How self-driving vehicles should be allowed on public roads is a topic of regular debate. However, self-driving technology not only makes the use of cars on public roads different, but also the testing of such vehicles. A new algorithm predicts when components of a self-driving car are in danger of breaking down. This allows the vehicle to come to a safe and controlled stop.
The algorithm was developed by Sweden's Chalmers University of Technology. With it, the university aims to contribute to the safe testing of vehicles even if there is no human driver in the vehicle.
Recognise signals in time
Testing cars is currently the work of test drivers. A key skill of these drivers is being able to recognise signals in time that a significant part of the vehicle is in danger of breaking down. This allows test drivers to anticipate this, stop their vehicle in time and avoid accidents.
However, testing of self-driving vehicles is far from always taking place in the presence of a human driver. It is therefore important to notice such signals in a different way. Consider vibrations, anomalous readings, rising temperatures and other factors.
Testing in the same place as regular vehicles
This is particularly important as in the future, self-driving vehicles are expected to be tested on the same test tracks as regular vehicles with drivers. With the safety of these drivers in mind, it is important that problems with self-driving vehicles are detected in time.
The research project is called 'Enablers for testing autonomous vehicles at existing testing grounds'. The research identifies the challenges that arise around testing vehicles at existing test tracks as cars become increasingly autonomous. Researchers from Chalmers University of Technology worked with truck manufacturer Volvo and car manufacturer Volvo Cars on this project.
Real-time warning system
The parties are now presenting an automated system that alerts in real-time if a mechanical component important to the safety of the vehicle is in danger of failing. "The problem lies in how best to monitor a vehicle if you don't know in advance what is going to break down," explains Tomas McKelvey, senior lecturer in signal processing in the Faculty of Electrical Engineering at Chalmers University of Technology.
Self-driving vehicles on the road equipped with a variety of safety systems. However, McKelvey stresses that while testing vehicles, systems whose presence on public roads is mandatory cannot be relied upon. This is because vehicle testing often takes place in early stages of vehicle development. Such autonomous functions are not always reliable then.
Changes in vibrations
Humans detect problems based on changes in vibrations, among other things. The algorithm proceeds in the same way. Using accelerometers at various points in the vehicle, the algorithm records vibrations in the engine and how these vibrations spread through components and the body. "Our hypothesis is that the nature of the vibrations changes if a problem arises on board and those changes can be recorded using instrumentation," McKelvey explains.
Vibrations can, of course, change due to other factors. For example, by variations in the road surface. A number of sensors placed near the wheels therefore measure the vibrations entering the vehicle. The other instruments look for changes in vibrations inside the vehicle.
Collecting data with test vehicles
In developing the algorithm, collecting the right data played an important role. Two test vehicles were used for this purpose: a truck and a passenger car. These were tested at a Volvo test track in Hällered. This is also where the AstaZero test facility of Chalmers University of Technology and RISE is located.
Based on data collected here, a mathematical description of how the different parts of the vehicle move together was created. Based on this data, a model has been created that describes the normal situation. Each value here can have a certain deviation.
Deliberately create technical problems
The researchers then had the two test vehicles drive over the same surface as before. To test the algorithm, they deliberately created technical problems in the vehicle. The system turned out to detect these problems quite easily. Even if they were problems that test drivers say are very difficult to detect.
In one case, the algorithm did not notice a problem. This involved a nut that had loosened slightly on the truck's shock absorber. "However, this was a problem that the test drivers could not sense either," McKelvey said.
Nature of problem not detected
The system currently only detects that a problem has occurred. What this problem is, the algorithm cannot figure out. The researchers expect that this can be developed, but did not have enough time to do so in this study. McKelvey: "From a safety perspective, detecting that a problem is occurring is the most important thing so that the vehicle can be taken off the test tracks."
McKelvey reports that there is a lot of interest in the project from other testing facilities. He hopes the method will be developed into an international standard for test circuits.
Further information about the algorithm is available here.
Author: Wouter Hoeffnagel
