PowerVLM:基于Federated Learning与模型剪枝的电力视觉语言大模型OA

The rapid evolution of smart grids has produced massive volumes of multimodal,heterogeneous power-system data,posing new challenges for AI models in complex electric-field perception.Meanwhile,the sensitivity of industry data and stringent privacy-preservation requirements further restrict the cross-scenario transferability of general-purpose models in the power domain.To address these issues,we propose a federated-learning and model-pruning framework for a power-domain vision-language large model.Specifically,we introduce PowerVLM,a class-guided vision-language model that incorporates a novel class-guided enhancement module to strengthen its comprehension and question-answering capabilities on power-related image-text pairs.A reinforcement-learning-driven federated-training strategy is adopted to mitigate domain gaps while strictly preserving data privacy.Finally,an information-resolution-based pruning algorithm is designed to enable efficient fine-tuning with significantly reduced trainable parameters.Extensive experiments on three representative power scenarios—substation inspection,transmission-line inspection,and operation safety supervision—demonstrate that our method achieves superior performance on all key metrics(METEOR,BLEU,and CIDEr)in multimodal power-domain question-answering tasks,offering a new technical paradigm and practical support for intelligent perception in power systems.