[Paper Review] Federated Battery Diagnosis and Prognosis

Nur Banu Altinpulluk, Deniz Altinpulluk+5 more|arXiv (Cornell University)|Oct 14, 2023

Advanced Battery Technologies ResearchEngineering3 citations

TL;DR

This paper proposes a federated learning framework for lithium-ion battery diagnosis and prognosis that preserves data privacy by training models locally and sharing only model parameters, not raw data. The approach achieves competitive remaining useful life prediction accuracy comparable to centralized methods while eliminating data transfer and privacy risks, demonstrating strong performance across multiple battery datasets with minimal communication overhead.

ABSTRACT

Battery diagnosis, prognosis and health management models play a critical role in the integration of battery systems in energy and mobility fields. However, large-scale deployment of these models is hindered by a myriad of challenges centered around data ownership, privacy, communication, and processing. State-of-the-art battery diagnosis and prognosis methods require centralized collection of data, which further aggravates these challenges. Here we propose a federated battery prognosis model, which distributes the processing of battery standard current-voltage-time-usage data in a privacy-preserving manner. Instead of exchanging raw standard current-voltage-time-usage data, our model communicates only the model parameters, thus reducing communication load and preserving data confidentiality. The proposed model offers a paradigm shift in battery health management through privacy-preserving distributed methods for battery data processing and remaining lifetime prediction.

Motivation & Objective

• To address the critical challenge of data privacy and communication overhead in centralized battery health management systems.
• To enable collaborative battery prognosis across multiple entities without sharing raw current-voltage-time-usage data.
• To develop a federated learning framework that maintains model accuracy while ensuring data confidentiality and compliance with data residency regulations.
• To evaluate the performance of the federated model against centralized and baseline periodic replacement policies in real-world battery datasets.
• To quantify the trade-offs between model accuracy, communication efficiency, and data privacy in battery health management.

Proposed method

• The framework employs a federated autoencoder architecture to learn low-dimensional representations of battery current-voltage-time-usage data in a decentralized manner.
• Local models are trained on-site at each data source using only local data, with model weights aggregated centrally to update a global model.
• Information diodes are implemented to prevent raw data transfer, allowing only model parameters to be exchanged, thus ensuring data privacy.
• Statistical and temporal features are extracted from raw voltage and current data, including mean, standard deviation, and capacity/energy metrics, to enhance model input representation.
• The system uses a batch-federated training strategy with varying cluster sizes to simulate different levels of data aggregation and communication efficiency.
• Model performance is evaluated using remaining useful life (RUL) prediction error and cost rate metrics, comparing federated learning to centralized and age-based replacement benchmarks.

Experimental results

Research questions

• RQ1 Can a federated learning framework achieve comparable RUL prediction accuracy to centralized models without transferring raw battery data?
• RQ2 How does the federated approach perform in terms of communication efficiency and data privacy compared to centralized data collection?
• RQ3 What is the impact of varying data aggregation levels (cluster size) on model accuracy and convergence in a federated battery prognosis system?
• RQ4 How does the federated model compare to age-based periodic replacement policies in terms of unused battery life and operational cost?
• RQ5 To what extent can statistical and temporal features derived from current-voltage-time data improve the performance of federated battery prognosis models?

Key findings

• The fully federated model achieved a cost rate of 20.3 on the Nature Energy database with a 100-day threshold, outperforming age-based policies and minimizing unused battery life.
• On the Argonne NMC532 dataset, the federated approach reduced unused life to 124.6 days (vs. 551.8 under age-based policy) and cost rate to 21.4, demonstrating significant economic benefits.
• The batch-federated experiments showed that as the number of clusters increased (reducing data aggregation), prediction error increased, confirming that higher data locality improves model performance.
• The fully centralized model had the lowest error, but the federated model approached its performance with significantly reduced communication and data transfer.
• The information diode mechanism successfully prevented raw data leakage, enabling secure collaboration across entities with competing interests.
• The model reduced average unavailable days to 1.3–3.9 across datasets, indicating minimal operational disruption from premature or delayed battery replacements.