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

The Power Plant AX team at Narae Energy Services implemented automation for the existing Remote Vibration Monitoring System (RVMS) using Amazon Bedrock AgentCore and Strands Agents. While the existing RVMS only performed alarm detection and relied on experts for root cause analysis, this implementation resolved the issue with an architecture that separates judgment procedures from data access.

Architecture and Security Configuration

The system consists of Agent Skills (Markdown-based definitions of expert judgment procedures) and read-only MCP tools (data access). Through a model-driven execution loop, it observes tool results to determine the next action, and halts answer generation if evidence is insufficient. On the security front, VPC endpoints, internal ALBs, Direct Connect, or Site-to-Site VPNs were used to securely connect to AWS services within a closed network.

Separation of Search Tracks and Performance Validation

The initial single vector index approach resulted in only 0 out of 21 evaluation questions succeeding due to similarity errors in identifier data such as KKS tags. To resolve this, four search tracks were designed as follows:

• Track A: Normalization and exact match lookup for KKS tags/alarm names (including SQL aggregation)
• Track B: Vector search and full manual text search for startup sequences/concepts
• Track C: Graph adjacency traversal for signal lineage/drawing connections
• Track D: Vector search and metadata filtering for general product platform operations

Evaluation Results and Implications

In tests based on a 21-item evaluation set, the search layer achieved 21/21 success, and the response layer achieved 18/21 success. The three failures were response quality issues such as formatting or missing evidence references. Notably, all three trap questions, including those with non-existent tags or ambiguous queries, passed, proving the effectiveness of the design in suppressing ungrounded answer generation. This case demonstrates that vector search alone is insufficient for identifier-centric data, and a design combining exact matches with relationship queries is essential.