Autonomous End-to-End SOH Prediction Services for Battery Systems via Temporal-Contrastive Representation Learning

Technical Summary: TC-SOH for Autonomous SOH Prediction

Problem Statement
Accurate State of Health (SOH) estimation is critical for lithium-ion battery management systems (BMS) to ensure safety and lifecycle value. However, current data-driven diagnostic services face two primary barriers to scalable industrial deployment:

1. Reliance on Manual Feature Engineering: Existing methods depend on labor-intensive, bespoke extraction of indirect health indicators (e.g., SOC, thermal variations). These methods often discard subtle, non-linear degradation dynamics or arbitrarily expand feature sets, increasing computational overhead.
2. Lack of Interpretability: Deep learning models often function as "black boxes," creating a trust deficit in safety-critical environments. The inability to trace predictions back to physical electrochemical root causes or to verify that learned features align with degradation trajectories limits their adoption.

Methodology: TC-SOH Framework
The authors propose TC-SOH, a modular, plug-and-play service architecture for autonomous, end-to-end SOH prediction. The framework operates on raw operational data from the Constant Current-Constant Voltage (CC-CV) charging phase, bypassing the need for manual feature engineering.

The architecture consists of three integrated components:

1. CNN Module ($f(\cdot)$): A 1D-CNN encoder that transforms raw electrical signals (current, voltage, SOC) from individual CC-CV cycles into high-dimensional latent representations ($z_c$).
2. Temporal-Contrastive (TC) Module ($g(\cdot)$): A sequence transformer that executes a cross-window prediction pretext task. It summarizes a sequence of consecutive cycle representations ($z_{\le t}$) into a context vector ($c_t$) to predict subsequent representations ($z_{t+1}, \dots, z_{t+K}$).
   • Loss Function: The model optimizes a joint objective combining a prediction loss and a contrastive loss ($\mathcal{L}_{TC}$). The contrastive loss maximizes the similarity between predicted and true representations while minimizing similarity with negative samples (other batches), enforcing the learning of sequential degradation dynamics.
   • Augmentation: To prevent encoding bias, augmentation (jitter and scaling) is applied to the learned representations rather than raw signals.
3. Output Module ($o(\cdot)$): A linear layer mapping the refined temporal context to the final SOH value.

Diagnostic Chain for Transparency
To address the "black box" issue, the authors introduce a diagnostic chain to analyze the learned representations:

• Visualization: Using t-SNE to verify smooth degradation trajectories.
• Sensitivity Analysis: Checking responsiveness in healthy and early-decay regimes where SOH changes are subtle.
• Redundancy & Probing: Testing if learned features merely replicate manual descriptors or contain additional SOH-relevant variation.
• Future-SOH Probing & Temporal Shuffling: Evaluating whether the ordered temporal context ($c_t$) predicts future SOH better than unordered statistics or single-cycle readouts.

Key Contributions

1. Plug-and-Play Diagnostic Framework: An end-to-end service that extracts non-linear aging dynamics from routine CC-CV data, significantly reducing reliance on manual feature engineering.
2. Service Trust and Transparency: A diagnostic interpretation chain demonstrating that learned features overlap with physical descriptors while retaining unique predictive variation, and that ordered temporal context is crucial for prediction.
3. Service Efficacy and Robustness: Validation across four large-scale public datasets (XJTU, TJU, HUST, MIT) covering diverse chemistries and operating conditions.

Experimental Results
The model was evaluated against Physics-Informed Neural Networks (PINN) and Manual-MLP baselines under strict battery-level splits (avoiding data leakage).

• Performance: TC-SOH outperformed all baselines. On average, it reduced MAPE by 1.91 times and RMSE by 2.13 times compared to the considered baselines. For instance, on the HUST dataset, the Auto-LSTM variant achieved a MAPE of $0.102 \pm 0.020\%$ and RMSE of $0.0011 \pm 0.0002$.
• Data Efficiency: The model achieves near-full-data performance using only 1–4 batteries for training, indicating reduced data dependence.
• Cross-Domain Adaptation: In transfer learning scenarios (source-only vs. fine-tuning), source-only models failed catastrophically under large domain shifts. However, adapting with just one or two target batteries reduced errors by one to two orders of magnitude. Mixture training (joint source-target) achieved the lowest errors ($<0.012$ RMSE).
• Diagnostic Findings:
  • Learned features showed smoother degradation trajectories than manual features.
  • The model demonstrated superior sensitivity in early-decay regimes (SOH > 0.95).
  • Redundancy analysis showed manual descriptors were largely redundant with learned features, not complementary.
  • Temporal Shuffling: Destroying the order of cycle history significantly degraded future-SOH prediction, confirming that the TC module effectively utilizes ordered temporal context rather than just unordered statistics.

Significance and Claims
The paper claims that TC-SOH provides a practical, end-to-end alternative to hand-crafted feature design for battery health diagnostics. Its significance lies not only in superior predictive accuracy but in its ability to rebuild end-user trust through interpretability. By explicitly diagnosing the learned representation, the authors demonstrate that the model:

• Captures smooth degradation trajectories and early-stage sensitivity.
• Encodes information that overlaps with physical descriptors but retains additional, non-redundant SOH-relevant variation.
• Relies on ordered temporal context to improve subsequent SOH prediction, a capability that unordered or single-cycle features lack.

The authors conclude that this approach addresses the urgent need for scalable, robust, and transparent asset management in modern energy services, though they note future work should connect these findings with controlled electrochemical stress tests.