Abstract: To reliably assess mechanical threats to transmission lines in complex outdoor environments remains a critical bottleneck urgently requiring a breakthrough in current smart inspection within the power sector． Although existing deep learning-based monitoring systems have achieved preliminary binary detection regarding the“presence or absence”of mechanical intrusions, field operators have a pressing demand for fine-grained and highly interpretable quantitative methods to assess“threat severity”． To address this demand, in the present work, a framework is proposed for mechanical intrusion threat assessment based on collaboration of small and large models． Firstly, a small object detection model is employed to locate accurately power facilities and mechanical equipment． Subsequently, a vision-language model (VLM) is utilized to infer threat levels ranging from A to D． During the large model inference stage, a specialized cognitive reasoning chain is designed to deconstruct the scene into four interpretable dimensions: spatial analysis, power facility localization, target positioning, and mechanical feature recognition． This will enable VLM to fully integrate visual information with semantic context for reasoning． Experimental results from an expert-annotated dataset demonstrate that the proposed method achieves an accuracy of over 90% in the threat target recognition task and over 80% in the quantitative threat level assessment task, to effectively verify its effectiveness and engineering practicality．