2. Crash test data fundamentals

Correlation between a PAM-CRASH simulation and a physical crash test is only as trustworthy as the measurement chain that produced the reference data. Before any model calibration begins, the engineer must understand what each sensor actually measures, how the data acquisition system (DAS) samples and stores it, which filters have already been applied, and how the raw channels are transformed into the signals that will be overlaid against simulation output. This section establishes that foundation.

Sensor Types and What They Physically Measure

Full-scale vehicle crash tests rely on a small set of transducer families, each with distinct physical principles and limitations.

• Accelerometers (piezoresistive or capacitive): Measure acceleration along one or more axes. They are mounted on the vehicle body, sled, or crash barrier. Typical full-scale ranges are ±500 g to ±2000 g for structural locations and ±50 g to ±250 g for occupant-compartment or sled measurements. Their usable bandwidth is limited by mounting resonance, often 5–10 kHz.
• Load cells: Measure force directly, commonly in the barrier face, seat belt anchorages, or on a pedestrian impactor. A barrier load cell wall may contain hundreds of individual cells, each sampled independently.
• String potentiometers and displacement transducers: Measure linear displacement, for example steering column intrusion or pedal travel. They are relatively low-bandwidth (typically below 100 Hz) and prone to mounting compliance errors.
• Strain gauges: Measure local surface strain on structural members. They require temperature compensation and careful bridge wiring.
• Angular rate sensors and gyroscopes: Measure rotational velocity, increasingly common for capturing yaw and pitch during oblique impacts.
• High-speed cameras and DIC: Provide full-field displacement and deformation data, often used to validate simulation contour plots rather than point signals.

The critical point for correlation is that each sensor has a defined measurement location, axis convention, and polarity. A simulation node must be placed at the same physical point and oriented consistently. A sign error in one axis will produce a mirror-image curve that no amount of material tuning can fix.

Data Acquisition: Sampling, Resolution, and Anti-Alias Protection

Crash DAS units sample all channels simultaneously at a common rate, typically 10 kHz to 20 kHz for full-scale tests, with 16-bit or 24-bit resolution. The sampling theorem requires that the sample rate exceed twice the highest frequency of interest; in practice, engineers target at least 10 samples per cycle of the highest meaningful frequency. For a 1 kHz structural oscillation, 10 kHz sampling gives adequate resolution.

Before the analog-to-digital converter, every channel passes through an analog anti-alias filter, usually a low-pass Butterworth with a cutoff at roughly one-third to one-half of the sample rate. This prevents high-frequency content from folding back into the recorded band. The engineer must know this cutoff frequency because it defines the maximum bandwidth that can ever be recovered from the data.

Filtering: Channel Frequency Class and Its Consequences

Raw accelerometer signals contain high-frequency noise from sensor resonance, cable vibration, and structural ringing that is not relevant to occupant injury or structural deformation. SAE J211 defines Channel Frequency Class (CFC) filters that standardize how this content is removed:

The filter is applied in post-processing, not during acquisition. This means the same raw channel can be filtered at different CFC levels for different purposes. For correlation, the simulation output must be filtered with the same CFC and the same filter implementation (zero-phase, forward-backward) as the test data. Comparing an unfiltered simulation curve against a CFC 60 test curve is a common and serious error.

Preprocessing Steps for Correlation-Ready Data

Between the raw DAS file and a curve that can be overlaid on simulation output, several transformations are required. The following sequence reflects standard practice in automotive crash correlation.

1. Zero offset correction: Determine the pre-trigger mean value over a quiet window (e.g., 50 ms before impact) and subtract it from the entire channel. This removes DC bias from sensor electronics.
2. Time-zero alignment: Define t = 0 at a physically meaningful event, typically the first contact between vehicle and barrier. This may come from a contact switch, a high-speed camera frame, or a threshold crossing on a barrier load cell. All channels and the simulation must share this reference.
3. Polarity and axis verification: Confirm that positive acceleration in the test corresponds to positive acceleration in the simulation coordinate system. Document any sign flips explicitly.
4. Filtering: Apply the appropriate CFC filter using a zero-phase implementation to avoid introducing artificial time shifts.
5. Integration (when required): Derive velocity and displacement from acceleration using cumulative trapezoidal integration. Apply the same integration scheme to both test and simulation data. Drift in integrated displacement is a common artifact and should be noted, not hidden.
6. Resampling: If test and simulation are sampled at different rates, resample both to a common time base before comparison. Use a consistent interpolation method.

From Raw Channels to Correlation Metrics

Once the test channels are preprocessed, they are paired with corresponding simulation outputs. The pairing is defined by a correlation map: a table that links each test sensor to a specific simulation node or element output, with its axis, CFC class, and time window of interest. This map is the contract between the test and simulation worlds.

Typical quantitative metrics include peak value difference, time-of-peak difference, and curve-matching scores such as the CORrelation and Analysis (CORA) rating, which combines phase, magnitude, and shape components into a single score. These metrics are only meaningful when the preprocessing is consistent. A CORA score computed on mismatched filters is worse than useless—it gives false confidence or false alarm.

Practical Example: Preparing a Body Acceleration Channel

Consider a full-scale frontal impact test with a body-mounted accelerometer at the B-pillar. The DAS records at 20 kHz with a 5 kHz anti-alias filter. The raw channel shows a pre-trigger offset of +0.8 g and a noisy signal with visible resonance above 2 kHz.

Preprocessing proceeds as follows. First, the pre-trigger window from −100 ms to −10 ms is averaged, yielding +0.8 g, which is subtracted from the entire channel. Second, time-zero is set from the barrier contact switch, which triggers at the first electrical contact between the vehicle bumper and the barrier face. Third, the channel is filtered with a CFC 180 zero-phase Butterworth filter (cutoff 300 Hz). Fourth, the filtered acceleration is integrated to velocity and displacement using the trapezoidal rule, with the initial velocity set to the measured impact speed (e.g., 56 km/h or 15.56 m/s).

The resulting displacement curve should return to a plausible permanent deformation value after the crash pulse ends. If the integrated displacement drifts unrealistically—for example, continuing to increase after the vehicle has come to rest—this indicates a residual DC offset or an integration error. The fix is to re-examine the zero-offset window or apply a linear trend removal to the acceleration before integration.

On the simulation side, the corresponding node output must be extracted at the same physical location, filtered with the same CFC 180 implementation, and integrated with the same trapezoidal scheme. Only then can the two curves be overlaid and a CORA score computed. If the test curve peaks at 42 g and 38 ms while the simulation peaks at 48 g and 35 ms, the correlation map should record these differences explicitly, and the calibration effort should focus on the structural stiffness or mass distribution that controls the pulse shape—not on the preprocessing, which is now known to be consistent.