Most drone inspections still require a human to manually inspect the video for defects. Computer vision can help automate and accelerate this inspection process.
Overview
Exelon is leveraging synthetic data generation using NVIDIA Omniverse to automate drone inspections of grid infrastructure. This approach aims to enhance the efficiency of defect detection in grid assets, ultimately improving reliability and reducing manual labor.
What You'll Learn
1
How to utilize synthetic data generation for training computer vision models
2
Why using segmentation masks improves defect detection accuracy
3
How to implement a scalable ecosystem for asset inspection using drones
Prerequisites & Requirements
- Understanding of computer vision concepts and AI model training
- Familiarity with NVIDIA Omniverse and Omniverse Replicator(optional)
Key Questions Answered
How is Exelon using synthetic data for drone inspections?
Exelon is using synthetic data generation in NVIDIA Omniverse to create thousands of labeled, photorealistic examples of grid asset defects. This synthetic data helps train computer vision models for automated defect detection during drone inspections, improving grid maintenance and reliability.
What challenges did Exelon face in training their AI model?
Exelon faced the challenge of needing a large pool of labeled real-world defect data for training and testing their AI model. They explored synthetic data generated in Omniverse Replicator as a solution to address this data scarcity.
What are the benefits of automating drone inspections?
Automating drone inspections reduces exposure to in-field hazards for crews, minimizes manual labor for image review, and accelerates the timeline from image capture to defect resolution, thereby enhancing grid reliability and resiliency.
How does Exelon plan to scale their inspection process?
Exelon aims to create an end-to-end scalable ecosystem that can be applied to various transmission and distribution assets, starting with cross-arms and potentially expanding to poles, transformers, and other grid components.
Key Statistics & Figures
Customer base served by Exelon
more than 10M customers
Exelon is the largest regulated electric utility in the United States, serving customers across multiple states.
Emission reduction goals
50% reduction by 2030 and net-zero emissions by 2050
These goals are part of Exelon's Path to Clean initiative aimed at enhancing sustainability.
Technologies & Tools
Platform
Nvidia Omniverse
Used for synthetic data generation and creating photorealistic environments for training AI models.
Tool
Omniverse Replicator
A core extension of Omniverse for generating physically accurate synthetic data for training computer vision models.
Software
Autodesk Maya
Used by Deloitte to build 3D models of assets and defects.
Software
Epic's Unreal Engine
Utilized for developing photorealistic environments with accurate lighting and scene conditions.
Key Actionable Insights
1Implementing synthetic data generation can significantly enhance the training of AI models, especially in fields with limited labeled data.This approach allows organizations to create diverse datasets that improve model accuracy and performance, making it particularly useful in industries like utilities where real-world data may be scarce.
2Using segmentation masks for defect detection can lead to improved identification of specific defects in images.Segmentation masks allow for precise pixel-level identification, which is crucial for detecting fine details like cracks or joins in grid assets, ultimately leading to better maintenance outcomes.
3Collaborating with third-party vendors can expedite the development of complex 3D models and synthetic data.By leveraging external expertise, organizations can focus on their core competencies while still achieving high-quality results in areas like 3D modeling and photorealistic environment creation.
Common Pitfalls
1
Relying solely on real-world data for training AI models can lead to insufficient datasets and poor model performance.
This is often due to the challenges in collecting labeled data for every possible defect, which can hinder the effectiveness of AI solutions.
2
Underestimating the complexity of 3D modeling can result in delays and increased costs in projects.
Organizations may not have the in-house expertise required for high-quality 3D modeling, necessitating partnerships with specialized vendors.
Related Concepts
Synthetic Data Generation
Computer Vision In Utilities
AI Model Training Techniques
3d Modeling And Simulation