⚡ electrical engineering

A Degradation-Aware SegRNN Method Optimized by ASHA for Lithium-Ion Battery SOH Prediction

This paper proposes a degradation-aware SegRNN method optimized by the asynchronous successive halving algorithm (ASHA), which integrates reversible instance normalization, segment-based GRU encoding, a cumulative degradation increment prediction mechanism, and a degradation-aware training strategy to achieve robust and continuous multi-step state-of-health prediction for lithium-ion batteries under small-sample and noisy conditions.

Original authors: fubin zhang

Published 2026-09-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: fubin zhang

Original paper licensed under CC BY 4.0 (https://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 heart of modern life, powering everything from the smartphones in our pockets to the electric cars on our roads. Inside these devices, lithium-ion batteries store energy through a complex chemical dance, but like any living thing, they age. Over thousands of charge and discharge cycles, they slowly lose their ability to hold a full charge, a process known as degradation. Engineers call the measure of this remaining health the "state of health." Knowing exactly how much life is left in a battery is critical; it ensures safety, prevents unexpected failures, and helps manage energy efficiently. However, predicting this decline is notoriously difficult. Real-world data is often messy, filled with random noise and sudden, confusing fluctuations that make it hard to see the true, slow downward trend of a battery's life.

To solve this problem, a researcher named Fubin Zhang from the Zhangzhou Institute of Technology has developed a new way to forecast how a battery will age. The challenge lies in the fact that batteries do not degrade in a perfectly straight line. Their health drops steadily over time, but along the way, they experience small, erratic jumps and dips caused by measurement errors or temporary chemical shifts. Traditional computer models often get confused by these short-term glitches, leading to predictions that wobble or drift away from reality. Zhang's approach treats the battery's history not as a single, continuous stream of numbers, but as a series of distinct chapters. By breaking the past into smaller, manageable segments, the new method can focus on the local story of each part while still understanding the overall plot of the battery's life.

The core of this new method is a system that learns to ignore the noise and focus on the pattern. First, the computer takes a window of the battery's recent history and normalizes it, effectively smoothing out the differences in scale so that a small fluctuation doesn't look like a major crisis. Then, it chops this window into several consecutive pieces, or segments. Instead of trying to memorize every single data point, the system analyzes how these segments relate to one another over time. It uses a specialized type of neural network, a form of artificial intelligence designed to recognize sequences, to understand how the battery's condition evolves from one segment to the next. This allows the model to capture the true degradation state, filtering out the static of random noise.

Once the system understands the current state of the battery, it predicts the future not by guessing the next number directly, but by calculating the small steps of decline that will happen next. Imagine the battery's future health as a staircase going down; instead of guessing the height of the final step, the model predicts the height of each individual step and then adds them up to find the total drop. This "cumulative" approach ensures that the predicted path remains smooth and continuous, respecting the natural, gradual nature of battery aging. To make sure the model learns this correctly, the researchers added a special training rule that forces the computer to pay attention not just to the final numbers, but to the direction and slope of the curve. This ensures the prediction follows the same gentle downward trend as the real battery, even when the data is messy.

Because there are many ways to set up such a computer model, the researchers used a smart search algorithm to find the perfect settings for their specific task. This algorithm, known as ASHA, acts like a highly efficient scout. It tests hundreds of different configurations, quickly discarding the ones that perform poorly and focusing its energy on the most promising candidates. This process ensures that the final model is finely tuned to the specific characteristics of the battery data it is analyzing. The researchers tested this method on two well-known collections of battery data: one from NASA and another from the University of Maryland. These datasets included both short, simple battery histories and long, complex ones with significant fluctuations.

The results showed that this new method outperformed several existing, sophisticated models. On the short, noisy datasets, the new approach reduced the prediction error significantly, achieving a level of accuracy that was noticeably better than the next best method. On the longer, more complex datasets, where batteries showed strange behaviors like temporary recoveries in capacity, the model still managed to track the true health curve closely. The researchers found that every part of their new system contributed to this success. Removing any single component, such as the way the data was chopped into segments or the method used to smooth the predictions, resulted in a less accurate model. This confirmed that the combination of segmenting the data, predicting step-by-step increments, and using the smart search algorithm was essential for handling the messy reality of battery aging.

Ultimately, this work offers a more reliable way to look into the future of a battery's life. By breaking the problem down into smaller, understandable pieces and forcing the computer to respect the continuous nature of degradation, the method provides a stable and accurate picture of battery health. While the current study focused on data from specific battery types and did not yet include other factors like temperature or voltage, the approach demonstrates a clear path forward. It suggests that by teaching machines to see the story in the noise, we can better manage the energy systems that power our world, ensuring they remain safe and efficient for as long as possible.

Technical Summary: A Degradation-Aware SegRNN Method Optimized by ASHA for Lithium-Ion Battery SOH Prediction

Problem Statement

Accurate multi-step State of Health (SOH) prediction for lithium-ion batteries is critical for maintenance and energy management but faces significant challenges in practical applications. Existing data-driven methods often struggle under small-sample conditions and noisy environments. Key limitations include:

Susceptibility to Local Fluctuations: Models often fail to distinguish between measurement noise/capacity regeneration and actual degradation trends.
Insufficient Trajectory Continuity: Direct multi-step prediction methods frequently produce unstable trajectories that lack consistency with the continuous physical degradation process of batteries.
Overfitting and Parameter Sensitivity: Complex deep learning models are prone to overfitting on limited datasets, and their performance is highly sensitive to hyperparameter selection and window length configurations.

Methodology

The paper proposes a Degradation-Aware SegRNN framework, optimized by the Asynchronous Successive Halving Algorithm (ASHA), to address these challenges. The methodology consists of four core components:

1. Data Preprocessing and Segmentation

Noisy Sequence Construction: Original SOH sequences are augmented with noise to simulate real-world observation conditions.
Reversible Instance Normalization (RevIN): Applied to input windows to mitigate local distribution discrepancies caused by different degradation stages.
Segmented Representation: The historical window is divided into consecutive degradation segments. A linear embedding layer extracts local representations, which are then processed by a Gated Recurrent Unit (GRU) encoder. This encodes the temporal evolution among segments to extract the current degradation state ( $z_t$ ).

2. Cumulative Degradation Increment Prediction (CDIP)

Instead of predicting absolute SOH values directly, the model employs a CDIP mechanism:

The model predicts the SOH increment ( $\Delta \hat{s}$ ) for each future step.
The future SOH trajectory is reconstructed by cumulatively summing these increments starting from the last observed value.
This approach enforces temporal continuity, ensuring the predicted trajectory reflects the gradual nature of battery degradation.

3. Degradation-Aware Training Strategy (DATS)

To improve trend consistency under noisy conditions, a composite loss function is introduced:

Horizon-Weighted MSE ( $\mathcal{L}_{wmse}$ ): Penalizes point-wise errors with weights assigned to different prediction horizons.
Slope-Consistency Loss ( $\mathcal{L}_{slope}$ ): Constrains the local variation trend by minimizing the difference between the slopes of the predicted and true trajectories.
Total Loss: $\mathcal{L} = \mathcal{L}_{wmse} + \lambda \mathcal{L}_{slope}$ , balancing point accuracy with trajectory smoothness.

4. ASHA-Based Hyperparameter Optimization

The Asynchronous Successive Halving Algorithm (ASHA) is used to adaptively optimize key hyperparameters (e.g., window length $L$ , segment length $l_s$ , learning rate, hidden size). ASHA efficiently searches the hyperparameter space by:

Training multiple candidate configurations with a small initial budget.
Eliminating poor-performing candidates early.
Allocating larger training budgets to promising configurations.
This ensures the model adapts to the specific characteristics of different battery degradation sequences (short vs. long).

Key Contributions

Degradation-Aware SegRNN Model: A novel architecture that extracts local degradation segments and models their temporal evolution via GRU, enhancing modeling capability under small-sample conditions.
CDIP and DATS Mechanisms: The design of a Cumulative Degradation Increment Prediction mechanism to ensure trajectory continuity, coupled with a Degradation-Aware Training Strategy to enforce slope consistency in noisy environments.
ASHA Integration: The application of ASHA for adaptive hyperparameter optimization, improving the model's robustness across different datasets and degradation patterns.

Experimental Results

The proposed method was evaluated on the NASA (short-sequence, B0006, B0007) and CALCE (long-sequence, CS2-36, CS2-38) datasets, comparing against Transformer, TabM, HistGBR, and PatchTSMixer.

NASA Dataset (Small-Sample/Short-Sequence):
- The proposed model achieved the lowest error rates. For B0006, it reached an R² of 0.942 and RMSE of 7.60E-03, outperforming the best baseline (HistGBR) by reducing RMSE by approximately 26.2%.
- For B0007, it achieved an R² of 0.915 and RMSE of 4.99E-03.
CALCE Dataset (Long-Sequence/Complex Fluctuations):
- The model demonstrated superior adaptability to capacity regeneration and local fluctuations.
- For CS2-36, it achieved an R² of 0.991 and RMSE of 1.60E-02.
- For CS2-38, it achieved an R² of 0.988 and RMSE of 1.70E-02.
Ablation Studies:
- Experiments confirmed that each module (RevIN, CDIP, DATS, ASHA) contributes incrementally to performance. The full model (V4) consistently outperformed ablated variants, validating the synergy of the proposed components.

Significance and Claims

The paper claims that the proposed method effectively addresses the instability and lack of continuity in multi-step SOH prediction under noisy and small-sample conditions. By explicitly modeling degradation segments and enforcing trajectory consistency through CDIP and DATS, the method achieves higher fitting accuracy and lower prediction errors compared to state-of-the-art deep learning and machine learning baselines.

The author notes that while the method is effective, it currently relies solely on SOH sequences. Future work is identified as needing to incorporate multi-source operating features (voltage, current, temperature) and validate generalization across different battery material systems and operating conditions. The study concludes that the ASHA-optimized SegRNN provides a robust approach for battery health assessment but acknowledges the computational cost of the optimization process as a factor for future consideration.