When Smaller Wins: Dual-Stage Distillation and Pareto-Guided Compression of Liquid Neural Networks for Edge Battery Prognostics

The paper introduces DLNet, a framework utilizing dual-stage distillation of liquid neural networks for compact battery health prediction models suitable for edge deployment. By reformulating liquid dynamics through Euler discretization and employing Pareto-guided selection, DLNet reduces model size from 616 kB to 94 kB (84.7% reduction) while achieving a low prediction error of 0.0066, outperforming the larger teacher model by 15.4%. This work demonstrates that smaller models can effectively meet stringent on-device constraints, with implications for various industrial analytics tasks requiring efficient deployment.