4. Physical test data for corridor development

A validation corridor is only as trustworthy as the physical data used to construct it. Before any lower or upper bound is drawn around a response channel, the engineer must answer a deceptively simple question: how much of the spread in the test data is real physical variation, and how much is an artifact of the measurement and processing chain? Corridor bounds that are too tight will flag acceptable CAE models as failures; bounds that are too loose will pass models that do not represent the occupant response. This chapter examines the four sources that dominate corridor uncertainty — test repeatability, instrumentation uncertainty, signal processing, and filtering — and shows how each one propagates into the corridor envelope.

Separating Repeatability from Reproducibility

Repeatability describes scatter when the same sled, the same dummy, the same instrumentation, and the same procedure are used across consecutive runs. Reproducibility describes scatter when any of those elements change: a different dummy serial number, a different lab, a different fixture build, or a different day. For corridor development, repeatability sets the minimum achievable corridor width, while reproducibility sets the width that a CAE model must actually satisfy, because the model is compared against the population of tests, not a single run.

A practical way to quantify this is to run a dedicated repeatability series before the main test matrix. For a frontal sled pulse at 50 km/h with a Hybrid III 50th percentile occupant, a typical series might be five runs at identical input conditions. For each channel of interest — head resultant acceleration, chest deflection, femur load — compute the mean and standard deviation at each time step, then form the coefficient of variation at the peak. As an illustrative example, if chest deflection peaks average 42.0 mm with a standard deviation of 1.3 mm across five runs, the peak CV is about 3.1%. That number is a floor: no corridor should be narrower than roughly ±2σ of the repeatability distribution at the peak, or it will reject the test population itself.

Instrumentation Uncertainty and Channel-Level Error Budgets

Every measurement channel carries a documented uncertainty that must be folded into the corridor. The dominant contributors are sensor nonlinearity and hysteresis, shunt calibration error, DAQ resolution and gain error, and dummy-specific effects such as mass properties and joint friction. Manufacturer specifications give per-channel bounds, but the effective uncertainty at the corridor level is a combination, not a single number.

Build a channel-level error budget before setting bounds. For a uniaxial accelerometer with ±2% full-scale nonlinearity on a 500 g range, the sensor contribution is ±10 g. Add ±1% shunt calibration (±5 g), ±0.5% DAQ gain (±2.5 g), and a dummy mass-property contribution estimated at ±3 g from build sheets. Combining in root-sum-square gives approximately ±11.8 g. That figure is the instrumentation floor for that channel; the corridor half-width at the peak must be at least this large before any test-to-test scatter is added.

Signal Processing: Resampling, Alignment, and Differentiation

Raw channels rarely arrive in a form suitable for corridor construction. Three processing steps matter most. First, resampling: test data at 10 kHz and CAE output at 1 kHz must be brought to a common time base. Resample with an anti-aliasing filter, not by simple decimation, or high-frequency content will fold back into the passband and corrupt peak values. Second, time alignment: corridors are built in a common time reference, usually tied to the sled pulse onset (t = 0 at the 1 g rising edge of the sled acceleration). Misalignment of even 2–3 ms shifts peaks and artificially widens the envelope. Third, differentiation: deriving velocity or displacement from acceleration amplifies noise, so integrate rather than differentiate where possible, and apply the same integration scheme to test and CAE data.

# Illustrative processing sequence for a chest acceleration channel
# 1. Anti-alias and resample to common 1 kHz grid
y = resample(raw, fs_in=10000, fs_out=1000, anti_alias=True)
# 2. Align to sled pulse onset (1 g crossing on rising edge)
t0 = find_first_crossing(sled_accel, threshold=1.0 * g)
y_aligned = shift(y, -t0)
# 3. Apply CFC filter (see next section) before peak extraction
y_filt = cfc_filter(y_aligned, class_=180)
peak = max(abs(y_filt))

Each of these steps changes the peak value. A 2 ms misalignment on a sharp chest acceleration peak can shift the measured maximum by several percent; resampling without anti-aliasing can inflate it further. The corridor must therefore be built from data processed with exactly the same chain that will be applied to CAE results — otherwise the comparison is not like-for-like.

Filtering Effects and Channel Frequency Class

Occupant signals are conventionally filtered to a Channel Frequency Class (CFC) per SAE J211. The choice of CFC is not cosmetic: it directly sets the corridor width at high-frequency features. A CFC 1000 filter preserves sharp transients and impact spikes; a CFC 180 filter smooths them, lowering peaks and narrowing the apparent scatter. If test data are filtered at CFC 180 but CAE output is compared unfiltered, the model will appear to overshoot on every spike.

The practical rule is to fix the filter class per channel in the corridor specification and apply it identically to test and simulation. For head resultant acceleration, CFC 1000 is standard; for chest acceleration and deflection, CFC 180 is common; for femur load, CFC 600. Document the filter class, the filter implementation (analog prototype, phase treatment), and the zero-phase vs. causal choice, because a causal filter introduces a phase lag that shifts peaks in time and can widen the corridor if test and CAE use different implementations.

Propagating Uncertainty into Corridor Bounds

With repeatability, instrumentation, processing, and filtering characterized, the corridor half-width at each time step can be assembled as a combined uncertainty. A defensible construction is to take the test population mean μ(t), add the repeatability band k·σ(t) (k typically 2), and then add the instrumentation and processing contributions in root-sum-square where they are independent. The result is a time-varying envelope that reflects the actual evidence behind the data rather than an arbitrary percentage.

The figure above shows how the four evidence streams combine into a time-varying envelope. Note that the band is not a constant percentage of the mean: it widens where repeatability scatter is large (typically around sharp transients) and narrows in the quasi-static plateau. This is the correct behavior — a corridor that is uniformly ±10% everywhere is almost always either too tight at the peaks or too loose in the plateau.

Practical Checks Before Freezing the Corridor

• Filter parity: confirm test and CAE channels use the same CFC class and the same filter implementation; verify by comparing a synthetic step response.
• Alignment audit: overlay all repeat runs on the common time base and check that peaks coincide within one sample; misalignment shows up as a smeared peak in the population.
• Error budget sign-off: document the per-channel instrumentation budget and confirm the corridor half-width at the peak is at least the RSS of instrumentation and processing contributions.
• Population size: with fewer than three repeat runs, σ is poorly estimated; treat the corridor as provisional and widen it conservatively.
• Channel-specific review: do not apply a single global corridor rule across head, chest, and femur channels — their repeatability and filtering behavior differ substantially.

When these checks are complete, the corridor is traceable: every unit of width can be attributed to a measured source, and any future disagreement between CAE and test can be triaged against a known uncertainty budget rather than argued qualitatively. That traceability is what separates a credibility corridor from a hand-drawn tolerance band.