New AI method cuts cost of battery health checks without sacrificing accuracy
Researchers have developed a neural network that estimates lithium-ion battery degradation faster and cheaper than existing methods, eliminating the need for expensive data collection across different operating conditions. The breakthrough could accelerate adoption of battery management systems in electric vehicles and grid storage, reducing downtime and maintenance costs.
Originaltitel: ONET: Operator network for randomized and robust battery health estimation using operation condition and cycling data matching
Robust state-of-health (SOH) estimation is critical for ensuring the reliability and safety of lithium-ion batteries. However, robust estimation across diverse operating conditions often necessitates computationally intensive strategies, such as transfer learning and associated data acquisition cost and time, challenging their practical deployment. This study presents an operator learning neural network (ONET) enabled SOH estimation, which leverages the conditional dependencies between operation conditions and resulted battery operating data to help explicit utilization of operation conditions as an additional information source for state estimation, which is currently underinvestigated. The ONET consists of a trunk network captures the underlying degradation patterns from partial charging segments, while a branch network model the impact of operational conditions on these patterns. An attention-based multi-feature fusion (AMFF) was proposed to produce operation-condition dependable SOH estimates, which adaptively combines outputs of the trunk and branch networks by dynamically learning attention weights to assess their relative importance. Validated on a two chemistries dataset (NCA and NCM battery chemistry) under diverse temperatures (25, 35, and 45 °C ) and charging rates (0.25, 0.5, and 1C), the proposed ONET exhibits highly accurate SOH estimation using 0.2 V partial charging segments, achieving performance to state-of-the-art models with a mean absolute error (MAE) of 0.438%, a mean absolute percentage error (MAPE) of 0.496%, and an improved coefficient of determination (R 2 ) of 0.991 across different dataset splits. Practically, ONET is lightweight with a memory size of 230 KB, an 88.7% reduction in memory compared to a reference model, making it suitable for deployment on resource-constrained edge devices. Moreover, AMFF-ONET demonstrates robustness against noise injections, maintaining accuracy (R 2 > 0.75) under a low signal-to-noise ratio up to 30 dB. Broadly, proposed ONET demonstrates advantages of learning inherited conditional matching between battery operation conditions and resulted cycling data, providing robust while lightweight solutions to health state monitoring and evaluation of critical energy infrastructures. • ONET-based framework mapping degradations to the operational conditions. • Achieving end-to-end SOH estimation using only short 0.2V charging segments. • Integrating attention mechanism for robust performance across diverse conditions. • Compact model (<230 KB) enabling efficient edge deployment in battery management systems. • Evaluated on 121 cells with two chemistries under varied operating conditions.