DefectCoT: Evaluation and Analysis of Interpretability Reasoning Capability for Multimodal Large Language Models in Power Equipment Defect Detection

Abstract

Accurate detection of power equipment defects is a critical link in ensuring the safe operation of power grids. With the rapid advancement of Multimodal Large Language Models (MLLMs), general-purpose Vision Language Models (VLMs) have demonstrated strong capabilities in natural image understanding. Nevertheless, in the identification of power equipment defects, existing models frequently lack interpretable reasoning for subtle anomalies, and thus struggle to meet the stringent requirements of intelligent inspection for highly reliable defect classification. To this end, we construct DefectCoT, a Chain-of-Thought (CoT) reasoning dataset tailored for detection of power equipment defects. This dataset covers 9 typical defect categories. Through a systematic construction pipeline, each sample not only provides fine-grained descriptions of anomalous features but also includes three candidate defect categories, a reasoning chain for stepwise evidence matching and elimination, as well as the final defect category and the corresponding defect coordinate localization. Based on the DefectCoT dataset, we conduct evaluations on several mainstream VLMs, with a focus on assessing their capabilities in interpretable defect reasoning and localization. Experimental results demonstrate that the reasoning performance of open-source models is generally low under direct evaluation; after supervised fine-tuning, both the defect classification accuracy and localization precision of all open-source models are significantly improved. The experimental results demonstrate that InternVL3.5-14B achieves the best performance in the reasoning of the defect categories, with an accuracy of 87.20%, representing a 28.65% improvement over the baseline model. The DefectCoT dataset proposed in this paper provides a benchmark for evaluating fine-grained, interpretable visual reasoning on power equipment defects, facilitating the reliable deployment of multimodal large language models in intelligent inspection scenarios.