- The paper presents a multi-agent O&M assistant that integrates retrieval-augmented generation with schema-constrained SQL to provide explainable fault diagnosis in BESS.
- The approach uses complexity-aware routing and hybrid text-image retrieval, significantly improving routing accuracy from 20% to 70% and diagnostic quality.
- Experimental results highlight 100% safe SQL query plans and enhanced diagnostic report quality through coordinated multi-agent synthesis.
Traceable Retrieval-Augmented Multi-Agent Fault Diagnosis for Battery Energy Storage Systems
Problem Context and Motivation
Battery Energy Storage Systems (BESS) are critical infrastructures in modern power grids, supporting renewable integration, frequency regulation, and peak-shaving. However, scalable and reliable BESS operation is challenged by the diversity and complexity of faults that may emerge across subsystems (cells, modules, racks, BMS, PCS, EMS) as subtle, persistent, and often weakly correlated signals—such as cell-voltage inconsistency, resistance drift, and thermal anomalies. Current O&M workflows are fragmented: operators traverse disparate platforms to correlate alarms, analyze measurements, consult manuals, and ultimately rely on human expertise for actionable decisions. Conventional ML approaches may flag anomalies or estimate state-of-health but generally lack traceability, contextual reasoning, or auditable evidence chains to support high-stakes maintenance actions. LLM-based agents, in turn, are known to hallucinate when queried beyond their training distribution or domain knowledge.
System Design Overview
This work introduces a BESS-oriented O&M assistant that integrates retrieval-augmented generation (RAG), multi-agent reasoning, schema-constrained access to operational databases, and hybrid text-image document retrieval. The assistant aligns all pipeline steps—input routing, data access, evidence aggregation, and answer synthesis—with the regime of O&M in grid-scale BESS, emphasizing explainability, traceability, and auditability.
Central to the system is a multi-path routing module: simple tasks (e.g., time-series queries, alarm statistics) are dispatched to a single-agent fast path leveraging schema-constrained natural language SQL generation, hybrid retrieval, and lightweight visualization. In contrast, complex diagnosis (e.g., root-cause analysis spanning documents and operational logs) triggers a multi-agent workflow, wherein supervisory and researcher agents coordinate evidence acquisition, deduplication, ranking, and synthesis into structured diagnostic reports.
Methodology
Complexity-Aware Routing
Task routing is explicitly aligned with BESS business processes: alerting, troubleshooting, and time-series analysis. Input queries are decomposed into task-relevant components, mapped to corresponding logical tables (alarm events, diagnostics, measurement streams), and classified according to diagnostic complexity.
Schema-Constrained Database Access
LLM-predicted query plans are subjected to deterministic schema validation. Each plan (table, field, filter, time-window, ordering, limit) is checked against allowlists and route-table mappings; illegal fields, excessive limits, and unsupported queries are pruned or corrected. This enforcement prevents SQL injection, semantic collisions, and loss of traceability—a crucial requirement for safety-critical environments.
Hybrid Text-Image Retrieval-Augmented Generation
Text chunks from manuals, standards, maintenance cases, and operational reports are indexed alongside linked image metadata. Retrieval combines BM25 for exact-ID/term recall, and dense embeddings for semantic similarity:
s(d)=αBM25(d)+(1−α)Emb(d)
Dense retrieval and BM25 results are normalized and fused to rank relevant passages and image evidence. Retrieved materials are synthesized with structured findings and recommendations.
Multi-Agent Deep-Research Path
For non-trivial queries (ambiguous symptoms, multi-device root-cause diagnosis), the assistant invokes a supervisor/researcher multi-agent protocol. The supervisor decomposes subtasks, coordinates retrieval and initial answer drafting, and iteratively compresses and ranks supporting findings. The final output comprises the diagnostic conclusion, linked evidence, explicit risk judgments, and maintenance recommendations, all logged for audit and expert review.
Experimental Results
An internal evaluation leveraged anonymized BESS operational data and a domain-specific knowledge base (99 documents, 6,741 text chunks, 717 images). The test regime covered routing, database querying, and diagnostic inference.
- Routing: The system achieved 70.0% action accuracy (vs. 20.0% without routing), quality score 4.44 (vs. 3.20), and nearly halved average latency.
- Schema-Constrained SQL: 100% safe query-plan success (vs. 0% without schema validation), confirming that deterministic validation is essential for auditability and safety.
- Diagnosis: Multi-agent synthesis yielded answer quality 4.80 (vs. 3.60 for ablated single-agent inference), indicating a substantial gain in comprehensiveness and traceability.
Illustrative case studies demonstrated the assistant’s ability to translate unstructured or natural language O&M queries into schema-validated database operations, hybrid retrieval, visual analytics, and evidence-grounded maintenance recommendations.
Implications and Future Directions
This research demonstrates that reliable, explainable, and traceable O&M support for BESS can be achieved via modular integration of retrieval-augmented LLMs, deterministic schema validation, and multi-agent diagnostic coordination. The strong numerical results for routing accuracy, SQL plan safety, and diagnostic report quality underscore the practical viability of such domain-grounded, tool-augmented assistant architectures.
The implications for energy storage O&M are substantial: the approach offers a principled path to unifying multi-modal data, streamlining complex diagnostic workflows, reducing expert workload, and enabling audit-ready maintenance justifications—key enablers for safe and scalable grid-integration of storage assets. From a theoretical perspective, the work exemplifies best practices in aligning LLM outputs with domain schemas and tool traces, rigorously closing the gap between ML-based inference and actionable engineering decisions.
Future work includes scaling to additional BESS deployments, expanding the knowledge base and document/image corpus, integrating expert-labeled benchmarks for more rigorous external validation, and incorporating real-time or continuous-learning updates without compromising explainability or safety. Extensions to cross-site comparison, proactive anomaly detection, and autonomous O&M decision-making are promising next steps.
Conclusion
The traceable, retrieval-augmented multi-agent O&M assistant for BESS establishes a comprehensive, domain-specific approach to explainable fault diagnosis and maintenance support. By tightly coupling complexity-aware task routing, schema-constrained data access, hybrid RAG, and multi-agent synthesis, the system achieves high traceability, accuracy, and utility for large-scale storage operations. This methodology sets the stage for highly reliable, interpretable AI-driven maintenance and health management in critical energy infrastructure contexts (2607.01992).