2. Physical and engineering background

Before any neural network can predict how an airbag behaves, we must define precisely what is being predicted and why the physical system behaves the way it does. An airbag is not a single object but a pyrotechnic-mechanical assembly whose response is governed by gas dynamics, material deformation, and occupant kinematics. This section establishes the physical vocabulary and the measurable quantities that later chapters will treat as regression targets.

System Architecture and Components

A frontal driver-side airbag module consists of several interacting subsystems. The inflator is the energy source: either a pyrotechnic device burning solid propellant, a hybrid unit combining compressed gas with a small pyrotechnic charge, or a cold-gas inflator. The airbag cushion is a woven fabric envelope, typically nylon 6.6 coated with silicone or neoprene to control permeability. The module housing retains the folded cushion and directs gas flow. A retainer ring clamps the cushion throat to the housing, and the cover tears along a designed seam when internal pressure rises. The Electronic Control Unit (ECU) hosts accelerometers and the crash algorithm that issues the fire command.

Two physical parameters dominate the response. The first is inflator mass flow rate ṁ(t), the rate at which gas enters the cushion, usually expressed in kg/s or as a pressure-time curve from a tank test. The second is cushion venting, characterized by vent hole area and fabric permeability, which removes gas as the occupant loads the bag. The balance between these two determines internal pressure and thus the force transmitted to the occupant.

Deployment Stages

Deployment proceeds through four overlapping stages, each with a distinct timescale measured from the fire command (t = 0):

1. Initiation (0–5 ms): The ECU closes the firing circuit; the squib ignites the propellant. Gas generation begins but the cushion is still folded.
2. Inflation (5–30 ms): The cushion unfolds and fills. Internal pressure rises rapidly, peaking near 200–400 kPa in a typical driver module. The cover tears open.
3. Occupant interaction (30–80 ms): The occupant contacts the cushion. Gas is vented and fabric stretches; pressure may drop or plateau depending on venting design.
4. Deflation (80–150 ms): Gas escapes through vents and permeable fabric, and the cushion collapses, absorbing residual energy.

These stages are not rigidly sequential; inflation and occupant interaction overlap significantly in real crashes. The key engineering insight is that the timing of peak pressure relative to occupant contact determines injury risk. An airbag that peaks too early is already deflating when the occupant arrives; one that peaks too late strikes the occupant while still expanding.

Key Response Metrics

The quantities we ultimately want the model to predict fall into three groups:

In a data-driven workflow, P_max, t_Pmax, and a_chest are the most common regression targets because they are directly measurable in a sled test and strongly correlated with occupant injury. Volume V(t) is harder to measure directly and is usually inferred from pressure and mass flow via the ideal gas law.

From Physics to Features

The practical consequence for machine learning is that raw sensor channels are not the best inputs. Instead, we construct physics-informed features that encode the governing relationships. For example, the isentropic relation for gas entering a control volume gives a first-order estimate of pressure rise:

dP/dt = (γ - 1)/V * (ṁ * c_p * T_in - P * dV/dt)

where:
  γ   = ratio of specific heats (≈1.4 for air, ≈1.3 for propellant gas)
  V   = instantaneous cushion volume (L)
  ṁ   = inflator mass flow rate (kg/s)
  c_p = specific heat at constant pressure (kJ/kg·K)
  T_in= gas temperature at inflator exit (K)
  P   = gauge pressure (kPa)

This equation tells us that pressure rise depends on the ratio of mass inflow to volume growth. A model that receives ṁ(t), V(t), and T_in as separate features can learn the residual effects of venting and fabric stretch that the ideal equation ignores. Conversely, a model fed only raw accelerometer traces must rediscover this relationship from data, which requires far more samples.

Engineering Example: Interpreting a Pressure Trace

Consider a driver-side hybrid inflator tested in a 50 km/h frontal sled. The measured cushion pressure rises from 0 kPa at t = 0 to 280 kPa at t = 22 ms, then falls to 180 kPa by t = 45 ms as the occupant loads the bag. The chest deceleration peaks at 42 g at t = 38 ms. From these numbers we can extract the feature vector [P_max=280, t_Pmax=22, P_contact≈210, a_chest=42]. A model trained on hundreds of such vectors can then predict a_chest for a new inflator charge mass without running a full sled test.

Typical failure symptoms to watch for: if P_max is high but t_Pmax is very short (<15 ms), the bag may be too aggressive for an out-of-position occupant. If P_contact is below 100 kPa, the bag may be under-inflated and bottom out against the steering column. These physical checks validate whether a predicted response is plausible before it is trusted.

The figure above ties the three layers together: the component stack on the left, the four deployment stages in the center, and the measurable response metrics on the right, all mapped onto a representative pressure-time curve. The dashed vertical lines mark the two most important events — peak pressure and occupant contact — whose separation in time is the single strongest predictor of occupant loading. In the next section we will formalize how these physical quantities become the input and output tensors of a deep learning model.