YOLOv8n thermal solar-PV fault detector. Bootstrap model for ThermalVariations (Track B) until labelled in-house fault data is available.
Classes: hotspot, diode, misc (renamed from the source Hotspot/Diode/Anomaly to match the app taxonomy)
Training data: public "Solar thermal image defects" (Roboflow Universe), 1685 train / 100 val; DJI Mavic 3T thermal frames.
Val metrics (best.pt): mAP50 all 0.52 - hotspot 0.75, diode 0.45, misc 0.35.
Intended use: demo/bootstrap only. Single small site; under-fires on other palettes (e.g. white-hot). Not production-accurate.
App wiring: worker loads via hf_hub_download. Set TV_INFERENCE_BACKEND=yolo, TV_INFERENCE_MODEL_REPO=ThermalVariations/tv-thermal-yolov8n, TV_INFERENCE_MODEL_FILENAME=best.pt.
