GRAPE-UPS: Direction-Constrained Peer Anomaly Detection for Weak Batteries in Data-Center UPS Strings

Overview

In a series-connected UPS battery string the same discharge current flows through every unit, so the first unit to reach an unsafe voltage limits the usable capacity of the whole string. The difficulty is that load and environmental variation can hide the weak unit's degradation signature. GRAPE-UPS uses each shared load event to define a robust peer reference and keeps only departures in physically adverse directions; the directed peer score determines maintenance ranking, while one-class reconstruction and labelled validation support a separate, weakly supervised alert pathway.

Evaluation is almost entirely on simulated strings plus a separate electrochemical simulation benchmark. Across 15 independent simulated realisations per difficulty level, mean AUPRC was 0.904, 0.839 and 0.823 for easy, medium and hard conditions and exceeded all 18 comparison methods, with paired differences significant after correction across 54 comparisons. On the independent electrochemical benchmark, directed peer scoring achieved AUPRC 0.848 under nominal conditions and 0.850-0.877 under operating shifts.

The method requires comparable units and a healthy majority in the string. The available field record had no independent health labels, so prospective validation against capacity tests is still needed; results were evaluated under simulation conditions and transfer to real installations is not proven.

The paper is therefore best read as an operations paper with limited design content. Its design relevance is in showing how valuable per-block battery monitoring and peer comparison on each load event are for detecting weak units.

Key points

1. In a series UPS string the same discharge current flows through every unit, so the first unit to reach an unsafe voltage limits the usable capacity of the whole string.
2. Across 15 simulated realisations per difficulty level, mean AUPRC was 0.904, 0.839 and 0.823 for easy, medium and hard conditions, better than all 18 comparison methods.
3. All 54 paired comparisons were significant after Holm correction; direction and reference controls showed the contribution of combining event-conditioned peers with adverse-direction evidence.
4. On an independent electrochemical simulation benchmark, directed peer scoring achieved AUPRC 0.848 under nominal conditions and 0.850-0.877 under operating shifts.
5. The method requires comparable units and a healthy majority in the string.
6. The field record lacked independent health labels, so prospective validation against capacity tests is still needed.

Using it in design

• Require block-level (per-unit) voltage and resistance monitoring in UPS battery monitoring rather than string totals, citing that one weak unit in a series string limits the whole string's capacity.
• Produce the weak-unit ranking by relative comparison with peers on the same load event, and turn only adverse-direction deviations (such as higher voltage drop) into alerts.
• Validate monitoring results with periodic capacity tests; do not treat thresholds based only on simulation as definitive in the field.

Who should read it

Facility engineers and operations teams setting UPS battery monitoring and maintenance strategy in data centres.

Document contents

1. I. Introduction
2. II. Related Work (A. Threshold-, trend-, and model-based battery diagnosis, B. Statistical and learned pack-relative diagnosis, C. General multivariate time-series anomaly detection, D. Research gap and positioning of GRAPE-UPS)
3. III. GRAPE-UPS Method (A. Problem formulation and design rationale, B. Direction-constrained group-relative evidence, C. Decoupled reconstruction and decision, D. Temporal evidence accumulation, E. End-to-end decision procedure)
4. IV. Experimental Design (A. Simulation datasets, B. Baselines and evaluation discipline, C. Evaluation metrics and sensitivity analyses, D. Physical and operational robustness)
5. V. Results
6. VI. Discussion (A. What the results establish, B. Relation to existing battery and time-series methods, C. Operational interpretation, D. Limitations and next validation step)
7. VII. Conclusion

Read the document

Publication details

Authors

Shibo Luo; Bin Shen; Xiaoning Chen; Yiquan Fang; Huifeng Wang

Publisher

IEEE Access (Institute of Electrical and Electronics Engineers (IEEE))

Year

2026

Pages

1-1

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1109/ACCESS.2026.3753456

Licence

CC BY 4.0 — share with attribution

Licence evidence

ieeexplore.ieee.org

Citation