Open Access

ARTICLE

A Three-Layer Multi-Agent Framework for PHM-Enabling Autonomous Condition Monitoring of Power ICT Infrastructure in Underground Facilities

Jaekyung Lee^1,2, Byungsung Ko², Jiwon Lee², Jaeheon Park², Taewon Kim², Seoktae Kim², Wonhee Kim^3,*

1 Department of Energy Systems Engineering, Chung-Ang University, Seoul, Republic of Korea
2 KEPCO Research Institute, Daejeon, Republic of Korea
3 School of Energy Systems Engineering, Chung-Ang University, Seoul, Republic of Korea

* Corresponding Author: Wonhee Kim. Email:

(This article belongs to the Special Issue: AI-Enabled Prognostics and Health Management: Advanced Methodologies, Intelligent Systems, and Field Applications)

Computers, Materials & Continua 2026, 89(2), 31 https://doi.org/10.32604/cmc.2026.085203

Received 07 May 2026; Accepted 22 July 2026; Issue published 15 September 2026

Abstract

This study proposes the Artificial Intelligence-integrated Inspection Ecosystem (AIIE) as an autonomous condition monitoring platform to enable Prognostics and Health Management (PHM) for underground infrastructure facilities at the Korea Electric Power Corporation (KEPCO) power Information and Communication Technology (ICT) center. To address the environmental dependency of conventional systems, which necessitate extensive control logic redesigns upon changes in target facilities or environments, a three-layer abstraction architecture separating directive, orchestration, and execution roles is established, integrating a quadrupedal robot with heterogeneous sensors into a unified control structure. To overcome the limitation of relying on one general-purpose model for all inspection tasks, this study introduces a collaborative multi-agent system as its core component. This system is individually optimized for analog gauges, digital gauges, switch/LED states, and qualitative appearance defects. The proposed system achieves an Analog Gauge (AG) inspection error of 1.66% of full scale (%FS) (95% confidence interval (CI): [1.31, 2.11]) and a Digital Gauge (DG) recognition rate of 83.8% ($n=130$) using a small vision-language model (sVLM). A three-way validation score of ${\kappa }_w=0.724$ was obtained in identifying qualitative appearance defects, such as oil/water leaks, cleanliness, and insulation degradation. Field evaluations using 11,692 cases of real operational data confirm the system’s effectiveness, achieving an automatic inspection success rate of 87.6%. Downstream PHM prognostic functions, such as anomaly detection, degradation modeling, and remaining useful life (RUL) estimation, remain as future research directions.

---

Keywords

---

Supplementary Material

Supplementary Material File

---