Source-linked AI summary
Closed-loop domain generation enhances battery health diagnosis
Yue Wu, Ziyu Mo, Jinze Dai, Ziqing Xia, Shengyu Tao, Yusheng Zheng, Yunhong Che
TL;DR
Electrode-level battery diagnosis must contend with operating-condition shifts and scarce full-lifespan labels in new target domains. The paper uses VAE-based latent extrapolation and Bayesian optimization to generate physically constrained cross-domain samples, reporting lower diagnosis errors across cross-domain evaluations. Its generation module is presented as a plug-in data enhancer for existing diagnosis networks.
Problem
Dynamic operating conditions can shift degradation trajectories, while full-lifespan labeled data for each new target domain are costly and time-consuming to collect.
Source: §1 · p. 2Method
Results
Across four cross-domain scenarios the method is reported superior; in the typical Lab2Real scenario its RMSE values are 1.20%, 2.36%, and 1.70% for LAMpe, LAMne, and LLI, respectively.
Source: §5 · p. 14Takeaways & Limitations
The domain-generation module is decoupled from downstream diagnosis networks and is presented as a plug-and-play data enhancer for cross-domain battery management systems.
Source: §5 · p. 14
Abstract
from arXiv · showhide
Reliable electrode-level degradation diagnosis is essential for accurate health estimation and non-destructive monitoring in advanced battery management. Diagnosis models developed under controlled laboratory conditions can fail when operating conditions change, while acquiring sufficient full-lifespan data in new target domains remains prohibitively time-consuming and practically unfeasible. Here, we propose a closed-loop domain generation framework to enable cross-domain battery health diagnosis under dynamic conditions. Physics-constrained domain generation is achieved through Variational Autoencoder-based latent extrapolation, which maps the source data and few-shot early-life target domain data into a shared latent space and generates cross-domain samples within physical boundaries. Bayesian optimization closes the generation-diagnosis loop by automatically calibrating the sampling, latent regularization, extrapolation, perturbation, and generation-volume hyperparameters using leakage-free development validation, thereby strictly excluding later-life target data from optimization and ensuring the quality and quantity of generated data. Numerical results demonstrate that, across four cross-domain cases, the proposed framework achieves average root mean square errors of 1.53%, 2.15%, and 2.02% for positive electrode active material loss, negative electrode active material loss, and lithium inventory loss, respectively, whereas traditional methods suffer from notably higher errors or even complete failure. This framework provides a general pathway for onboard and physics-consistent electrode-level degradation diagnosis, independent of downstream diagnosis models.
1. Introduction
Electrode-level diagnosis can reveal degradation modes obscured by capacity-only estimates, but existing approaches face domain shifts and limited target-domain labels under realistic dynamic operation. This paper proposes physics-constrained domain generation to improve cross-domain diagnosis data quality.
1. Introduction
- Electrode-level diagnosis quantifies lithium-inventory and active-material losses by reconstructing full-cell OCV from half-cell potentials, whereas macro-capacity can obscure internal degradation mechanisms.Prior work also estimates electrode aging from partial charging profiles and develops material-specific diagnostic models.Source: §1 · p. 1
- Dynamic currents, partial operating windows, temperature variation, noise, and changed degradation trajectories challenge diagnosis methods built around standardized laboratory discharges.Existing responses include field-data integration, transformed high-rate features, residual learning, random partial data, and physics-informed learning.Source: §1 · p. 2
- Cross-domain diagnosis often has only limited early-life target data, while existing few-shot and generative studies mainly address macroscopic health indicators rather than electrode-level degradation modes.Domain shifts can arise from operating profiles, current rates, temperatures, chemistries, and manufacturers.Source: §1 · p. 2
- Battery domain generation must preserve coupled electrode capacities, lithium inventory, utilization windows, and full-cell OCV while bridging feature shifts with limited later-life labels.The open challenge is jointly controlling physical consistency, target-domain relevance, and the quantity of generated samples.Source: §1 · p. 2
- The paper proposes a cross-domain diagnosis framework that improves data quality through a physics-constrained generation mechanism intended to extend across diagnostic methods.Its contributions include VAE-based latent extrapolation and automated Bayesian optimization of generation boundaries and volume under leakage-free validation.Source: §1 · p. 2
2. Physics-informed data foundation and domain definition
The section establishes the physical basis for electrode-aging labels, describes dynamic-discharge feature extraction, and defines cross-domain evaluations using a 92-cell aging dataset.
2.1. Battery aging mechanisms and half-cell OCV model
- Full-cell OCV is reconstructed by matching positive and negative half-cell potentials, and fitted electrode parameters support quantification of LLI, LAMpe, and LAMne.For the fitted model, the cathode utilization parameter y100 is set to zero to avoid multi-solution problems because the cathode shifts minimally across cycling.Source: §2.1 · pp. 2–3
2.2. Diagnosis feature extraction under dynamic discharging profiles
- Randomly sampled dynamic-discharge segments provide voltage, local cumulative capacity, and equivalent full cycle features, resampled to 100 points for model inputs.Source: §2.2 · p. 3
2.3. Dataset description and cross-domain definition
- The dataset contains 92 cells spanning conventional, pulsed, synthetic, and real-world driving discharges at C/16, C/10, C/5, and C/2; five cases test same-condition, Lab2Real, rate-transfer, and unseen-rate shifts.Case 3 is an extreme data-scarcity setting with only 8 training cells; 90 cells are used after excluding two failed-protocol duplicates.Source: §2.3 · p. 4
3. Battery health diagnosis based on closed-loop domain generation
The framework combines few-shot early-life target data with source data to generate physics-constrained synthetic samples, then uses leakage-free Bayesian optimization to calibrate generation for downstream electrode-level diagnosis.
3. Battery health diagnosis based on closed-loop domain generation
- The framework uses shared-latent VAE encoding, latent mixup, controlled extrapolation, and perturbation to generate synthetic samples for diagnosis of LAMpe, LAMne, and LLI.Only early-life target samples inform generation; the remaining target lifespan is held out for final testing.Source: §3 · p. 5
3.1. Early few-shot sampling of the target domain
- The sampling rule selects initial target-domain sequences as few-shot training data and strictly isolates all remaining sequences for unseen testing.The motivation identifies the first 5% of lifespan as a realistic early-life data setting.Source: §3.1 · p. 5
3.2.1. Cross-domain data representation and balanced sampling
- A VAE maps source and few-shot target data into a shared latent space using concatenated dynamic features and three electrode-level degradation labels.The input contains a 300-dimensional flattened feature vector and a three-dimensional diagnosis label vector.Source: §3.2.1 · p. 6
3.2.2. Physics-constrained joint loss reconstruction
- The VAE loss combines feature and label reconstruction with KL divergence, weighting label reconstruction to keep generated features aligned with physical degradation states.The feature and label reconstruction losses are MSE terms.Source: §3.2.2 · p. 6
3.2.3. Latent space extrapolation and data decoding
- Latent extrapolation varies a scaling factor, mixes the encoded mean with a shuffled mean sequence, and decodes perturbed latent vectors into synthetic samples.The generation procedure applies the reparameterization trick and a standard-deviation scaling coefficient.Source: §3.2.3 · p. 6
3.3. Automated hyperparameter search via Bayesian optimization
- Bayesian optimization minimizes a leakage-free objective balancing lifespan extrapolation and target-domain adaptation, using source-proxy and visible early-life target validation while excluding later-life target data.Source cells are split into base and proxy groups; all remaining 95% of target-cell data is reserved for final independent testing.Source: §3.3 · pp. 7–8
- The search tunes seven parameters controlling target-sample mixing, reconstruction and latent regularization, extrapolation, perturbation, and synthetic-data volume.A Tree-structured Parzen Estimator selects configurations by maximizing expected improvement.Source: §3.3 · p. 8
3.4. Downstream battery health diagnosis model based on deep learning
- The generation framework is independent of the downstream diagnosis architecture; the study evaluates it with an MLP trained on source, early-life target, and generated samples.Evaluation uses only the isolated remaining target-domain test set.Source: §3.4 · p. 8
4. Results and discussion
The results evaluate the framework under cross-domain conditions, from a source-only baseline through physics-constrained generation and ablations. Case 2 is presented in detail, with additional diagnosis results across other cases.
4.1. Baseline diagnosis failure in out-of-distribution scenarios
- The source-only baseline reaches RMSEs of 6.41% for LAMpe, 7.76% for LAMne, and 7.60% for LLI, with entirely negative R2 values.Under Case 2, training uses CC discharging profiles and testing uses unseen synthetic and real-world driving profiles; estimates show severe overestimation and dispersion.Source: §4.1 · p. 9, Figure 6 · p. 9
4.2. Visualization and evaluation of physics-constrained domain generation
- Bayesian optimization searches seven bounded hyperparameters for 300 trials; the best development objective improves progressively before reaching a stable plateau.Mgen is the most influential hyperparameter in Case 2, followed by β; trials with similar objectives can use very different parameter combinations.Source: §4.2 · pp. 9–10
- Generated samples extend from the source distribution toward the target-oriented latent region, filling intermediate space without visually apparent unrestricted extrapolation.The target samples cover only the first 5% of lifespan; generated samples retain broader degradation coverage while enriching density toward the target.Source: §4.2 · p. 10, Figure 8 · p. 11
4.3. Diagnostic performance enhancement driven by physics-constrained domain generation
- In Case 2, the proposed method achieves RMSEs of 1.20% for LAMpe, 2.36% for LAMne, and 1.70% for LLI on the unseen target domain.The corresponding R2 values are 0.93, 0.66, and 0.89; estimates are more closely aligned with the reference line than the baseline.Source: §4.3 · p. 11, Figure 9 · p. 11
- Half-cell OCV and dV/dQ analyses show that estimated degradation trajectories closely follow fitted references and reproduce aging-related shifts in electrode reaction windows.The high-gradient feature shifts from approximately 0.01–0.02 early in life to 0.21–0.23 near end of life, while the mid-peak position moves from approximately 0.35 toward 0.5.Source: §4.3 · pp. 11–12, Figure 10 · p. 12
4.4. Ablation study on data composition and domain generation
- In the Case 2 ablation, early-life target fine-tuning alone leaves LAMpe RMSE at 4.67%, while removing source data causes model collapse.The proposed method outperforms Methods 1–3; Method 4 uses target data across the full lifespan and is described as a leakage-based theoretical upper bound, with LAMpe RMSE of 1.19% versus 1.20% for the proposed method.Source: §4.4 · pp. 12–13, Table 1 · p. 12, Table 2 · p. 13
- Across Cases 3–5, proposed-method RMSEs remain low, although Method 2 slightly outperforms it for LLI in Case 4, at 1.65% versus 1.67%.The proposed method’s LAMpe, LAMne, and LLI RMSEs are 1.86%, 2.78%, and 3.14% in Case 3; 1.48%, 1.80%, and 1.67% in Case 4; and 1.59%, 1.66%, and 1.56% in Case 5.Source: §4.4 · p. 13
- The preferred generation strategy varies by source-data richness: Mgen dominates in data-scarce Case 3, whereas target sampling probability Psample dominates in data-richer Case 5.The authors report more balanced sensitivity across multiple hyperparameters in Case 4 and conclude that generation strategy should be optimized for each scenario.Source: §4.4 · pp. 13–14
5. Conclusion
The framework combines physics-constrained latent extrapolation with Bayesian optimization for cross-domain battery diagnosis. Evaluations span four scenarios, and the conclusion presents domain generation as a plug-and-play data enhancer.
5. Conclusion
- The framework uses VAE-based latent extrapolation and Bayesian optimization to generate data within physically feasible degradation boundaries.Optimization calibrates sample mixing, latent regularization, extrapolation, perturbation, and generation volume.Source: §5 · p. 14
- Across four cross-domain scenarios, the method is reported superior; in Lab2Real, its RMSE values are 1.20%, 2.36%, and 1.70%, respectively.The passage does not identify which degradation modes correspond to these three values.Source: §5 · p. 14
- Ablations indicate that early-life few-shot generation reconstructs the missing target domain, while differential-voltage reconstruction and phase-transition peak analysis demonstrate physical consistency.The peak analysis captures phase-transition evolution over battery lifespan.Source: §5 · p. 14
- Decoupled from downstream diagnosis networks, the domain-generation module is presented as a plug-and-play data enhancer for existing data-driven battery management systems.The paper frames physics-constrained data-distribution enhancement as an alternative pathway to focusing on network architecture or dynamic features.Source: §5 · p. 14
Appendix
Across Cases 2/3/4/5, the method reports average RMSEs for three electrode-level degradation modes.
Appendix
- 1.53% average RMSE is reported for loss of active material in the positive electrode across Cases 2/3/4/5.Source: Appendix · p. 14
- 2.15% average RMSE is reported for loss of active material in the negative electrode across Cases 2/3/4/5.Source: Appendix · p. 14
- 2.02% average RMSE is reported for loss of lithium inventory across Cases 2/3/4/5.Source: Appendix · p. 14