2. Sensing and data acquisition for airbag pressure and deformation

Predictive models for airbag pressure and deformation are only as trustworthy as the measurements used to train and validate them. This chapter examines the three sensing modalities that form the backbone of a typical airbag test campaign: pressure transducers for capturing the internal inflation history, high-speed cameras for recording the visible deployment kinematics, and digital image correlation (DIC) for converting those images into quantitative displacement and strain fields. The goal is not merely to list instruments, but to establish how each channel is configured, synchronized, and conditioned so that the resulting dataset is physically consistent and suitable for machine-learning pipelines.

Pressure Transducers: Capturing the Inflation History

Airbag inflation is a fast, highly transient event. A driver-side airbag may reach peak internal pressure within 20–40 ms of the trigger signal, with the pressure curve exhibiting a sharp rise, a brief plateau, and a decay phase as gas vents through the fabric or relief vents. Capturing this curve faithfully requires transducers with sufficient bandwidth and appropriate mounting.

Piezoresistive transducers are the most common choice because they offer good linearity, a wide dynamic range, and a flat frequency response well beyond the few kilohertz needed here. A typical sensor might have a 0–500 kPa range, ±0.5% full-scale accuracy, and a natural frequency above 50 kHz. The critical practical detail is mounting location: the transducer must be flush with the inner surface of the inflator housing or the module cavity, with a short, rigid port to avoid acoustic resonance in the connecting passage. A long flexible tube between the bag interior and the sensor acts as a low-pass filter and can distort the peak by 10–20%, which is unacceptable when the peak pressure is a key label for the model.

Sampling rate matters as much as bandwidth. A rule of thumb is to sample at least ten times the highest frequency of interest; for a 2 kHz pressure event, a 25–50 kHz sample rate is reasonable. The following example shows a minimal acquisition configuration for a single pressure channel:

# Example pressure acquisition configuration (illustrative values)
sensor:        piezoresistive, 0-500 kPa, flush-mounted
sample_rate:   50 kHz
anti_alias:    10 kHz low-pass (hardware)
trigger:       rising edge on inflator current, threshold 2 A
pre_trigger:   5 ms
record_length: 200 ms
calibration:   two-point, 0 kPa and 300 kPa reference

Always record a pre-trigger window so the baseline and any pre-ignition noise are visible. During post-processing, verify that the pressure trace returns to ambient within the record window; a non-zero tail usually indicates a blocked vent or a sensor offset that was not zeroed before the test.

High-Speed Imaging: Framing the Deployment

High-speed cameras provide the spatial context that pressure sensors cannot. To resolve fabric motion during deployment, frame rates of 1,000–10,000 fps are typical, with exposure times short enough to freeze motion—often 10–50 µs. At 5,000 fps and a 1,024 × 1,024 pixel sensor, the field of view must be sized so that the fully deployed bag fills roughly 70–80% of the frame, leaving margin for out-of-plane motion.

Lighting is the most underappreciated variable. Because exposure times are short, continuous illumination is usually insufficient; high-output LED arrays or pulsed metal-halide sources are used, and they must be flicker-free at the chosen frame rate. A common failure mode is intensity beating, where the lamp's ripple frequency interacts with the frame rate and produces a periodic brightness modulation that corrupts DIC correlation. Setting the camera to external sync with the light source, or using DC-driven LEDs, eliminates this artifact.

Camera placement should be perpendicular to the primary deformation plane. For a passenger-side bag deploying toward the instrument panel, a front view captures the forward expansion, while a side view captures the vertical spread. If stereo DIC is planned, two cameras are mounted on a rigid bar with a calibrated baseline, typically 300–600 mm, and their relative orientation is fixed before the test.

Digital Image Correlation: From Pixels to Displacement Fields

DIC works by tracking the motion of a random speckle pattern applied to the airbag fabric. The pattern must be fine enough that each subset—typically 15–31 pixels square—contains several distinct speckles, yet coarse enough that the pattern survives the large strains of deployment. A practical approach is to airbrush a matte black-and-white speckle onto the fabric before folding, then cure it so it does not flake during inflation.

The correlation process divides each image into subsets and finds the displacement that maximizes the normalized cross-correlation between the reference and deformed images. The output is a displacement vector at each subset center, from which strain is computed by differentiating the displacement field. Key parameters include subset size, step size (the spacing between subset centers), and the strain filter window. A smaller subset improves spatial resolution but increases noise; a step size of 5–10 pixels is a common compromise.

For airbag applications, the largest challenge is large deformation and out-of-plane motion. A planar DIC setup assumes the surface remains in the camera plane; when the bag bulges toward the camera, the apparent displacement is contaminated by perspective scaling. Stereo DIC, using two calibrated views, reconstructs the three-dimensional surface and removes this error. The calibration target should be imaged at multiple orientations before the test, and the reprojection error should be checked—values below 0.1 pixels indicate a sound calibration.

Synchronizing the Channels

The three modalities must share a common time base. The standard practice is to use the camera's trigger output as the master clock: the inflator current signal fires the camera, and the pressure DAQ is triggered from the camera's exposure-active signal. This guarantees that frame n corresponds to a known timestamp relative to ignition. A residual timing uncertainty of one frame interval (e.g., 200 µs at 5,000 fps) is usually acceptable, but it should be recorded and, if possible, reduced by using a hardware trigger rather than software timing.

The figure below summarizes the acquisition chain and the synchronization path from trigger to synchronized dataset.

Data Conditioning and Quality Checks

Raw channels rarely go straight into a model. Pressure traces should be zeroed against the pre-trigger baseline, then low-pass filtered at a cutoff below the sensor's anti-alias limit to remove electrical noise without attenuating the physical rise. DIC displacement fields require outlier removal—typically a local median filter over a 3 × 3 subset neighborhood—and a check that the strain field is continuous across the fabric surface. Discontinuities often indicate a lost speckle patch or a specular reflection, both of which should be flagged rather than silently interpolated.

A practical quality gate before a dataset is accepted for training includes: peak pressure within the sensor's calibrated range, no saturation in any camera frame, DIC correlation coefficient above 0.95 in the region of interest, and a static zero-strain check below 50 µε. If any of these fail, the test is repeated rather than corrected in post-processing, because model errors introduced by bad labels are difficult to detect later.

With synchronized, conditioned channels in hand, the dataset is ready for the feature engineering and model training stages that follow. The discipline established here—flush mounting, flicker-free lighting, calibrated stereo DIC, and a single master clock—determines whether the pressure and deformation labels are physically meaningful or merely plausible-looking numbers.