AI Helps to Measure Temperature

AI helps ZF engineers to understand processes where cost or time constraints prevent them from developing a physically reliable model. The problem is that the temperature inside the rotor cannot be measured directly while the electric motor operates in the vehicle. Exact thermal measurements therefore require a very complex setup on the test bench. Furthermore, an FEM simulation cannot reliably determine the effect of the oil cooling. "The test bench showed that we have a too large deviation with our simulation models – this is not precise enough," says Hoffmann.

On the other hand, there is enough information and a lot of measurement data. Both are determined during complex functional tests on the test bench and later also in the test vehicles. Temperature measurements from the environment – such as the temperature of the oil in the oil pan – are available as well as the oil quantity used for cooling. The rotor speeds are also measured continuously.

At first glance, the figures might seem manageable – but, in fact, you are dealing with a flood of data resulting from a wide variety of possible operating points and their progression over time. These figures depend on whether and when drivers call up full power or cruise along at walking pace. This results in millions of possible combinations of the measured parameters. A human being cannot look at this data and find recurring patterns.

This support by artificial intelligence has enormous potential for ZF. On the one hand, it slashes development times for electric motors used in passenger cars. "Time-consuming special tests that we used with earlier procedures to ‘protect’ the physical properties of wide-ranging components and so check the simulation quality, can now be omitted," explains Hoffmann. Another aspect is that temperature measurement in inaccessible, dynamic areas is an important factor not only for electric motors for passenger car drives: "The AI model could be applied to many ZF products – whenever the operating temperature is essential for function development and exact measurement during operation is not possible," adds Dr Martina Flatscher, AI Strategist in Corporate R&D at ZF

However, AI support does not come for "free”. Tailoring the data to AI entails lots of work. "Data preparation takes up as much as 90 percent of the time; the actual training of the AI then only accounts for ten percent of the total time," according to Flatscher's rule of thumb. If the algorithm programmed by ZF is to compare the terabytes of measurement data, this data must be as uniform as possible – which is not usually the case. In addition to divisional differences at ZF, there are also customer-specific deviations – both in terms of signal standards and metadata. "The designation for motor speed, for example, changes depending on the brand of the test vehicle," Hoffmann says.

AI can only provide meaningful support where valid measurement data is available as a reference and where standardized amounts of data are available for evaluation. As such, "human" colleagues are also becoming increasingly important. Measuring technology expertise and good measurement data handling are crucial here. It is worth rethinking priorities. The question: "What does it take to optimally teach an AI algorithm?" could play a much more important role in future development work.