Automated Detection of Tribologically Relevant Brake Torque Plateaus: A Two-Stage Approach for Flywheel Dynamometer Testing

Reliable identification of the tribologically relevant braking phase in torque signals recorded on flywheel dynamometers is a prerequisite for quantitative friction analysis and data-driven modeling of dry-running friction brakes. We define brake torque plateaus as intervals with quasi-constant surface pressure and appreciable sliding velocity in which fading or drift of the coefficient of friction is explicitly admissible, while rise and decay ramps dominated by actuator dynamics are excluded. To automate this extraction across large industrial data sets, we propose a two-stage detection algorithm that sequentially narrows the search space using physics-based amplitude, gradient, and stability criteria, complemented by a Pruned Exact Linear Time (PELT)-based fallback for difficult cycles. Evaluation on 10,386 brake cycles, including 275 expert-annotated ground-truth cycles validated by a second independent expert, shows that the proposed method reaches 95% of the inter-annotator agreement ceiling on 75 held-out cycles, achieves a median Intersection-over-Union of 0.893 (11 percentage points above the strongest baseline), and a mean quality score of 9.18/10 across all cycles at under 1 ms per cycle (signals averaging 951 samples), outperforming six baseline configurations in both detection quality and runtime.