🤖 AI

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

This paper presents a traceable fault-diagnosis assistant for large-scale battery energy storage systems that leverages retrieval-augmented multi-agent reasoning to integrate operational data, domain knowledge, and visual evidence for explainable O&M decision-making.

Original authors: Jiangdi Ru, Bing Li, Yage Huang, Ding Wang, Keru Hua

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiangdi Ru, Bing Li, Yage Huang, Ding Wang, Keru Hua

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive battery energy storage system (BESS) as a giant, complex orchestra of thousands of individual batteries working together to power the grid. When something goes wrong—like a violin string snapping or a drum going out of tune—the system sounds an alarm. But here's the problem: current systems are like a frantic stage manager who just shouts, "Something is wrong!" without telling the musicians what is wrong, where it is, or how to fix it.

This paper introduces a new "Smart O&M Assistant" designed to be the ultimate conductor and detective for these battery orchestras. Here is how it works, broken down into simple concepts:

1. The Problem: Too Many Clues, No Detective

Right now, if a battery acts up, human operators have to play "connect the dots" across different rooms. They check one screen for alarms, another screen for graphs, a filing cabinet for manuals, and then call an expert to guess the solution. It's like trying to solve a mystery by looking at a map, a weather report, and a phone book separately, hoping they all make sense together.

2. The Solution: A Team of Specialized Detectives

The authors built a digital assistant that acts like a team of specialized detectives working together, rather than one confused generalist.

The Traffic Cop (The Router): When a human asks a question (e.g., "Why is the temperature rising?"), a smart "traffic cop" first figures out what kind of detective is needed. Is this a simple check? A deep investigation? Or a request for a chart?
The Librarian (Retrieval): If the team needs to know how to fix a specific part, the "Librarian" doesn't just guess. It dives into a massive library of manuals, past repair reports, and diagrams. It uses a special search that looks at both words and pictures (like finding a photo of a burnt wire in a manual) to get the right evidence.
The Accountant (Database Access): If the team needs to check the battery's vital signs (voltage, temperature), they can't just ask the database "anything." The system uses a strict "schema-constrained" method. Think of this as a translator who ensures the detective asks the database in a language it understands perfectly, preventing the AI from making up fake numbers or asking for data that doesn't exist.
The Chief Investigator (Multi-Agent Reasoning): For tough cases where the clues are scattered, a "Chief Investigator" coordinates the team. It breaks the big mystery into small tasks, sends researchers to find clues, and then stitches all the findings together into a single, traceable report.

3. Why It's "Traceable"

The most important feature is that this assistant doesn't just give an answer; it shows its homework. If it says, "This battery has a short-circuit risk," it points to the exact graph, the specific alarm log, and the page in the manual that led to that conclusion. It's like a student showing their work on a math test, proving they didn't just guess the answer.

4. What the Paper Actually Found

The authors tested this system internally with real (but anonymized) battery data and a private library of maintenance documents.

Accuracy: When they added the "strict translator" (schema validation) for the database, the system went from failing completely (0% success) to succeeding perfectly (100%) in generating safe, usable database queries.
Speed & Quality: The "traffic cop" (routing) made the system much faster and more accurate at picking the right tools.
Deep Thinking: When the system used the "team of detectives" (multi-agent) for complex problems, the quality of the final diagnosis jumped significantly compared to a single AI trying to do it alone.

The Bottom Line

This paper presents a prototype for a smart assistant that helps battery maintenance workers. It connects the dots between raw data, historical records, and expert manuals to provide clear, evidence-backed reasons for why a battery is acting up and what to do about it. The authors emphasize that this is a preliminary internal test, and while the early results are promising, they plan to share more detailed field statistics and expert reviews in a future, full paper.

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

Problem Statement

Large-scale Battery Energy Storage Systems (BESSs) require Operation and Maintenance (O&M) decisions that synthesize diverse data sources, including alarms, cell-level measurements, device topology, diagnostic tables, historical cases, and maintenance documents. While current monitoring platforms can flag threshold violations, they frequently fail to provide traceable explanations for complex issues such as voltage inconsistency, resistance drift, short-circuit risks, capacity divergence, or thermal abnormalities.

Existing workflows are fragmented; operators must manually inspect alarms in one platform, retrieve time-series curves from databases, search external manuals, and consult experts for final judgments. Furthermore, while data-driven models can detect anomalies or estimate State of Health (SOH), they typically output scores or labels rather than traceable maintenance decisions. General Large Language Models (LLMs) are deemed unsafe for this domain without private knowledge grounding, controlled database access, conservative risk reasoning, and auditable evidence.

Methodology

The proposed system is a traceable BESS fault-diagnosis assistant built on a retrieval-augmented multi-agent reasoning framework. It integrates operational data, domain knowledge, visual evidence, and report generation.

System Architecture and Task Model

The assistant organizes user requests into three specific BESS business routes:

Alerting: Analyzing alarm events and abnormality tables.
Troubleshooting: Investigating observed symptoms or weak signals using diagnostic tables and RAG evidence.
Station/Device Time-Series Analysis: Examining equipment measurements over time.

A complexity-aware router directs tasks:

Simple Tasks: Sent to a single-agent supervisor for direct actions (DATABASE, RETRIEVE, CHART, or DIRECT).
Complex Diagnostic Questions: Routed to a "deep-research" path involving pre-retrieval, draft generation, supervisor-researcher coordination, finding compression, and final report synthesis.

Core Modules

Schema-Constrained Database Access:
- Utilizes a "plan-then-execute" design where the LLM predicts a business route and a structured query plan (tables, fields, filters, time ranges).
- A deterministic sanitizer validates the plan against schema allowlists and route-table mappings, removes illegal fields, normalizes aliases, and clamps unsafe limits before generating executable SQL.
- Returned rows are converted into an evidence bundle for answer generation.
Hybrid Text-Image Retrieval (RAG):
- Offline, manuals, standards, procedures, and cases are parsed into text chunks with linked image metadata.
- Online retrieval fuses dense retrieval and BM25 using a weighted scoring function: $s(d) = \alpha \cdot \text{BM25}(d) + (1-\alpha) \cdot \text{Emb}(d)$ .
- Text grounds the answer, while linked images provide visual evidence for procedures, morphology analysis, and waveform interpretation.
Multi-Agent Evidence Synthesis:
- Reserved for high-complexity questions (e.g., cross-document root-cause analysis).
- A supervisor agent decomposes requests, dispatches researcher agents, compresses findings, and generates a diagnostic report containing conclusions, evidence, risk judgments, and recommended actions.

Key Contributions

The paper outlines four primary contributions:

BESS O&M Task Model: A framework featuring complexity-aware routing tailored to battery energy storage operations.
Schema-Constrained Database Access: A mechanism ensuring safe and accurate natural-language-to-SQL translation via strict schema validation.
Hybrid Text-Image RAG: A retrieval system capable of accessing manuals, procedures, and failure-analysis figures to support multimodal reasoning.
Multi-Agent Evidence Synthesis: A workflow for generating traceable conclusions and actions by synthesizing evidence from multiple sources.

Preliminary Evaluation and Results

The system was evaluated using anonymized operational data and a private BESS maintenance knowledge base containing three business routes, seven queryable tables, 99 documents, 6,741 text chunks, 717 images, and 486 image-linked chunks.

Quantitative Findings:

Routing: The full system achieved 70.0% action accuracy with a quality score of 4.44 and a latency of 40.01s. In contrast, the ablated variant without routing dropped to 20.0% accuracy and 87.45s latency.
Database Access: The schema-constrained module achieved 100% success in generating safe, executable SQL query plans. Without schema validation, the success rate dropped to 0% due to invalid fields and route-table mismatches.
Diagnosis: The full multi-agent system achieved a quality score of 4.80 (95.33s latency), compared to 3.60 (60.99s) for the variant without multi-agent reasoning. The authors note that the multi-agent approach improves completeness and traceability despite higher latency.

Case Studies:
The prototype successfully generated four representative cases covering:

ISC-risk evidence retrieval based on micro-short-circuit scores.
Cluster-level telemetry query and trend visualization.
BESS knowledge-base retrieval with document and image evidence.
Recurrent alarm and abnormality analysis for a battery pack.

Significance and Claims

The paper positions this work as a bridge between algorithmic battery diagnosis and practical O&M decision support. The system is designed to keep energy-storage safety, operational data, fault categories, and maintenance actions at the center, using LLM agents, RAG, and tools as supporting mechanisms rather than replacements for domain logic.

The authors claim the system improves reliability through BESS-specific task routing, schema-constrained database access, hybrid text-image retrieval, and evidence-based answer synthesis. The paper explicitly states that detailed field statistics and additional expert-labeled case studies will be reported in the final paper, maintaining a modest stance on the current preliminary nature of the evaluation.