A paper in the Journal of Energy Storage proposes to find battery faults in electric vehicles by turning numeric battery signals into plain text and letting a large language model reason over it, and its tests cover ten vehicles and three human experts. The agent, called VBFDD-Agent, is meant to do more than flag an anomaly: it also names a likely cause, grades the risk and drafts a maintenance recommendation. The abstract says it performs anomaly monitoring accurately but gives no accuracy figure.
What the agent does
| Step | What the paper describes |
|---|---|
| Input | Monitoring signals, statistical features, anomaly records and state assessments from a vehicle’s battery |
| Text modelling | Conversion into structured, readable natural language descriptions, guided by domain rules and expert judgments |
| Retrieval | Historical cases and local maintenance manuals, assembled from fault handling manuals of several vehicle makers |
| Reasoning | A language model (DeepSeek-V3.2 in the experiments) produces multi label fault diagnosis and maintenance recommendations |
| Output | Structured diagnostic results with cause tracing, risk grading, impact assessment and a maintenance suggestion |
The authors argue that conventional battery fault detection, whether model based, statistical or machine learning, follows a perception and classification pattern: it takes voltage, current, temperature and related features and returns a label, a score or an alarm. Those outputs still need a person to interpret them before a technician can act. The paper aims at the step after the alarm.
Why text
Two obstacles motivate the approach. Real after sales records and fault handling procedures are confidential at vehicle makers, so textual data about battery faults is rarely public. And a label or score alone does not explain a fault to an operator. The authors therefore build a language corpus from publicly available monitoring data by describing each battery state in words, with rule informed preliminary judgments embedded in the text. They say the corpus and the generated maintenance recommendations will be released to support later work, which would let others test the method.
The introduction positions the language model as an organiser of engineering evidence, not a replacement for numeric diagnostic models. In the authors’ design the numeric signals reveal the deviation and the language layer connects it to cases and manuals, which fits a safety relevant domain.
What was tested, and what was not reported
The experiments use operating data from ten electric vehicles, obtained from the National Data Alliance of New Energy Vehicles, and compare the agent with conventional baselines that take numeric features and return binary multi label alarm vectors. For the maintenance advice, three battery domain experts with no conflict of interest evaluated the recommendations independently, and the paper reports that they recognised its practical value.
The limits are visible in those numbers. Ten vehicles is a small fleet, and three experts is a small panel for judging advice that will be acted on. The abstract and the sections read for this analysis give no accuracy, precision or recall values, no count of faults detected and no comparison figures against the baselines, so the claim of accurate performance cannot be quantified here. How the model behaves when the descriptive text is incomplete or the retrieved manual does not match the vehicle is not described in the material read.
How this relates to recalls and field data
Battery safety is a live topic: Volkswagen’s recall of 22,524 ID.4 vehicles turns on charging profiles and a state of charge above 80 percent, a condition that fleet monitoring could in principle track. Whether a language model layer adds value over thresholds and numeric detectors is the question such a paper has to answer with data, and this paper’s evidence is a small step in that direction.
The check is replication on more vehicles and more fault types, with published quantitative metrics and expert panels larger than three, and the release of the corpus the authors promise. Source: Chan, Chen and Pan, Electric vehicle battery fault detection and diagnosis agent: descriptive text modeling of battery digital signals, Journal of Energy Storage 182 (2026) 124866. The work was supported by Shanghai Jiao Tong University and the National Natural Science Foundation of China, and the authors declare no competing interests. The sections read were the abstract, highlights, introduction and section snippets.
Add EnergyNews.biz as a preferred source on Google
