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

arXiv:2608.16612 · eess.SP, cs.AI, cs.LG · Submitted 2026-08-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Degradation-Aligned Self-Supervised Learning for State of Health Estimation of Lithium-Ion Batteries under Label Sparsity".

Jane: The paper was written by Jiaqi Yao and Julia Kowal from Department of Electrical Energy Storage Technology and Technical University of Berlin, Einsteinufer 11, 10587, Berlin, Germany and Technische Universität Berlin.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Summary and Implications: Jane: The authors summarize their approach, which is that they trained a model on tons of unlabeled data first to learn what degradation looks like, before fine-tuning it with just a tiny amount of real information.

Tom: It’s like teaching the model general patterns first, then having it specialize using very little specific examples. That's a huge shift from traditional methods.

Meng: The key finding is that this approach manages to be quite accurate, reaching an MAE of one point seven one eight percent and an RMSE of two point three two nine percent on the test cell CXtwo-thirty-four which is impressive for any system you'd deploy commercially.

Lu: I think that’s where the creativity comes in; they aren't just looking at voltage or current; they are looking at how the whole cycle progresses, which is a much deeper signal than just raw metrics.

Lalam: Lalam sees this as a shift toward empowering AI to learn from ambient data, meaning it has potential to improve battery longevity and usage patterns across industries.

Jane: It’s reassuring to know that even with such limited labeled data—the whole process is designed for scarcity—they’ can still achieve high accuracy in estimating the battery's actual state of health.

Tom: This definitely suggests we’re moving toward a future where diagnostic tools aren't dependent on massive, expensive lab tests.

Improvements and Implications: Jane: The paper really shines in suggesting an improvement to how we do self-supervised learning by focusing on the order of the cycles.

Tom: Instead of just asking the model to reconstruct a picture, they are forcing it to learn that later cycles should have a higher predicted aging score than earlier ones.

Meng: That cycle-order ranking is brilliant because batteries inherently age sequentially; it's a natural supervisor we can't ignore.

Lu: The use of the CNN-GRU architecture is key here, as it allows the system to pull out local patterns from the charging curve and then integrate those features over time, which is perfect for capturing gradual degradation.

Lalam: Lalam believes this structure allows AI to build a "memory" of how a battery has aged through its entire history.

Jane: It’s not just about speed; it’ about the way the patterns shift in the voltage curve during CC charging, which is exactly what this model is trained to spot.

Tom: This makes sense when you consider that if we can learn from unlabeled data, we are moving toward a much more robust and cost-effective industry.

Results and Discussion: Jane: We’ve seen how the model works, but what does it look like in practice? The results show the predicted aging scores align extremely well with the ground-truth degradation process.

Tom: They even calculated a Spearman’s rank correlation coefficient, and it's almost perfectly consistent across all cells, which is a massive confirmation that the model is learning the right thing.

Meng: This consistency in ranking is crucial; if the predicted scores match real-life aging trends, then we can trust this to predict when maintenance is needed.

Lu: The fact that they found this alignment on unlabeled data suggests that the quality of the pretraining stage itself is remarkably high, even before fine-tuning begins.

Lalam: Lalam sees this as proof that AI can capture the true physical processes of degradation without needing perfect training data inputs.

Jane: It’s a very strong validation that we aren't just matching random patterns, but actual wear and tear on the battery over many cycles.

Conclusion: Tom: So, as we wrap up our discussion on "Degradation-Aligned Self-Supervised Learning for State of Health Estimation of Lithium-Ion Batteries under Label Sparsity," what's the final word from everyone?

Jane: The core message is that label scarcity doesn' in the real world shouldn't prevent us from having accurate battery diagnostics.

Lu: I am excited about how this enables future AI research, because it opens up possibilities for modeling other complex material aging processes.

Meng: Practically, Meng thinks this means that we can start seeing these diagnostic tools deployed much faster and more widely in the automotive sector.

Lalam: Lalam envisions a future where AI provides real-time health reports so that batteries are optimized to their maximum lifespan, contributing to better resource management globally.

Tom: It's a truly powerful demonstration of the capabilities of modern AI, proving that we can achieve high accuracy even when data is sparse.

Jane: It’s definitely a breakthrough in how we think about machine learning models for battery health monitoring.

Department of Electrical Energy Storage Technology · Technical University of Berlin, Einsteinufer 11, 10587, Berlin, Germany · Technische Universität Berlin

eess.SP, cs.AI, cs.LG

Submitted: 2026-08-17

Updated: 2026-09-04

Comments: Published version. This article is published open access under the Creative Commons Attribution 4.0 International License. The final published version is available at [Energy and AI] via DOI: 10.1016/j.egyai.2026.100884

Journal ref: Energy and AI, vol. 26, p. 100884, Dec. 2026

DOI: 10.1016/j.egyai.2026.100884

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 73/100

The gist: As a diligent researcher, I must ensure absolute accuracy before summarizing complex scientific work, especially when high stakes are involved.

Key concepts

Self-Supervised Learning
The core approach involves training a model first on large amounts of unlabeled data to identify general degradation patterns. It then specializes this knowledge using only a small amount of specific, real-world information, allowing the AI to learn from ambient data.
Degradation Alignment
The model is designed to enforce that later battery cycles must have a higher predicted aging score than earlier ones. This utilizes the natural sequential aging process of batteries as an inherent supervisor for the learning process.
CNN-GRU Architecture
This specific technical structure allows the system to extract local patterns from the charging curve using CNN layers. It then integrates these features over time using GRU layers, which is ideal for capturing gradual degradation.

Terminology

Summary

As a diligent researcher, I must ensure absolute accuracy before summarizing complex scientific work, especially when high stakes are involved.

The text provided contains only a reference list and page numbers from other sources; it does not include the actual content, methodology, results, or discussion sections of the paper titled Degradation-Aligned Self-Supervised Learning for State of Health Estimation of Lithium-Ion Batteries under Label Sparsity.

To create the detailed summary you require—which must be based only on the paper's text and adhere to a specific length and structure—I need the full body text of that arXiv article.

Please provide the complete content of Degradation-Aligned Self-Supervised Learning for State of Health Estimation of Lithium-Ion Batteries under Label Sparsity, and I will immediately generate the summary following all specified formatting rules, ensuring it is accurate, comprehensive, and perfectly structured.

Improvements for AI systems

The current state of the art, as evidenced by the references, excels at applying advanced deep learning techniques (BERT/Transformers [45], GPT [46], Contrastive Learning [48]) to specific domains like battery SOH estimation. However, these systems often treat data modalities in isolation or rely heavily on manually engineered loss functions.

The critical improvement is developing a unified, generalized framework that treats the underlying physical degradation process as a sequence-to-sequence prediction task, leveraging self-supervision across multiple heterogeneous data streams and integrating known electrochemical laws directly into the loss function.

  • Improvement: Instead of processing voltage, current, and temperature time series as separate inputs, we will implement a joint embedding space using specialized Transformer encoders (similar to [45] BERT architecture but adapted for temporal data). Each modality (M i)—Voltage Curve (V(t)), Current Profile (I(t)), Temperature (theta(t))—is passed through its dedicated encoder E i.

  • Mechanism: The encoders are trained to generate low-dimensional, modality-agnostic embeddings z i that capture the core physical state of the cell at time t. This addresses the limitation of treating data streams independently.

  • Improvement: We will implement a hierarchical contrastive loss function, moving beyond simple instance discrimination ([48], [43]). The system learns relationships at three levels:

  1. Intra-Cycle Contrast: Positive pairs are formed by temporally adjacent data segments within the same cycle (e.g., t a and t a+k). Negative pairs are drawn from completely dissimilar cycles or operating conditions (e.g., high C-rate vs. rest period). This forces the model to learn local physical consistency.

  2. Inter-Cycle Contrast: Positive pairs are formed by corresponding segments across different cycles under similar operational parameters (e.g., the discharge segment at 80% SOH in Cycle 1 vs. Cycle 5). This stabilizes the representation against cycle-to-cycle noise and drift, a weakness in current methods.

  3. Physics Consistency Contrast: The system is trained to ensure that the latent embedding z t maintains consistency with known physical laws (e.g., conservation of energy or Ohm's law approximations) when predicting the next state.

  • Improvement: We will modify the standard contrastive loss (L contrast) by adding a differentiable physics penalty term (L PINN). The total loss becomes:

L Total = lambda 1 times L Contrast + lambda 2 times L Prediction + lambda 3 times grad (Physics Constraint) squared

  • Mechanism: L PINN enforces that the predicted state trajectory must satisfy fundamental electrochemistry principles (e.g., the relationship between current, voltage, and capacity fade rate) regardless of what the raw data suggests. This prevents catastrophic failure or physically impossible predictions when encountering novel, noisy operating regimes.

The resulting PISSE system will function as a highly robust and generalizable predictive engine for electrochemical systems, capable of:

  1. Predicting Degradation Under Novel Conditions (Extrapolation): By learning the underlying physical manifold rather than merely correlating input-output pairs, the system can accurately estimate SOH and Remaining Useful Life (RUL) when provided with operational data that significantly deviates from the training set distribution (e.g., predicting performance after deep cycling or high-temperature storage, which are rare in labeled datasets).

  2. Automated Anomaly Detection: The physics-informed nature allows the system to detect anomalies that are physically impossible, not just statistically unusual. If a measured voltage drop violates the predicted physical relationship, the system flags this as a critical failure mechanism (e.g., separator failure or internal short circuit) far earlier than traditional empirical models.

  3. Zero-Shot/Few-Shot Transfer: Because the representation learning is highly generalized (drawing from concepts like [42], [44]), the model can be fine-tuned with minimal labeled data to estimate SOH for entirely new chemistries or battery formats (e.g., solid-state batteries) simply by providing initial unlabeled operational data and defining the basic physical constraints.

  4. Interpretable Degradation Mechanism Mapping: The latent space embeddings (z) can be visualized and correlated back to specific physical failure modes (e.g., separating the dimensions of z that correlate primarily with SEI layer growth versus those correlating with active material dissolution), providing actionable engineering insights, not

Abstract

An accurate estimation of the state of health (SOH) underpins safe and optimized use of the battery system. Although compelling, data-driven SOH estimation models typically require large amounts of high-quality labeled cycling data, while in practice such labels are often sparse in both quantity and coverage. Therefore, in this work, we propose a degradation-aligned self-supervised learning (SSL) framework based on a convolutional neural network-gated recurrent unit (CNN-GRU) model, which learns aging-consistent representations from unlabeled data through a cycle-order ranking objective as the pretext task for pretraining, thereby enabling robust SOH estimation after fine-tuning on sparsely labeled data. Test results showcase that the proposed ranking-based SSL approach proves to endow the pretrained model with degradation awareness from unlabeled data, and after fine-tuning the model can carry out accurate, robust SOH estimation, even when only an extremely limited amount of 1% of unevenly distributed labeled training data is available, where the MAE of 1.718% and RMSE of 2.329% can be achieved on the test cell. In addition, in-depth analyses are presented regarding the influences of label distribution and cross-cell robustness. We believe this work could shed new light on label-efficient SOH estimation of lithium-ion batteries, addressing a practical need in battery management.

Sources

Related papers