EXPLANATION PRESERVING KNOWLEDGE DISTILLATION FOR LIGHTWEIGHT BATTERY STATE OF HEALTH ESTIMATION ON EDGE DEVICES

Deploying battery state of health (SOH) estimators on edge devices requires model compression, but similar predictions do not necessarily imply similar explanations. This paper proposes an explanation preserving knowledge distillation framework that transfers both regression outputs and input attributions from a temporal convolutional teacher to a smaller student. The framework separates supervised estimation, prediction distillation, and signed attribution alignment, followed by integer quantization and a second explanation audit. A reproducible validation protocol is specified for public lithium ion battery aging data, with cell separated testing, chronological evaluation, and training only preprocessing. Evaluation combines SOH error in percentage points, attribution agreement, perturbation based faithfulness, model size, peak memory, and batch one latency on a Raspberry Pi. Analytical examination shows that output matching alone cannot constrain local input sensitivities and that attribution normalization can conceal changes in explanation magnitude. Consequently, the protocol includes signed and magnitude sensitive metrics, a method independent perturbation audit, and explicit checks for nearly zero attribution maps. No model training or hardware measurements are reported. The contribution is a testable method and evaluation design rather than an experimentally established improvement. The framework supports a constrained investigation of whether explanation alignment survives compression and quantization without unacceptable loss of estimation accuracy or edge efficiency.