TA-TFN: Task-Adaptive Frequency Decoupling for Generalizable Battery State Estimation across Chemistries

Achieving generalizable battery state estimation across heterogeneous chemistries is fundamentally challenged by severe non-stationarity and the entanglement of short-term electrochemical dynamics with long-term aging-induced drifts in voltage signals. Existing data-driven approaches largely ignore the intrinsic frequency mismatch between state-of-charge (SOC) estimation and state-of-health (SOH) monitoring, leading to degraded robustness under chemistry shifts and complex operating conditions. We propose TA-TFN, a Task-Adaptive Time-Frequency Network that explicitly models task-specific frequency allocation for unified SOC/SOH estimation. TA-TFN introduces an end-to-end differentiable parameterized time-frequency convolution with learnable center frequencies and bandwidths, enabling adaptive decomposition of non-stationary current–voltage sequences without handcrafted features. Building on this representation, a task-adaptive gating mechanism dynamically regulates the contribution of distinct frequency components, directing SOC prediction toward aging-insensitive transient dynamics while enhancing SOH sensitivity to low-frequency degradation patterns. Extensive experiments on the large-scale BatteryLife benchmark, covering multiple chemistries and operating regimes, demonstrate that TA-TFN consistently outperforms existing methods in joint SOC/SOH estimation on heterogeneous and mixed-chemistry datasets, while achieving strong cross-chemistry generalization for SOC estimation, highlighting frequency-aware task decoupling as an effective pathway toward generalizable battery state estimation.