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

Accurate estimation of the state of health (SOH) is critical for the safe and efficient operation of lithium-ion batteries in systems such as electric vehicles and grid storage. Traditional data-driven models rely heavily on large volumes of high-quality labeled cycling data across the entire battery lifespan. However, obtaining these labels in real-world applications is expensive, time-consuming, and results in severe label sparsity, particularly in later aging stages. This paper introduces a self-supervised learning (SSL) framework designed to leverage abundant unlabeled operational data to learn aging-consistent representations before fine-tuning on sparsely labeled datasets.

To overcome the limitations of standard reconstruction-based SSL approaches that may not align with actual battery degradation pathways, the authors propose a ranking-based pretext task. Because battery aging is an inherently ordered process, the model is trained using a cycle-order ranking objective on constant current (CC) charging curves. This forces the network to capture relative aging trajectories and learn degradation-aligned representations from unlabeled cycling data without requiring manual annotations.

The core network architecture integrates convolutional neural networks (CNNs) and gated recurrent units (GRUs) into a unified CNN-GRU model. The CNN blocks use local receptive fields and same padding to extract local patterns from structured 1D charging curves, while the GRU layer sequentially integrates these features to model temporal dependencies across cycles. During the pretraining phase, the network utilizes a ranking head to output aging scores. Once pretraining is complete, the encoder parameters are transferred to the downstream task, and the model is fine-tuned using an extremely limited amount of labeled data to perform SOH regression.