Traceable Fault Diagnosis for Battery Energy Storage Systems via Retrieval-Augmented Multi-Agent O&M Assistant¶

🕒 Published (v1): 2026-07-02 10:26 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper presents a multi-agent retrieval-augmented assistant for battery energy storage system (BESS) operations and maintenance (O&M), designed to deliver traceable fault diagnoses from natural-language queries. It combines schema-constrained SQL generation, hybrid text-image RAG, and supervisor-researcher agent coordination to turn fragmented O&M workflows into auditable diagnostic reports. Preliminary evaluation shows that schema validation is necessary for safe database access (100% vs. 0% without it) and that multi-agent synthesis meaningfully improves diagnostic quality.

Problem¶

Current BESS O&M workflows are fragmented: operators must manually cross-reference alarm platforms, time-series databases, maintenance manuals, and expert judgment to diagnose faults such as voltage inconsistency, resistance drift, internal short-circuit risk, or thermal abnormality. Data-driven models output scores or labels but not traceable maintenance decisions, and general LLMs are unsafe without domain knowledge grounding, controlled database access, and auditable evidence chains.

Method¶

The system routes incoming O&M requests via a complexity-aware router into either a single-agent path (for simple alarm, trend, or retrieval tasks) or a multi-agent deep-research path (for cross-document root-cause analysis). Key components:

Schema-constrained database access: The LLM generates a structured query plan (table, fields, filters, time range, ordering, limit); a deterministic sanitizer checks allowlists, normalizes aliases, clamps limits, and rejects invalid fields before building executable SQL.
Hybrid text-image RAG: Offline, manuals/procedures/cases are chunked and indexed; online retrieval fuses BM25 and dense embeddings as \(s(d) = \alpha \cdot \text{BM25}(d) + (1-\alpha) \cdot \text{Emb}(d)\) after normalization, with image metadata linked to text chunks for visual evidence.
Multi-agent synthesis: For complex queries, a supervisor decomposes the request, dispatches researcher agents, compresses findings, and assembles a structured diagnostic report (conclusion, evidence, risk judgment, recommended action) logged for audit.

Key Contributions¶

BESS-specific task routing that maps queries to three business routes (alerting, troubleshooting, time-series analysis) with complexity-aware dispatch.
Schema-constrained natural-language-to-SQL pipeline with deterministic sanitization, achieving 100% safe SQL-ready plan success versus 0% without schema validation.
Hybrid text-image RAG pipeline combining BM25 and dense retrieval with linked image evidence for maintenance procedures and failure-analysis figures.
Supervisor-researcher multi-agent evidence synthesis producing traceable, auditable diagnostic reports rather than bare labels or scores.

Results¶

Routing: Full system achieves 70.0% action accuracy (quality 4.44, 40.01 s) vs. 20.0% accuracy (quality 3.20, 87.45 s) without routing.
Database access: 100% safe SQL-ready plan success with schema validation vs. 0% without (raw LLM plans produce invalid fields, route-table mismatches, or unsupported scopes).
Diagnostic quality: Full multi-agent system scores 4.80 quality vs. 3.60 without multi-agent reasoning (95.33 s vs. 60.99 s).
Qualitative cases cover ISC-risk retrieval, cluster telemetry query/visualization, knowledge-base retrieval with document and image evidence, and recurrent alarm analysis.

Note: evaluation is preliminary internal only; full field statistics and expert-labeling protocol are deferred to the final paper.

Limitations¶

Evaluation is internal and preliminary; no public benchmark, no hold-out test set with published expert labels, and station scale/sampling details are withheld pending clearance.
Quality scores are on an unspecified scale with no inter-rater agreement reported; methodology for ground-truth construction is not described.
The hybrid retrieval fusion weight \(\alpha\) is not ablated or tuned in reported experiments.
Multi-agent path adds significant latency (95 s vs. 61 s for ablated variant); real-time O&M use cases may be latency-sensitive.
System is evaluated only on a private internal knowledge base (99 documents, 7 tables); generalizability to other BESS vendors or chemistries is unvalidated.

Relevance to Agentic AI / LLM Agents¶

This work is a concrete domain deployment of multi-agent RAG architecture in a safety-critical industrial setting, demonstrating how complexity-aware routing, tool-constrained agents, and evidence-bundle synthesis can replace fragmented expert workflows. The schema-constrained SQL generation pattern—where a deterministic sanitizer enforces correctness rather than trusting raw LLM output—is a practical guardrail strategy directly applicable to any agentic system with database tool access. The supervisor-researcher decomposition follows established ReAct/multi-agent paradigms but grounds them in a domain-specific task ontology (alarm, troubleshooting, trend), illustrating how domain structure improves both routing accuracy and latency. For researchers tracking agentic AI, the paper provides a worked example of traceability and auditability requirements in high-stakes agentic deployments.