Lightweight Multimodal Workflow for Automated Fouling Detection and Decision Making in PV Solar Thermal Power Plants

Photovoltaic solar thermal hybrid power plants face significant energy losses due to surface fouling, necessitating frequent visual inspections and maintenance interventions. While advanced visual language multimodal models have demonstrated robust capabilities for detailed visual analysis, their computational demands limit their direct applicability in resource-constrained field scenarios. To address this gap, we introduce SFFlow, a lightweight multimodal workflow designed for automated fouling detection, segmentation, and maintenance decision-making in solar installations. By strategically integrating a quantized visual language model for rapid initial fouling assessment with an optimized, specialized fouling segmentation network, our system achieves accurate, real-time fouling analysis on edge computing devices deployed on inspection robots. Comprehensive experimental evaluations demonstrate that SFFlow achieves competitive segmentation accuracy (57.55% IoU) compared to baselines, while significantly reducing computational cost (only 1.93 GFLOPs). Additionally, our multimodal decision-making strategy generates maintenance recommendations that align closely with expert-generated plans. This integrated, computationally efficient approach represents a significant advancement towards fully automated, intelligent fouling management workflows suitable for large-scale PV solar thermal power plants.