3. Data acquisition for crash deformation

Deep learning models for crash deformation prediction are only as good as the data they are trained on. A model that predicts intrusion depth, deformation pattern, or residual crush from a given impact condition must learn the mapping between loading parameters and structural response. That mapping is established empirically through three complementary experimental tiers: full-scale vehicle crash tests, sled tests, and component-level experiments. Each tier occupies a distinct position on the trade-off between physical fidelity and controllability, and a well-constructed dataset for deformation prediction typically draws from all three.

Full-Scale Crash Tests: The Reference Standard

Full-scale crash tests subject a complete vehicle to a controlled impact and represent the highest-fidelity data source. The test article is a production or prototype vehicle instrumented with accelerometers, load cells, string potentiometers, and high-speed cameras. Standard configurations include frontal rigid barrier at 56 km/h (FMVSS 208 / Euro NCAP offset deformable barrier), side pole impact at 32 km/h, and rear moving deformable barrier at 80 km/h. For deformation prediction specifically, the most valuable measurements are the post-crash deformation map — the residual displacement field of the body-in-white — and the time-resolved intrusion history at selected structural points.

Deformation is typically captured through two methods. First, a coordinate measurement machine (CMM) or structured-light scanner records the residual surface geometry of the vehicle after impact, producing a point cloud that can be registered to the pre-crash CAD model. Second, high-speed stereo photogrammetry tracks the displacement of painted targets on the structure during the crash event, yielding a time series of 3D displacement vectors at each target. The difference between pre-crash and post-crash target positions defines the residual deformation, while the time series captures the deformation rate and sequence of structural collapse.

A practical example: a frontal offset test at 64 km/h with 40% overlap produces a deformation field concentrated on the impacted side. The left longitudinal rail may crush by 350–450 mm, the left apron panel buckles, and the firewall intrudes 80–150 mm into the occupant compartment. These values are not arbitrary — they are the kind of numbers a deformation prediction model must learn to output. The dataset entry for this test would include the impact condition vector (velocity, overlap, barrier type, impact angle), the vehicle configuration (mass, stiffness class, material grades), and the deformation field as a structured output (e.g., a grid of displacement values or a set of keypoint displacements).

Sled Tests: Controlled Loading of Subsystems

Sled tests occupy the middle ground. Instead of crashing a full vehicle, a sled carriage carries a body-in-white or a structural subsystem and decelerates it according to a prescribed pulse. The pulse is typically derived from a full-scale test accelerometer trace, so the sled reproduces the inertial loading that the structure would experience in a real crash without requiring a complete vehicle. This makes sled tests cheaper, more repeatable, and easier to instrument with dense sensor arrays.

For deformation prediction, sled tests are particularly useful for isolating the effect of a single variable. If you want to understand how the longitudinal rail deforms under different pulse shapes, you can run a series of sled tests where only the pulse is varied while the structure remains identical. The deformation is measured with the same photogrammetry or CMM methods used in full-scale tests, but the boundary conditions are simpler and the noise sources are fewer. A typical sled test might use a pulse with a peak deceleration of 30–40 g and a duration of 80–120 ms, producing rail crush of 200–300 mm depending on the pulse shape and the rail's axial crush initiators.

The limitation of sled tests is that they do not capture the full vehicle's load paths. A rail that crushes in a sled test may behave differently when connected to a complete front-end structure with crossmembers, suspension, and powertrain. This is why sled test data is best used to augment full-scale data, not replace it. In a deep learning context, sled tests provide clean, high-volume samples of the structure's intrinsic deformation behavior, while full-scale tests provide the realistic boundary conditions.

Component-Level Experiments: Material and Joint Behavior

At the finest scale, component-level experiments characterize the deformation behavior of individual structural elements: sheet metal coupons, spot-welded joints, adhesive bonds, and small structural assemblies such as a single rail section or a B-pillar subassembly. These tests are essential because the deformation of a full vehicle is ultimately the sum of many local material and joint responses, and a model that has never seen the underlying constitutive behavior will struggle to generalize.

Common component tests include:

• Tensile tests on sheet metal specimens (e.g., DP590, DP780, 22MnB5) to obtain stress–strain curves at strain rates from 0.001/s to 1000/s, since crash deformation involves high-rate loading.
• Spot-weld shear and peel tests to determine the force–displacement response and failure envelope of the joints that hold the body-in-white together.
• Axial crush tests on single rail sections or top-hat specimens, where a hydraulic press or drop tower compresses the specimen at a controlled rate and the force–displacement curve is recorded. A typical top-hat axial crush test on a 1.5 mm DP780 section might show an initial peak force of 80–120 kN, followed by a mean crush force of 40–60 kN over 150 mm of stroke.
• Three-point bending tests on B-pillar or roof rail sections to characterize bending collapse and the formation of plastic hinges.

These tests produce the material cards and joint models that feed into finite element simulations, which in turn generate the large synthetic datasets that deep learning models often require. In a purely data-driven workflow, component test data can also be used directly to train models that predict local deformation from local loading, which are then composed into a full-vehicle prediction.

Practical Considerations for Dataset Construction

When assembling a dataset from these three tiers, several practical issues arise. First, the deformation measurements from different tiers are not directly comparable without normalization. A full-scale test reports deformation in millimeters on a complete vehicle, while a component test reports force and displacement on a coupon. The deep learning architecture must either operate on a common representation (e.g., a normalized deformation field on a fixed grid) or use separate input branches for each tier.

Second, the number of samples per tier is highly imbalanced. A typical research program might have 20–50 full-scale tests, 100–300 sled tests, and thousands of component tests. Training a single model on this combined dataset requires careful weighting or transfer learning: pre-train on component and sled data, then fine-tune on full-scale data.

Third, the impact conditions must be parameterized consistently. A full-scale frontal offset test at 64 km/h and 40% overlap is not the same as a sled test with a pulse derived from that test, even if the pulse is identical, because the full-scale test includes the rear structure, suspension, and powertrain. The dataset should record the tier as a categorical variable so the model can learn tier-specific corrections.

From Experiment to Training Sample

The final step in data acquisition is converting raw experimental measurements into training samples. For a full-scale test, this means registering the post-crash point cloud to the pre-crash geometry, computing the displacement field, and resampling it onto a regular grid or a set of keypoints. For a sled test, the same process applies but with a simpler geometry. For a component test, the force–displacement curve is typically used as a material calibration input rather than a direct training sample, unless the model is designed to predict local deformation from local loading.

A well-structured dataset entry for deformation prediction might look like this:

{
  "test_id": "FS-2024-017",
  "tier": "full_scale",
  "vehicle": {"make": "generic_sedan", "mass_kg": 1520, "year": 2024},
  "impact": {"velocity_kmh": 64, "overlap_pct": 40, "barrier": "ODB", "angle_deg": 0},
  "deformation": {
    "type": "keypoint_displacements",
    "points": [
      {"id": "rail_left_front", "dx_mm": -412, "dy_mm": 18, "dz_mm": -5},
      {"id": "rail_left_mid",   "dx_mm": -287, "dy_mm": 12, "dz_mm": -3},
      {"id": "firewall_left",   "dx_mm": -118, "dy_mm": 6,  "dz_mm": -2},
      {"id": "apron_left",      "dx_mm": -203, "dy_mm": 22, "dz_mm": -8}
    ]
  },
  "intrusion_history": {"firewall_left_max_mm": 142, "time_ms": 95}
}

This structured format allows the deep learning pipeline to ingest heterogeneous test data uniformly. The keypoint displacements serve as the prediction target, while the impact and vehicle parameters serve as inputs. Component and sled data can be mapped into the same schema with appropriate tier labels and reduced keypoint sets.

In summary, full-scale crash tests provide the ground truth for real-world deformation, sled tests provide controlled and repeatable loading scenarios, and component experiments provide the material and joint behavior that underpins the physics. A deep learning model trained on a thoughtfully fused dataset from all three tiers can learn to predict deformation across a wide range of impact conditions, but only if the data acquisition is systematic, the coordinate conventions are consistent, and the tier-specific limitations are explicitly represented in the dataset.