TransAgent: Multimodal Transformer Defect Recognition with an LLM Agent, Specialized Models and Evidence Fusion

Transformer defect recognition involves complementary information from inspection images, dissolved gas analysis (DGA), and operation logs. This paper presents TransAgent, a collaborative architecture in which a ReAct language-model agent coordinates smaller task-specialized models: BERT interprets logs, a CNN analyzes gas measurements, and YOLO detects visible defects. Their outputs enter a common ten-class diagnostic frame and are combined through prediction-derived weighting and Dempster–Shafer evidence fusion; the agent coordinates tool use and report generation. On 100 constructed cases, with ten cases per class, the complete TransAgent configuration achieves 87% classification accuracy and a macro-F1 of 0.861. The corresponding accuracies are 73% for Qwen, 64% for ERNIE, and 50% for the text-input DeepSeek condition. Images are available to TransAgent, Qwen and ERNIE; the DeepSeek condition uses logs and gas measurements. ERNIE has 96% coverage and 66.67% accepted-case accuracy, while TransAgent classifies all 100 cases. The results demonstrate a practical way to combine language-model coordination with specialized diagnostic models in a multimodal workflow and provide a classwise assessment of the resulting predictions.

Authors

Institutions

Publication Details

Journal
Electronics
Published
2026-11-16
DOI
https://doi.org/10.3390/electronics15225313
Primary Topic
Power Transformer Diagnostics and Insulation
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

TransAgent: Multimodal Transformer Defect Recognition with an LLM Agent, Specialized Models and Evidence Fusion

Shanfeng Liu, Weijian Zhang, Joshua Luhwago, Shengzhe Xi et al.
1 links
Electronics
Power Transformer Diagnostics and Insulation
article

TransAgent: Multimodal Transformer Defect Recognition with an LLM Agent, Specialized Models and Evidence Fusion

Shanfeng Liu, Weijian Zhang, Joshua Luhwago, Shengzhe Xi, Asad Ullah, Li Ruan
article en

Abstract

Transformer defect recognition involves complementary information from inspection images, dissolved gas analysis (DGA), and operation logs. This paper presents TransAgent, a collaborative architecture in which a ReAct language-model agent coordinates smaller task-specialized models: BERT interprets logs, a CNN analyzes gas measurements, and YOLO detects visible defects. Their outputs enter a common ten-class diagnostic frame and are combined through prediction-derived weighting and Dempster–Shafer evidence fusion; the agent coordinates tool use and report generation. On 100 constructed cases, with ten cases per class, the complete TransAgent configuration achieves 87% classification accuracy and a macro-F1 of 0.861. The corresponding accuracies are 73% for Qwen, 64% for ERNIE, and 50% for the text-input DeepSeek condition. Images are available to TransAgent, Qwen and ERNIE; the DeepSeek condition uses logs and gas measurements. ERNIE has 96% coverage and 66.67% accepted-case accuracy, while TransAgent classifies all 100 cases. The results demonstrate a practical way to combine language-model coordination with specialized diagnostic models in a multimodal workflow and provide a classwise assessment of the resulting predictions.

ElectronicsVol. 15(22)
Electric Power Research Institute (US), State Grid Corporation of China (China) (CN), State Grid Henan Electric Power Company (China) (CN), Beihang University (CN)
State Grid Henan Electric Power Company
Power Transformer Diagnostics and Insulation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon