4. Data acquisition and preprocessing for injury UQ

Uncertainty quantification (UQ) for occupant injury is only as credible as the data that feeds it. Three data streams dominate practice: controlled experimental crash tests, computational human body models (HBMs), and real-world field data such as crash databases and event data recorders (EDRs). Each stream carries a distinct uncertainty signature — measurement noise, model-form error, and sampling bias respectively — and each requires its own acquisition protocol and preprocessing chain before it can be fused into a probabilistic injury model. This section walks through what each source actually provides, how to extract usable signals from it, and where the pitfalls lie.

Experimental crash tests: from sled pulses to injury metrics

Laboratory crash tests supply the highest-fidelity ground truth available, but they are expensive and sparse. A typical frontal sled test with a Hybrid III 50th-percentile male ATD yields time histories for head acceleration (three axes), chest deflection, pelvis acceleration, and belt force, sampled at 10–20 kHz. The first preprocessing decision is filtering: injury criteria such as HIC and chest deflection are defined on signals processed per SAE J211, typically a CFC 1000 filter for head acceleration and CFC 180 for chest deflection. Applying the wrong channel filter class silently shifts HIC by several percent, so the filter class must be recorded as metadata alongside every metric.

Beyond filtering, experimental data must be time-aligned to the crash pulse (t = 0 at sled acceleration onset) and channel-calibrated against the transducer sensitivity used during the test. A practical extraction pipeline for a single test looks like this:

# Example: HIC15 extraction from a filtered head acceleration signal
# a_cg: resultant head CG acceleration in g, sampled at 10 kHz, CFC 1000 applied
# t: time vector in seconds, t=0 at pulse onset

def hic15(a_cg, t, window=0.015):
    best = 0.0
    dt = t[1] - t[0]
    n = int(window / dt)
    for i in range(len(a_cg) - n):
        seg = a_cg[i:i+n]
        integral = np.trapz(seg, dx=dt)
        hic = (seg.mean())**2.5 * integral
        best = max(best, hic)
    return best

# Example output for a 48 km/h frontal sled test (illustrative values):
# HIC15 = 412, peak chest deflection = 31.2 mm, peak pelvis accel = 42 g

The values above are illustrative, not measured. The key point is that every derived metric should be stored with its filter class, window length, and calibration reference so that downstream UQ can propagate processing uncertainty separately from physical test-to-test variability. When multiple repeat tests exist (typically 2–3 per configuration), the coefficient of variation across repeats gives a direct empirical estimate of aleatoric uncertainty for that configuration.

Human body models: sampling the model space

Finite-element HBMs such as THUMS, GHBMC, and the SAFER HBM provide dense spatial and temporal output but introduce model-form uncertainty that experiments do not. For UQ, the HBM is treated as a stochastic simulator: material parameters (bone modulus, ligament stiffness, muscle activation), geometry scaling (stature, BMI), and boundary conditions (belt routing, seat friction) are sampled from prior distributions, and each sample produces an injury metric realization.

A defensible sampling design matters more than raw sample count. Latin hypercube sampling over 8–12 active parameters with 200–500 runs is a common starting point for a single load case; screening via Sobol indices or Morris elementary effects first can cut the active set substantially. The preprocessing burden here is output standardization: HBM runs often output in different units, coordinate frames, and mesh resolutions than experimental channels, so a harmonization layer must map model outputs onto the same metric definitions (e.g., HIC computed on the model's head CG node set with the same CFC filter). Failure to harmonize creates artificial variance that swamps the physical uncertainty you are trying to quantify.

Field data: bias, missingness, and linkage

Field data — NASS/CDS, CISS, EDR records, hospital trauma registries — provide the real-world exposure distribution that experiments and models cannot. Their value for UQ is that they capture the joint distribution of crash severity, occupant characteristics, and restraint use. Their cost is selection bias: fatal and tow-away crashes are overrepresented, minor crashes underrepresented, and injury severity is often coded on the AIS scale by trained coders with known inter-rater variability (kappa typically 0.6–0.8 for AIS 2+ injuries).

Preprocessing field data for UQ involves three concrete steps. First, case weighting: NASS provides national weighting factors; ignoring them biases severity distributions toward the sampled strata. Second, missing-data imputation: delta-V is missing in a nontrivial fraction of cases, and mean imputation understates variance — multiple imputation with 20–50 draws is preferable, with the imputation model conditioned on crash type and vehicle class. Third, linkage: matching EDR delta-V to NASS case IDs requires VIN and event-time matching, and mismatches must be audited rather than silently dropped.

The figure below shows how the three streams converge into a single UQ-ready dataset, with the uncertainty each contributes and the harmonization step that aligns them.

Fusing the streams without double-counting uncertainty

The harmonization layer is where most UQ pipelines fail. If the same physical variability enters twice — once as experimental repeat scatter and again as an HBM prior spread calibrated to that same scatter — the posterior injury distribution is over-dispersed. The discipline is to assign each variance component to exactly one stream: test-to-test repeatability to the experimental stream, parameter and model-form uncertainty to the HBM stream, and population exposure and coding variability to the field stream. A simple bookkeeping table makes this auditable:

Once variance components are separated and documented, the dataset is ready for surrogate modeling and Bayesian calibration in later chapters. The practical test of a good acquisition and preprocessing stage is simple: given any injury metric in the final dataset, you should be able to trace it back to its raw channel, its filter class, its calibration reference, its sampling design, and its assigned variance component. If any of those links is missing, the UQ result downstream inherits an unquantified error that no amount of sophisticated inference can recover.