Solar Panel Inspection
Soiling, cracked modules and loose wiring in solar plants and rooftop arrays
The AI cuts out each photovoltaic panel from the image (drone, fixed camera or inspection photo) and checks for the non-conformities that hurt a plant's generation and safety the most: soiling (dust, bird droppings, leaves, snow), cracked glass or broken modules, and loose or exposed wiring. Each panel gets a status (Compliant, Soiling X%, Broken, Exposed wiring), generation loss is estimated and panels are ranked by maintenance priority — the O&M team goes straight to the modules that most need cleaning or replacement.
How it works
- 1Panel segmentation (YOLO11-seg): the outline of each module/table is cut out; in drone video each panel is tracked and counted only once.
- 2Defect segmentation (YOLO11-seg): dust, bird droppings, leaves and physical damage are cut out in the actual shape of the stain. Only what is INSIDE the panel counts — surrounding ground and roof are ignored.
- 3A second detector adds snow and confirms damage; CLIP (vision-language, no training) compares each panel with descriptions of clean, dirty, snow-covered and broken panels — damage is only confirmed with two votes.
- 4Soiling = % of the panel area covered by stains (dust, droppings, leaves, snow). For snow, coverage is measured by the panel's white pixels.
- 5Loose/exposed wiring: text-prompted detection (OWLv2) of hanging cables, loops and loose coils near the modules or on the structure.
- 6Estimated generation loss: proportional to the soiled area (see assumptions); a broken panel assumes the configured loss (default 33%, one of the three bypass diodes of a typical module).
- 7Maintenance ranking: broken panels and exposed wiring first (safety), then the panels with the highest estimated loss.
Benefits
- On-demand cleaning — only on panels/tables with relevant soiling, instead of a fixed schedule.
- Cracked modules identified early, before they become hot spots, water ingress or a fire risk.
- Loose wiring found before it causes arcing, abrasion wear or string downtime.
- Per-panel photographic evidence, with status and ranking, ready for the work order.
Generation loss from soiling is assumed proportional to the covered area (1% of area covered ≈ 1% loss) — a simplification: fine translucent dust loses less than that, while opaque stains over a cell (droppings, leaves) can trigger the bypass diode and lose more. A broken panel assumes the configured loss (default 33%). Exposed wiring is not included in the loss: it is a safety risk. Inspection is visual (RGB): internal microcracks and hot spots require electroluminescence or thermography (see Thermographic Analysis). Pre-trained models — "Solar Panel Inspection" (wcahca, CC BY 4.0, Roboflow data CC BY 4.0) and "solar-panel-od" (4keles, MIT); calibration with images from the plant itself is recommended before operational use.
