A new self-supervised training trick lets AI models judge lithium-ion battery health using almost no labeled data.
Researchers built a CNN-GRU model - a neural network that combines pattern recognition with sequence memory - and pretrained it on unlabeled battery cycling data using a ranking task: teaching the model to sort battery cycles by their order of degradation rather than requiring exact health labels for each one. After that pretraining, they fine-tuned the model on a sparse labeled dataset, using just 1% of available labels, unevenly distributed across the aging process. On test cells, the fine-tuned model estimated state of health with a mean absolute error of 1.718% and a root mean square error of 2.329%. The team also analyzed how the distribution of those sparse labels affects estimation accuracy.
This matters because state of health estimation is usually the bottleneck for real-world battery monitoring. Collecting large, well-labeled cycling datasets is expensive and slow, since it requires running batteries through full degradation cycles under controlled conditions - something manufacturers and fleet operators rarely have the budget for. A method that gets useful accuracy from 1% labeled data could make health monitoring viable for battery types and use cases that never had enough labeled data to justify it.
Self-supervised pretraining has already reshaped image and language models; batteries, with their slow, expensive-to-label degradation curves, are a natural next target - though real-world validation beyond test cells will decide if this holds up.