🤖 machine learning

Degradation-Aligned Self-Supervised Learning for State of Health Estimation of Lithium-Ion Batteries under Label Sparsity

This paper proposes a degradation-aligned self-supervised learning framework using a CNN-GRU model and a cycle-order ranking pretext task to enable accurate and robust lithium-ion battery state of health estimation even when trained on extremely sparse and unevenly distributed labeled data.

Original authors: Jiaqi Yao, Julia Kowal

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaqi Yao, Julia Kowal

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

Batteries are the silent engines of the modern world, powering everything from the smartphones in our pockets to the electric vehicles on our roads. At the heart of every battery is a simple but critical question: how much life is left? Engineers call this the "state of health." It is a measure of how much capacity a battery has lost since it was new, much like checking the tread on a tire to see how much road it can still handle. Knowing this number is vital for safety and efficiency, allowing us to predict when a battery needs maintenance or replacement before it fails. However, getting an accurate answer is surprisingly difficult. The true state of health is hidden inside the battery's chemistry and cannot be measured directly with a simple tool. Instead, it must be estimated by watching how the battery behaves while it charges and discharges.

For years, scientists have tried to teach computers to make these estimates by feeding them vast amounts of data. The problem is that this data usually needs to be carefully labeled with the correct answers, a process that requires running batteries through years of slow, expensive tests in a lab. In the real world, such perfect data is rare. Most batteries in use today have only a few known data points, while the rest of their history remains a mystery. This scarcity of labeled information has long been a bottleneck, making it hard to build reliable systems that can predict battery health in everyday situations.

A team of researchers at the Technical University of Berlin has found a way to work around this shortage. They developed a new method that allows a computer model to learn the secrets of battery aging using mostly unlabeled data. Imagine a student trying to learn a language. Traditionally, they would need a teacher to correct every sentence they speak. This new approach is more like letting the student listen to thousands of hours of native speakers and figure out the grammar and flow on their own, only needing a teacher to check their work occasionally. The researchers taught their model to recognize the natural order of aging. Since batteries always degrade in a specific sequence—getting older and weaker with every charge cycle—the model learned to rank charging sessions from "fresh" to "worn" without ever being told the exact age of the battery.

To do this, the researchers used a specific type of artificial intelligence that combines two powerful techniques. One part of the system acts like a microscope, zooming in on tiny, local patterns in the voltage curves that appear as the battery charges. The other part acts like a memory, connecting these small patterns over time to understand the bigger story of how the battery is changing. By training this system on a massive amount of unlabeled charging data, the model learned to identify the subtle shifts that happen as a battery ages. It learned that a slight change in the shape of the charging curve at the beginning of a cycle is different from the same change at the end of a cycle, even without knowing the exact number of cycles that had passed.

Once the model had learned these patterns through this self-guided practice, the researchers tested it with very little labeled data. They gave the model only one percent of the usual amount of labeled examples—just a tiny fraction of the battery's life history—to fine-tune its final predictions. The results were striking. Even with such sparse information, the model could estimate the battery's health with remarkable accuracy. In their tests, the average error was less than two percent, a level of precision that rivals methods requiring far more data. This suggests that the model had successfully absorbed the underlying logic of battery aging during its self-supervised training phase, allowing it to generalize well even when the specific examples were few and far between.

The study also explored how the distribution of these few labeled examples affected the outcome. They found that it mattered less if the labels were spread evenly across the battery's life and more if they were concentrated in the later stages of aging. When the model was trained on data from the deep-degradation phase, where the battery's behavior changes most dramatically, it performed even better. Conversely, when the limited labels came only from the early, stable years of the battery's life, the model struggled to learn the complex changes that happen later. This insight reveals that the most valuable information for teaching a computer about battery health lies in the moments when the battery is showing clear signs of wear.

This work offers a promising path forward for the future of battery management. By showing that machines can learn the story of battery aging from the raw, unlabeled data that is already available in the real world, the researchers have removed a major barrier to creating smarter, safer battery systems. Their approach does not require waiting for years of perfect laboratory data to become available. Instead, it leverages the vast oceans of operational data that are already being collected, teaching computers to understand the language of aging so that we can better care for the power sources that drive our lives.

Technical Summary: Degradation-Aligned Self-Supervised Learning for State of Health Estimation of Lithium-Ion Batteries under Label Sparsity

Problem Statement
Accurate State of Health (SOH) estimation is critical for the safety and optimization of lithium-ion battery systems. While data-driven deep learning approaches have shown promise, they typically rely on large volumes of high-quality, labeled cycling data. In practice, obtaining such labels is challenging because SOH is not directly measurable and requires time-consuming, expensive standardized checkup tests. Consequently, real-world battery aging data often suffer from "label sparsity," characterized by both a low quantity of labeled samples and uneven coverage across the battery's lifespan (e.g., a lack of data in deep-degradation regimes). Existing self-supervised learning (SSL) methods in this domain often focus on reconstruction-based pretext tasks, which may not align the learned representations with the specific task of SOH estimation, as battery aging is inherently an ordered, progressive process.

Methodology
The authors propose a Degradation-Aligned Self-Supervised Learning (SSL) framework designed to leverage abundant unlabeled data to learn aging-consistent representations, which are then fine-tuned on sparse labeled data.

Model Architecture (CNN-GRU):
- The core model consists of an Encoder and a Head.
- The Encoder utilizes a Convolutional Neural Network (CNN) with two stacked 1D convolutional blocks to extract local patterns from Constant Current (CC) charging voltage curves. These features are then sequentially integrated by a Gated Recurrent Unit (GRU) layer to capture temporal dependencies.
- The Head is a fully connected network used for different tasks depending on the training phase.
Pretext Task: Cycle-Order Ranking:
- Instead of reconstructing input data, the pretraining phase employs a cycle-order ranking objective.
- The model is trained to predict a dimensionless "aging score" ( $s_{age}$ ) such that the scores respect the relative chronological order of battery cycles.
- A pairwise logistic ranking loss is minimized, ensuring that for any two samples from the same cell, the one from a later cycle (higher degradation) receives a higher aging score.
- To ensure reliability, pairs with identical cycle numbers or cycle differences below a threshold ( $d_{min} = 50$ ) are excluded to avoid noise from capacity recovery phenomena.
Training Workflow:
- Self-Supervised Pretraining: The CNN-GRU model is trained on a large volume of unlabeled CC charging curves using the ranking-based SSL algorithm. The encoder learns to map charging curves to an aging score that correlates with degradation.
- Downstream Adaptation (Fine-Tuning): The pretrained encoder parameters are frozen. A new head is instantiated for SOH regression. The model is then fine-tuned using a small, sparsely labeled dataset (ranging from 1% to 100% of available labeled data).

Key Contributions

Ranking-Based SSL Framework: The paper introduces a novel SSL framework that uses cycle-order ranking as a pretext task. This approach explicitly aligns the learned representations with the intrinsic progressive nature of battery degradation, addressing the limitations of reconstruction-based methods.
CNN-GRU Architecture: A specific deep learning architecture is developed to extract local patterns from charging curves via CNNs and integrate them temporally via GRUs, suitable for both self-supervised pretraining and SOH regression.
Analysis of Label Distribution: The study provides an in-depth analysis of how the distribution of sparse labels (uniform, random, early-only, late-only) impacts model performance, offering insights into optimal data acquisition strategies under label scarcity.

Experimental Results
The framework was evaluated using the CALCE battery aging dataset (cells CX2-34, CX2-36, CX2-37, CX2-38).

Degradation Alignment: Before fine-tuning, the pretrained model's predicted aging scores showed an almost perfect monotonic inverse relationship with ground-truth SOH, with Spearman's rank correlation coefficients ( $\rho$ ) ranging from -0.996 to -0.998 across all cells. This confirms the model successfully captured degradation-aligned information from unlabeled data.
Performance under Label Sparsity:
- The proposed SSL-Rank model significantly outperformed both a pure Supervised Learning (SL) baseline and other SSL approaches (Reconstruction-based SSL and Multi-Objective SSL).
- While the pure SL model failed to estimate SOH accurately when labeled data dropped below 40%, the SSL-Rank model maintained robust performance even with only 1% of labeled data.
- At the 1% labeled data ratio, the SSL-Rank model achieved a Mean Absolute Error (MAE) of 1.718% and a Root Mean Square Error (RMSE) of 2.329% on the test cell, with an $R^2$ of 0.970.
- The SSL-Rank model consistently outperformed SSL-Recon and SSL-MO across all labeled data ratios (1% to 100%), demonstrating that the ranking objective provides superior initialization for downstream tasks.
Label Distribution Influence:
- Experiments revealed that "late-only" labeled data (samples from deep degradation stages) yielded better performance than "early-only" data, even with the same small sample size. This is attributed to the more pronounced aging-related pattern shifts in the voltage curves during late-life stages.
- Conversely, models fine-tuned on "early-only" data failed when the sample size was very small (e.g., 5%), highlighting the need for data diversity or quantity to compensate for limited degradation stages.

Significance and Claims
The authors claim that this work sheds new light on SOH estimation for lithium-ion batteries in real-world applications where labeled data is scarce. By exploiting the intrinsic cycle-order information in unlabeled data, the proposed framework enables robust SOH estimation with minimal labeled data (as low as 1%). The study demonstrates that leveraging unlabeled data through degradation-aligned pretext tasks not only mitigates label scarcity but also improves model initialization, leading to superior performance even when labeled data is relatively abundant. The authors note that the current study is limited to a single cell model under specific cycling conditions and suggest that future work should extend this framework to cross-cell and cross-condition settings to better reflect real-world variability.

More like this