In a newly formed research lab,artificial intelligence is training neural networks. The researchers see major benefits in the development of electric powertrains.

Powertrain developers see challenging times ahead. Even today, they are not just designing a gasoline or diesel engine to deliver the best possible performance and efficiency. They are also giving classic internal combustion engines more electric support.

Today’s powertrains feature micro hybrids with start-stop systems or mild hybrids with 48-volt systems, all the way to high-voltage hybrids with or without an external charging feature. There are six different architectures for the electrification of the internal combustion engine, ranging from the P0 arrangement with a belt-driven starter-generator (which replaces the alternator) to the P5 arrangement, where electric motors are housed in wheel hubs. The all-electric vehicle is at the far end of this spectrum. To be sure, it does without an internal combustion engine, but the powertrain itself can take a number of different forms. This makes relationships in the powertrain increasingly difficult to monitor.

One possible solution could be greater use of artificial intelligence in the development of powertrains. For example, teaming up with the German Research Center for Artificial Intelligence (DFKI), engineering services provider IAV formed the joint Research Lab for Learning from Test Data (FlaP). The lab has its own space at the DFKI facility in Kaiserslautern, Germany, and it is starting out with a staff of four. The team is due to expand quickly in coming months.

Improving data analysis

“Huge quantities of test data are produced in the development of control systems for powertrains,” said Andreas Dengel, who is in charge of the joint venture on the DFKI side. These test data are captured with numerous sensors and are stored in multidimensional data spaces. More than 100 sensors are used in today’s development equipment. “In our new research lab, we plan to use artificial intelligence and ‘deep learning’ methods to carry out so-called big data analyses in various dimensions so we can improve our understanding of the relationships within the engine,” Dengel said. The fundamental work should provide greater efficiency in the development process whileenabling the development of robust systems in an increasingly complex environment.

Matthias Schultalbers, who isin charge of the new research lab on behalf of Powertrain Mechatronics at IAV, expects the collaboration to provide better access to research results, especially on the use of AI in data analysis.A team at IAV’s company headquarters in Gifhorn in northern Germany has already been working on data analytics and its applications in control units in engines, transmissions and power electronics. “The new research lab is supposed to apply AI and deep learning methods to systematically evaluate IAV test data, filtering out deviations from normal operations and analyzing their causes,” Schultalbers said. To make this possible, IAV has built a high-performance GPU cluster that is capable of processing huge quantities of data.

Neural networks

The deep-learning approach uses neural networks that build up in numerous deep layers. The two partners would not say just how deep they go. But in theory, with the maximum amount of “learning data,” these layers learn independently how the modeling in a control device and the subsequent parametrization can occur. First the many individual layers of neurons separate individual tasks into their basic forms and then identify increasingly complex interrelationships in the deeper layers. This makes them easily scalable for increasingly complex tasks. A major advantage of neural networks is that computing power can be adjusted to their complexity. 

One disadvantage is that they are “black boxes.” It is certainly possible to validate a result. But precisely how neural networks achieve their results has not really been understood so far. Researchers at IAV and DFKI want to determine two things: First, the conditions under which a model does not correspond to reality; and second, the conditions within a model under which a deviation from the target behavior occurs. 

“In these cases, AI gives us a tremendous opportunity to gain new insights,” Schultalbers said. “For example, in regulating boost pressure, we can identify the situations where a regulator behaves in a certain way with much greater precision. If there are anomalies, the regulator parameters can be optimized and hardware defects identified in advance.”

Predictive health monitoring

Assuming they can get the right amount of computing power, the partners plan to take these methods into the field. They will function as predictive health monitoring. “It is like an early warning system,” Dengel said. “The data on vibrations, pressures and temperatures that an engine records over time is evaluated with AI techniques in the control device. This makes it possible to detect an error that is becoming established early.” A vehicle can be sent to a service shop promptly, or a system can be designed to be error-tolerant in advance. 

Looking further ahead, the two partners plan to use measurement data from the development process to define how a car’s electronic control units (ECUs) should perform. For example, an engine control unit would independently adjust to different environmental conditions — such as the hot dry climate of California or the moist, cold weather of Scandinavia. With the help of flexible control devices, the system could be restored to its ideal physical condition again and again, even in the event of sensor tolerances or wear. That assumes there will be sufficient computing power in the vehicle.

-By Laurin Paschek