealing with potential threats, increasing safety for workers and the public, and reducing inspection costs. It also allows predictive maintenance, predicts potential failures, and actively schedules maintenance.

Real-time intrusion detection using deep feature encoding

Scientists have developed a three-stage framework for real-time detection of foreign objects intrusion in power transmission systems, overcoming the limitations of existing methods of low-cost hardware. The system uses the Yolov7 segmentation model to generate a highly discernible embedding representing the unique properties of each detected object using a Convnext-based functional encoder trained with triplet loss. This innovative approach allows the system to effectively generalize without costly retraining, even if you encounter invisible objects before. The triplet loss function optimizes embedding to maximize object type separation, minimizes variance within each type, and improves the robustness of difficult conditions. To ensure reliable tracking of multiple objects under motion and occlusion, researchers implemented functionally assisted IOU-based trackers and leveraged the embedding of robust features to associate objects across consecutive frames. The entire pipeline is optimized to demonstrate real-time inference with lower-precision weights, such as the Jetson Orin Nano, for practical scalability for deployments to low-cost edge hardware, and for large field deployments.

Real-time external object detection in power systems

Scientists have developed a three-stage framework for real-time detection and tracking of foreign objects that invade power transmission systems. The system uses the Yolov7 segmentation model for rapid object localization, then generates 1×1024 feature embeddings using the finely tuned combonex model, encodes important visual information for robust identification between objects. This approach effectively captures discriminatory visual information, even when there are significant variations within the object class. The core of the system resides in feature similarity-based matching mechanisms, allowing rapid updates to the object database simply by adding new embeddings without computationally expensive model retraining. This allows the system to adapt to previously invisible objects, such as wind-blown debris and construction materials. The entire pipeline is designed to operate on low-cost edge hardware such as the Jetson Orin Nano, providing practical scalability for large-scale deployments and significantly improving the reliability and safety of the power transmission system.

Real-time intrusion detection using deep Feature Tracking

This study presents a three-stage framework designed for real-time detection and tracking of foreign object intrusions in power transmission systems. The system integrates a Yolov7 segmentation model for rapid object localization, a combonex-based feature extractor that utilizes the loss of triplets to create different object representations, and a functional support Iou tracker to maintain accurate tracking even when objects are partially obscure or moving. A key achievement is the system's ability to incorporate new object types without the need for full model retraining, allowing scalable deployment in dynamic environments. Extensive testing using real-world surveillance and drone footage demonstrates the high accuracy and robustness of the framework across a variety of challenging scenarios. The team successfully optimized the pipeline for deployment to low-cost edge hardware to ensure the practicality of their applications. Future research will expand the system's capabilities and include fire detection, automated camera control, and large field trials, ultimately aiming for fully autonomous monitoring to increase the safety and reliability of the power infrastructure.