2. Fundamentals of occupant restraint CAE-test correlation

CORA (Correlation) rating is only meaningful when the underlying signals are physically comparable. Before any objective rating can be computed, the CAE analyst and the test engineer must agree on what is being compared, how each signal is filtered and aligned, and which metrics quantify the agreement. This chapter establishes that foundation: the canonical crash-safety channels, the structure of their time histories, and the metric families used later in the CORA scoring workflow.

Key signals in an occupant restraint load case

A frontal occupant restraint model (belted Hybrid III 50th percentile, 56 km/h full-frontal rigid barrier, or a Euro NCAP MPDB variant) produces a small set of channels that dominate the correlation outcome. They fall into three groups:

  • Vehicle-level kinematics: barrier force F_barrier, vehicle CG longitudinal acceleration a_x, and the sled/buck pulse used to drive the occupant simulation. These define the boundary condition; if the pulse is wrong, every downstream channel is wrong.
  • Restraint system response: belt force at the retractor and at the shoulder/lap anchor, belt payout (spool-out), retractor locking time, pretensioner stroke, load-limiter force level, and airbag internal pressure plus venting behaviour.
  • Occupant response: head resultant acceleration a_head, chest resultant acceleration a_chest, chest deflection D_chest, pelvis acceleration, upper-neck force/moment, and occupant kinematics (head CG trajectory, HIC36, chest 3 ms clip).

Each channel has a characteristic frequency content. Belt force and chest deflection are dominated by low-frequency structural response (roughly 0–50 Hz), while head acceleration and barrier force contain meaningful energy up to 300–1000 Hz. This difference is the single most common source of spurious correlation failures: applying one filter to all channels either smears the head peak or leaves high-frequency noise on the belt trace.

Anatomy of a time history

Every crash channel can be decomposed into four phases that must be compared separately, because a single global metric hides phase-specific errors:

  1. Pre-trigger / quiet phase (t < 0): baseline noise and any pretensioner firing artefacts. Used to verify zero-offset and sensor polarity.
  2. Ramp / loading phase: from pulse onset to the first significant peak. Governed by restraint stiffness, pretensioner timing, and airbag deployment.
  3. Peak / interaction phase: the occupant–restraint interaction window where head and chest peaks occur. This is where HIC and chest deflection are decided.
  4. Rebound / unloading phase: belt spool-in, airbag deflation, and occupant rebound. Often neglected, yet it drives the second chest peak and neck loading.

Two preprocessing steps must be fixed before comparison. First, time-zero alignment: the test pulse and CAE pulse are cross-correlated to find the lag that maximises their overlap, and that single lag is applied to all channels. Second, filtering: apply the channel-appropriate filter (e.g. SAE J211 CFC 60 for belt force and chest deflection, CFC 1000 for head acceleration, CFC 180 for vehicle acceleration) to both curves identically. Never filter one side only.

Record the chosen filter class, time-zero lag, and any channel scaling in the correlation log. CORA scores are only reproducible if these preprocessing choices are documented alongside the curves.

Evaluation metrics and what each one captures

CORA combines two complementary metric families, each evaluated over a defined evaluation window:

Metric familyWhat it measuresTypical sensitivity
Magnitude / amplitude (corridor)Pointwise deviation of test and CAE curves, normalised by the test curve's rangePeak height errors, offset errors, sustained level differences
Phase / progressionTime lag between corresponding features (cross-correlation of the two curves)Late pretensioner, early airbag, shifted pulse

In addition, engineers routinely track derived scalar metrics that are physically meaningful to the restraint design and are reported alongside the CORA score:

  • Peak value ratio P_CAE / P_test for head acceleration, chest acceleration, chest deflection, and belt force.
  • Time-to-peak difference Δt_peak in milliseconds.
  • Impulse / integral difference over the interaction window, which captures energy absorption mismatch even when peaks agree.
  • HIC36 and chest 3 ms clip as injury-criterion-level comparisons, since a small curve deviation can flip an injury metric across a threshold.

The CORA rating itself is a weighted combination of the magnitude and phase sub-scores, mapped to a 0–1 scale and then to a grade (e.g. "excellent", "good", "acceptable", "poor"). The exact weighting and corridor definition are configuration parameters; what matters at this stage is that the analyst understands which physical error each sub-score penalises.

Worked example: reading a correlation result

Consider a belted 50th percentile frontal model compared against a sled test. Suppose the belt force trace shows a CAE peak of 4.6 kN versus a test peak of 4.2 kN, with the CAE peak arriving 8 ms late, while head acceleration peaks agree within 3% and 2 ms. The magnitude sub-score for belt force will be penalised by the ~10% peak overshoot; the phase sub-score will be penalised by the 8 ms lag. The head channel scores well on both. A naive average would call this "good", but the belt-force lag is a real physical signal: it usually points to a pretensioner firing later in CAE than in test, or to a load-limiter force level set too high. The correct response is to inspect the pretensioner trigger time and the load-limiter table, not to tune the belt stiffness to chase the peak.

Do not adjust model parameters to improve a CORA score without identifying the physical cause. A high score obtained by mistuning stiffness or timing is a false correlation and will fail on the next load case.

Signal chain from test to rating

The diagram below summarises how raw channels become a CORA rating, and where preprocessing decisions enter the chain.

Test channels belt force, a_head, a_chest, D_chest, F_barrier, pulse CFC-filtered, time-zeroed CAE channels same channel set, same filter class, same time-zero lag from sled/buck pulse Preprocessing • filter (CFC class) • time-zero alignment • offset / polarity check • evaluation window logged for reproducibility Magnitude metric pointwise corridor normalised by test range penalises peak & offset Phase metric cross-correlation lag between feature pairs penalises timing shift CORA rating weighted sub-scores → 0–1 scale → grade per channel, then aggregated Preprocessing choices (filter class, time-zero lag, evaluation window) propagate into every downstream metric — document them with the curves.

Practical checklist before running a correlation

The following items should be confirmed for every channel pair before any CORA score is trusted:

  • Both curves use the same physical unit and sign convention (e.g. chest deflection positive in compression, head acceleration positive forward).
  • Both curves are filtered with the same CFC class appropriate to the channel.
  • Time-zero is defined by the same event (pulse onset or barrier contact) on both sides, and the residual lag is recorded.
  • The evaluation window covers the loading and interaction phases, and is stated explicitly (e.g. 0–120 ms).
  • Derived scalars (peak ratio, Δt_peak, HIC36, chest 3 ms clip) are computed and reported alongside the CORA score.
  • Any channel excluded from the rating is justified in the correlation log.

With these conventions fixed, the CORA score becomes a diagnostic instrument rather than a single number: each sub-score points to a specific physical discrepancy, and the analyst can trace it back to the pretensioner, load limiter, airbag venting, or pulse input that produced it.

← Previous 📖 Contents Next →