Algorithm reveals faults in self-driving test vehicles

But how can this safety-critical professionalism be passed on to a future when more and more vehicles become self-driving? Is it possible to mix self-driving and driver-controlled vehicles on one and the same test track? How is safety affected when there is no longer anyone on board who can break or steer to the side to prevent an accident? These are some of the issues behind ETAVEP (Enablers for testing autonomous vehicles at existing testing grounds), a research project co-financed by Vinnova program Vehicle Strategic Research and Innovation, which takes a closer look at the challenges in vehicle testing that arise as cars become increasingly autonomous.

Researchers from Chalmers, together with the truck manufacturer AB Volvo and the car manufacturer Volvo Cars, have focused on the part of the project that deals with the fact that vehicles must be able to be monitored even when a person's ears, sensory organs and experience are not in place. 

The task has thus been to create an automated system that in real time can give an indication that some mechanical component that is important for safety is about to fail.

Tomas McKelvey, Professor in Signal processing at the Department of Electrical Engineering and, who has led the research within the subproject, explains:

"The problem lies in how to best monitor a vehicle when you do not know in advance what it is that will break.”

He emphasises that one cannot rely on the systems that need to be on board once the self-driving vehicles come out in normal traffic - this is because test drives can take place even in earlier phases of vehicle development before the autonomous functions are reliable.

You could say that the approach chosen is somewhat reminiscent of how the human perception works - we are usually quite good at paying attention to sensory impressions that deviate from what is expected. 

A few dozen accelerometers, placed in strategically important places in the vehicle, have been allowed to act as the system's tactile rod. They register how the vibrations that occur in the engine and in contact with the road surface propagate further out into the body and components.

“Our hypothesis has been that the nature of the vibrations will change when an error occurs on board and that the changes can be registered with the help of these instruments.” 

To prevent a changed road surface from triggering a "false alarm", some of the sensors, located near the vehicle's wheels, are used to register the input signal. The rest of the instruments should then be able to pick up such changes that are due to faults on board.

But to know how a broken car vibrates, an automatic system must first learn how a faultless car behaves. So, the collection of this data became the researchers' first task once the instrumentation of the two test vehicles - a heavy truck and a passenger car - was completed. 

"We achieve this by creating a so-called transfer function, a mathematical description of how different parts of the car move together.”

He adds that this type of analysis creates "an awful lot of data" - it is about 1,500 so-called complex numbers (numbers that have a real part and an imaginary part) per second. 

"From these data, we build a stochastic model that describes the normal case, where each value is allowed to have a certain variation.”

In the next stage, the researchers let the two test cars roll on the same surface as before, but now with various intentional errors introduced in the vehicles. 

Data from the accelerometers were analysed again - this time with the hope that the monitoring system could perceive changes in vibrations as an indication of the faults. 

Did it work? 

"Yes, the system could quite easily detect errors - even such errors that according to the test drivers are very difficult to detect.”

"Only one of the intentional faults did the system bite on - a nut on a shock absorber attachment that had been loosened a few turns on the heavy truck. But on the other hand, it was a fault that the test drivers could not feel either.”

The automatic monitoring system that the Chalmers researchers have developed reacts to something being wrong - but the system cannot tell exactly where the fault is. 

"You could probably work on this fur ther, but we have not had time for that in this study. From a safety point of view, the most important thing is to find out that an error has occurred, so that the vehicle can be taken off the test track.”

The practical experiments with the monitoring system took place last autumn at Volvo's test track in Hällered outside Borås, where the test facility AstaZero, run by Chalmers and RISE, is also located.

In addition to monitoring vehicle status, which the Chalmers researchers worked on in this sub-project, ETAVEP also includes methods and systems for traffic monitoring and vehicle control, including using radar, light radar (lidar) and camera surveillance. The systems are linked to communication based on 5G technology. 

According to Tomas McKelvey, the methodology developed within the overall ETAVEP project is the first among test facilities around the world, when it comes to allowing autonomous and driver-controlled vehicles to coexist. 

"The hope is that the methods developed will become a kind of international standard for test tracks.” 

He adds that representatives of other test facilities have shown great interest in the project. The reason is that there are both economic and environmental benefits from being able to use existing facilities, rather than having to sacrifice land and resources to build new test tracks only for autonomous vehicles. 

"In addition, the test driver's working life as it looks today is very physically stressful. So even from that aspect there is a sustainability perspective,” Tomas McKelvey concludes.

• The two-year research project ETAVEP (Enablers for testing autonomous vehicles at existing testing grounds) is part of the Vinnova program Vehicle Strategic Research and Innovation. 
• In addition to AB Volvo and Volvo Cars and Chalmers, the project participants consist of the research institute RISE, the test facility AstaZero and the research company SafeRadar. 
• Final report from the project will be written in the spring. 
• In addition to Tomas McKelvey, those who worked on the sub-project on automatic monitoring have been Project Manager Patrik Nordberg from Volvo Cars and Daniel McKelvey, a student in Engineering mathematics and computational science at Chalmers, who had the project as his master thesis. 

An instrument that measures acceleration in relation to free fall, often in three dimensions, and which is found in every modern mobile phone, among other things. An accelerometer at rest registers the earth's gravity. The instrument can also, as in this case, be used to accurately detect very small movements and vibrations.

Tomas Mckelvey, Full Professor in the Signal processing research group at the Department of Electrical Engineering, Chalmers

tomas.mckelvey@chalmers.se