Chained Models for High-Res Aerial Solar Fault Detection скачать в хорошем качестве

Chained Models for High-Res Aerial Solar Fault Detection

In this episode, I caught up with Jonathan Lwowski, Connor Wallace, and Isaac Corley to explore how Zeitview built an AI-powered system to monitor solar farms at continental scale. We dive into the North American Solar Scan, which surveyed every 1MW plus site using high-resolution aerial RGB and thermal-infrared imagery, then processed it through a chained ML pipeline that detects panel-level defects and fire risks. The team discusses the challenges of normalizing data across regions, why a modular cascaded model design outperforms monolithic end-to-end approaches, and how human-in-the-loop review ensures high precision. They also share insights from building a generalized ML library on top of Timm, Segmentation Models PyTorch, and Torchvision to accelerate model training and deployment, their philosophy of prioritizing data quality over chasing SOTA, and how the same framework extends to wind, telecom, real estate, and other renewable assets. https://www.zeitview.com/   / jonathan-lwowski     / conorw8     / isaaccorley   🚀 TIMELINE 0:09 Team intros — Jonathan (AI lead), Conor (AI engineer), Isaac (ex-Zite View, co-developed models). 1:57 Zite View inspects renewable energy assets (solar, wind, utilities) using aerial, drone, and ground robots. AI detects anomalies and prioritizes maintenance. 3:19 Launched North American Solar Scan — flew all 1MW solar sites, collected RGB/IR, trained AI to detect panels & issues, assigned health scores, used for sales + ops. 5:11 Multi-stage pipeline — YOLO-based object detection + supporting models to reduce false positives. Designed for scale, accuracy, robustness. 7:00 Target anomalies: diode failures, hot cells, multiple hot cells, panel outages, string outages, misaligned panels. 8:41 Generalized ML repo enables rapid model trials via config files. Focus was speed to solution, not cost. 10:19 Mostly batch jobs → accuracy & ease mattered more than speed/cost. 10:49 Annotations began pixel-level → derived simpler labels later. Started small subset, scaled nationwide with human-in-loop. 12:04 Scaling revealed normalization challenges across climates/sites. Non-panel hotspots filtered out to improve accuracy. 15:43 Multi-model approach easier to generalize, modular, and reusable (e.g., panel outlines for UI). 21:20 Models map anomalies to power loss + detect fire risks using heuristics on temperature deltas. 24:12 Humans stay in loop, but AI boosts speed & quality, avoids linear analyst scaling. 26:06 No full human-only benchmark, but AI+human much faster than human-only. 27:52 Presented at CVPR workshop on multispectral & thermal CV. 28:30 Zite View publishes openly, differentiating on data scale/quality not just models. 29:20 Prioritizes robust, modular, production-ready ML over chasing SOTA. 32:04 Tests new SOTA (e.g., Florence 2 VLM), integrates into generalized repo if promising. 33:39 Repo + pipeline ready for new domains (e.g., wind farms) in weeks. 33:51 Updates shared mainly on LinkedIn. Jonathan bio: Jonathan Lwowski is an accomplished AI leader and Director of AI/ML at Zeitview, where he guides high-performing machine learning teams to deliver scalable, real-world solutions. With deep experience spanning start-ups and enterprise environments, Jonathan bridges cutting-edge innovation with business strategy, ensuring AI efforts are aligned, impactful, and clearly communicated. He’s passionate about unlocking AI’s potential while fostering a culture of technical excellence, collaboration, and growth. Conor bio: Conor Wallace is a Machine Learning Scientist at Zeitview, where he develops computer vision systems - including vision-language models - for geospatial AI applications in aerial inspection and infrastructure monitoring. His work integrates visual, thermal, and spatial data to build scalable systems for analysing assets such as solar farms, wind turbines, and commercial rooftops. He is also completing a Ph.D. in Electrical Engineering, where his research focuses on agent modelling in multi-agent systems, emphasising behaviour prediction in dynamic, non-stationary environments. Conor is passionate about applying state-of-the-art machine learning to real-world challenges in remote sensing and intelligent decision-making. Isaac bio: Isaac Corley is a Senior Machine Learning Engineer at Wherobots, where he builds scalable geospatial AI systems. He holds a Ph.D. in Electrical Engineering with a focus on computer vision for remote sensing. Isaac previously worked as a Senior ML Scientist at Zeitview and a Research Intern at Microsoft's AI for Good Lab. He is a core maintainer of TorchGeo and is passionate about advancing open-source tools that make geospatial AI more accessible and production-ready.