Knowledge-Aware Modeling with Frequency Adaptive Learning for Battery Health Prognostics

TL;DR

Battery health prognostics remain challenging due to nonlinear degradation and noise. Karma addresses this by coupling frequency-aware learning with a knowledge-based degradation model, using Variational Mode Decomposition to obtain IMFs, a dual-stream CNN-LSTM/BiGRU architecture for low/high-frequency signals, and a double exponential degradation function optimized with Particle Filtering for uncertainty quantification. The approach yields substantial accuracy gains on NASA and CALCE datasets and provides physically meaningful, uncertainty-informed predictions for SoH and RUL. This framework offers robust, generalizable battery prognostics with practical implications for safety and maintenance decisions in diverse energy systems. C(k) = a \exp(b k) + c \exp(d k) captures degradation dynamics, while PF-based parameter estimation aligns data-driven forecasts with empirical knowledge, delivering reliable long-horizon prognostics and explicit prediction intervals.

Abstract

Battery health prognostics are critical for ensuring safety, efficiency, and sustainability in modern energy systems. However, it has been challenging to achieve accurate and robust prognostics due to complex battery degradation behaviors with nonlinearity, noise, capacity regeneration, etc. Existing data-driven models capture temporal degradation features but often lack knowledge guidance, which leads to unreliable long-term health prognostics. To overcome these limitations, we propose Karma, a knowledge-aware model with frequency-adaptive learning for battery capacity estimation and remaining useful life prediction. The model first performs signal decomposition to derive battery signals in different frequency bands. A dual-stream deep learning architecture is developed, where one stream captures long-term low-frequency degradation trends and the other models high-frequency short-term dynamics. Karma regulates the prognostics with knowledge, where battery degradation is modeled as a double exponential function based on empirical studies. Our dual-stream model is used to optimize the parameters of the knowledge with particle filters to ensure physically consistent and reliable prognostics and uncertainty quantification. Experimental study demonstrates Karma's superior performance, achieving average error reductions of 50.6% and 32.6% over state-of-the-art algorithms for battery health prediction on two mainstream datasets, respectively. These results highlight Karma's robustness, generalizability, and potential for safer and more reliable battery management across diverse applications.

Figures (4)

• Figure 1: An illustration of Karma's system architecture. It includes signal decomposition module, a dual-stream frequency adaptive model and a knowledge-based degradation model, which outputs the battery capacity estimation and RUL prediction.
• Figure 2: An illustration of the proposed dual-stream frequency adaptive learning model for battery capacity estimation in a given forecast horizon. The two parallel streams are based on CNN-LSTM and BiGRU for low- and high-frequency IMFs, respectively. The streams are fused into one before final processing including self-attention and model output.
• Figure 3: The degradation trajectories of the first NASA battery and the first CALCE battery in Figs. \ref{['fig:comp-soh-nasa-curve']} and \ref{['fig:comp-soh-cs-curve']}, respectively, with capacity estimations from Karma, CNN-LSTM and BiGRU. The corresponding prediction errors are visualized in Figs. \ref{['fig:comp-soh-nasa-err']} and \ref{['fig:comp-soh-cs-err']}, respectively. Karma aligns with the ground-truth the best with minimized prediction errors.
• Figure 4: Karma's battery capacity estimation with 95% CI uncertainty quantification and RUL prediction for NASA batteries at SP 50 in Fig. \ref{['fig:uncertainty-nasa']} and CALCE batteries at SP 100 in Fig. \ref{['fig:uncertainty-calce']}.