Vehicle-Conditional Split-Conformal Calibration for Risk-Budgeted Sub-Second Proxy-Triggered Vehicle Instability...

Area of Science:

• Vehicle dynamics and control systems
• Machine learning for safety-critical applications
• Real-time monitoring and warning systems

Background:

• Emergency maneuvers can rapidly induce severe vehicle instability.
• Existing systems lack sub-second warning capabilities with auditable false-alarm budgets.
• Need for reliable, last-moment alerts to prevent accidents.

Purpose of the Study:

• To develop and calibrate a lightweight hazard scorer for imminent vehicle instability.
• To provide a physics-defined proxy-triggered warning system with controlled false positive rates.
• To enable per-vehicle risk budgeting through vehicle-conditioned calibration.

Main Methods:

• Utilized a 0.1s past-only slice of onboard signals to predict a physics-defined instability proxy trigger within τ=0.2s.
• Employed split-conformal calibration on negative data slices to establish a slice-level false-alarm budget (α).
• Implemented vehicle-conditioned (Mondrian) thresholds to address fleet heterogeneity.

Main Results:

• The hazard scorer achieved high performance metrics: AUPRC ≈0.251, AUROC ≈0.986, and ECE ≈0.034.
• Calibrated system at α=5% maintained a slice-level false positive rate near the budget while achieving high true positive rate (≈0.982).
• Demonstrated effective fleet heterogeneity management via Mondrian calibration.

Conclusions:

• A transparent, risk-budgeted monitoring primitive for last-moment vehicle stability warning is presented.
• The system offers a calibrated warning for a surrogate instability event, not direct crash prediction.
• The approach ensures finite-sample, one-sided control of the marginal slice-level false positive rate under exchangeability assumptions.