In our latest interview, Michael, team lead “Crash Detection & Predictive Technologies”, explains why the intelligence in current crash sensors is integrated mainly into the vehicle body, how VAIVA is participating in the development of gender-pertinent technologies and why machine learning is the path to intelligent restraint systems, but not the sole answer to all issues.
Let’s start with a brief outline of the status quo of current airbag ECUs…
Exactly. The ECU recognizes if a collision is occurring – in other words a serious situation – and has to decide within a very short period of time – a couple of milliseconds – whether it has to deploy. The requirements for the necessary trigger times are largely derived from dynamic crash test situations stipulated by the government and consumer protection organizations.
The airbag ECU analyzes the increase in pressure and analyzes the remaining accelerometers in the vehicle with respect to the core question: predefined threshold exceeded? Yes or no?
If a threshold value is exceeded with a second signal, then you have a so-called plausibility check. The ECU then deploys the respective airbag. The point in time at which the threshold is exceeded is actually an indicator regarding the body design and therefore also about the body stiffness and deformation behavior. The speeds, the masses of the colliding vehicles – in other words the kinetic energy of the vehicles – and their own body stiffness, ultimately determine how the restraint components behave when deployed. So at the moment, the real “intelligence” – when and what has to be done to protect the occupant as best as possible – is not in the algorithm, but in the body stiffness. This brings us back to the world of physics and mechanics.
So if the intelligence is currently in the body of the vehicle, then the car is not yet able to recognize in real-time how serious the accident might be in order to adjust the restraint effect of the airbag accordingly?
Correct. First and foremost the ECU has a control function. It doesn’t regulate, even though some of the restraint components do have a regulating character. That has mainly to do with the physical and mechanical components and not the logic – the intelligence – that the restraint system controls.
Back to the side protection. Let’s take the side airbag as a concrete example. At the moment it has a chamber filled with a pyrotechnic generator. In other words, in most cases a defined outlet surface, called a vent, is activated so the gas can escape in a targeted manner. That’s all there is to it!
For years the industry has been dreaming about an adaptive side airbag that allows the gas to escape in a targeted manner. Loosely speaking: open the vent – close the vent – open the vent…
Is there any adaptivity with today’s restraint systems?
A certain degree of adaptivity can be found today in the front area in case of a crash. It takes into account specific passenger sizes for example. Women – generally smaller and lighter than men – are considered “five percent” or “50 percent” female. The system merely differentiates between the mass of the occupants.
As part of the current gender and diversity discussions, criticism is being leveled at the basic assumption that the occupants are male and that only male dummies are used.
There is an urgent need for action here!
You could easily represent gender-specific differences in “human models” for instance. These different body aspects (male-female, heavy-thin…) have to be taken into account in the underlying intelligence as well in order to control or deploy the restraint system. In order to become gender compliant in the technical area as well, we need even more information in the vehicle, such as from the interior monitoring and the occupant classification. Are the passengers big or small? What gender are they? Is the person pregnant? The more factors that can or must be used in the triggering decision, the more complex the underlying algorithm becomes..
You just talked about the human body model that plays a role in this issue. What other topics are important for the further development of crash sensors?
The human body model is certainly very important. Other things are highly relevant such as images from the vehicle interior in order to obtain information about the seat positions of the occupants. Registering the occupant kinematics during normal vehicle operation is also important. But not just registering the information, because you also need a corresponding prognosis. Predictors will generally become a big field, and not just in the area of vehicle safety.
The big issue is, what criteria are used to trigger a function? And not only from a technical standpoint. In the future there will have to be more monitoring, discussion and development regarding medical and even ethical criteria.
The technical criteria are already on-hand for the most part and are reflected in the functional and system development activities of restraint systems.
A developer optimizes the restraint system based on a specific position of the occupant. But the development frameworks stem from the laboratory scenarios, or from the basic accident research findings. We want to be even better in the field, out there in reality. What’s important for us is having a view of the situation in the field and what precisely makes sense in this concrete situation when a crash occurs.
So now, in order to know in the field what would be the best course of action in each possible situation, we need a lot of information: Where is the occupant seated? What is the prognosis in terms of crash severity? The keyword here is crash severity prognosis. But first and foremost, what restraint strategy will we use for the crash? Looking into the future when piloted vehicle driving has established itself: if the vehicle occupant is lying down during the drive, should you bring the passenger upright in case of a crash? Does it make sense to perhaps move the occupant in the vehicle interior? Does an additional restraint system make sense?
To calculate all of these scenarios in order to know which crash situations could occur and which restraint system strategy to use for the respective crash scenario, machine learning is the tool that makes it all possible in the first place. But it’s still important to understand that machine learning is not the sole answer to these many issues. We always need a control and plausibility component with machine learning.
This combination of machine learning and a rules-based algorithm – and being able to execute it in an ECU in the vehicle – will be an extremely important field of activity for VAIVA in the coming years and decades..
The combination you describe could also be referred to as an “intelligent restraint system”. What type of challenges do you see in developing such a system?
For me the big challenge is ensuring that the knowledge that already exists in the function development of restraint systems finds its way into the vehicle. Imagine for a moment that function and system developers are sitting in the vehicle – or more precisely in the ECU – and decide based on all of the experience and information they have, exactly which restraint system has to be deployed in this driving situation in order to offer the best protection for the occupants and other traffic participants.
Just for the decision as to whether a restraint system is deployed or not, we need a prognosis. What happens when I deploy? What happens when I don’t? It’s also an ethical trade-off. Is the worst case scenario a broken rib, or perhaps a severe head injury? With older people, a rib injury can be more threatening than a concussion.
So does that mean we should be working closer together with medical professionals in the future?
Definitely. We have to carry out risk analyses in terms of potential injuries. The potential health risks to the individuals have to be taken into account when deciding whether to deploy a restraint system. In the distant future, the health condition of the individual will already be included in the decision. For me that would then count as an intelligent restraint system.