Narae Energy Services Implements Automated Power Plant Equipment Diagnostics Using Strands Agents and Bedrock AgentCore

The Power Plant AX Team at Narae Energy Services completed a PoC to solve power plant equipment diagnostic issues using Strands Agents and Amazon Bedrock AgentCore. The existing Remote Vibration Monitoring System (RVMS) could detect alarms but had a high dependency on experts, who had to reference a list of 5,996 alarms, 64 manuals, and 15,614 logic drawings to determine the cause.

Architecture and Core Solution

As a solution, they designed an Operations Support Agent that separates the decision-making process from data access. Agent Skills define the expert's diagnostic sequence (Trend→FFT→Orbit, etc.), and data is accessed via read-only MCP (Model Context Protocol) tools. A model-driven execution loop was applied to stop answer generation when evidence is insufficient, and write tools and external web searches were blocked to protect source systems. In closed network environments, the AWS VPC and internal network were connected via VPC endpoints and Direct Connect.

Search Track Separation and Performance Improvement

The initial single vector index approach resulted in only 0 successes out of 21 evaluation questions due to similarity errors with identifier data such as KKS tags. To resolve this, the search was separated into 4 tracks based on data characteristics.

• Track A (Exact Match): Alarms and KKS tags are processed via exact match lookup after normalization
• Track B (Hybrid): Startup procedures, etc., use a combination of vector search and full-text manual search
• Track C (Graph): Signal and drawing relationships utilize graph adjacency
• Track D (General): Product information applies vector search and metadata filtering

With this track separation strategy, the search layer achieved a 21/21 (100%) success rate, and the response layer recorded 18/21. All 3 trap questions were passed, demonstrating the system's ability to prevent hallucinations.