Flexible Alternation of Structured Pruning and Knowledge Distillation with Recovery-Aware Scheduling for Lightweight Bearing-Fault Diagnosis

Deploying deep bearing-fault-diagnosis models on edge hardware requires combining structured pruning with knowledge distillation (KD). We present PruneDistill-Core, which separates a fixed optimization recipe—cross-entropy plus logit-level KD at constant temperature, learning rate, and KD weight—from a flexible alternating scheduler combining KD-gradient channel importance, an accuracy-gated per-layer pruning budget, a two-stage accuracy guard, recovery fine-tuning with snapshot rollback, and budget checkpointing. A full-system audit uncovered three evaluation pitfalls that had silently inflated earlier comparisons (a mask device-migration bug, recording-level split leakage, and a state-inconsistent rollback); all comparisons here use repaired protocols; recording-level leak-free splits apply to HIT, whereas HUST and SEU use block- and window-level splits. On three public bearing datasets—HIT (recording-level and cross-speed hold-out protocols), HUST (9-class), and SEU (5-class, two operating conditions)—with five seeds per cell (three for the low-data probes), PruneDistill-Core dominates all baselines on the leak-free in-distribution protocol (0.9885±0.0042 accuracy at 0.504 sparsity, 1.52M→604K effective parameters), reaches the deepest sparsity at the highest observed accuracy level in both natural-sparsity regimes on HIT (0.9888 at 0.595; 0.9432 at 0.607), ties the strongest baseline on HUST, and dominates all baselines in SEU’s natural regime. A controlled component-addition study on top of the scheduler—adding the optimizer/delivery machinery (half-stacked), then three extra loss components, recovering the fully stacked design—reduces accuracy and delivered sparsity monotonically (−4.8 pp in total), so the minimal configuration is retained as the final method. Boundary conditions (a 0.89 pp cross-speed gap and low-data collapse at 20% training data) are reported, and central processing unit (CPU) deployment of the delivered fair-regime model reproduces soft-mask accuracy exactly under 32-bit floating-point (fp32) physical compaction (max logit deviation 8.6×10−6) and to within 0.05 pp under 8-bit integer (INT8).