How do we validate e-drive and e-axle prototypes at Bosch Research?

Series inverters for motor vehicles are inexpensive, efficient, powerful, robust and very reliable. But they are highly optimized for their specific use case or for the requirements of a specific customer in addition to being very inflexible. Rapid prototyping inverters are low-volume products optimized for flexible use, but they are often expensive and not as efficient, reliable and powerful as high-volume automotive inverters.

To combine the advantages of both worlds, we use a self-designed adapter board, which enables a performant control of Bosch series and prototype inverters of different generations by a commercially available high performance rapid prototyping system. This concept enables switching frequencies of well over 50 kHz and the use of almost free pulse patterns. At the same time, the safety functions and reliability of the series hardware can be further used. The next illustration shows this self-developed adapter board mounted on a Bosch inverter.

For the control of the complete test stand with load machines, DC sources, cooling/heating and lubrication systems, and the coordination of the recording of the various measurements, we use a commercially available, open and very flexible automation system. This automation system also has the necessary certified safety chain that reacts very quickly in the event of a malfunction and switches off the power to the test stand. This means that it switches off the electrical energy, discharges electrical accumulators and brakes rotating masses. This automation system also enables the testing of equipment in complex driving cycles with automatic control of all relevant auxiliaries such as cooling and lubrication.

Certainly, it is best to perform measurements with the device under test in its real environment and with the intended use cases. This requires Hardware in the Loop (HiL) simulations where the test bench has to emulate the different environments of the device under test. This ranges from simple speed control of an examinee like an e-machine up to the emulation of complete wheel or vehicle dynamic behavior by one or more suitable load machines. It should be possible to operate the drive or axle on different driving cycles with different vehicles. In addition, the test bench must provide sufficient electrical power and emulate even the behavior of vehicle batteries and provide a suitable cooling power supply.

The HiL integration of an e-axle is briefly described below. This enables the modelling of the entire vehicle and driving situation in the software and the use of a real e-axle within this virtual car environment.

The vehicle model runs on a standard laptop with adequate software. The big issue here is the co-simulation with the real-time interface of the test bench. In stand-alone simulation, it is not critical if a simulation time step takes longer than in real-time. This can happen, for example, when the solver needs more time for calculation or there are background tasks like storing data. The test-bench control demands hard real-time over the complete operation time. At the same time, the vehicle model cannot provide results in advance because these results depend on the real-time information from the test bench.

The solution is to use a sufficiently fast computer for the vehicle model calculation and to reserve enough time for calculation even in worst-case scenarios. Still, there is a small chance that the necessary information will not be available in rare cases. In these instances, the test bench will have to run another timestep with old results if critical limits are not reached. If critical limits are reached, which is even much more seldom, because then a critical operating point must occur with a very rare time delay in the simulation, then the experiment must be aborted and repeated.

The following figure shows a rough overview over the realized HiL concept at the test bench. The used buses are mainly private CAN-buses, which are fast enough for the application.

Artificial intelligence is becoming increasingly important in the field of electrification. Data generation for artificial intelligence (AI) algorithms often leads to a high time requirement for test-bench use and thus to high test-bench costs. To meet these new and challenging requirements, we not only use the test bench for AI data generation, but also implement algorithms with integrated AI to improve our data generation process. These algorithms and methods are called safe active learning.

Data generation for AI models is often based on a classical experiment plan (e.g., grid measurements) or on a Design of Experiment (DoE) with a large amount of measurement data. Classically, DoE definition, data generation and AI model training are three separate domains with sequential and manual processing. To improve this process and adapt to the needs of artificial intelligence applications, we are developing a safe active learning approach, which includes all three areas in one automated process and combines this with AI model training. This approach, as shown in the diagram below, involves many steps that in the past had to be performed manually and required a lot of expert knowledge for each area and a deep understanding of the underlying system. With the new approach, only the basic constraints and operational limits need to be defined. The framework then generates trajectories that are likely to have the greatest relevance to the desired AI model, ensuring that critical regions under consideration are avoided during measurement. These short trajectories, called snippets, are then measured at the test bench. The AI regression model is then re-trained with the new measurement data, accuracy is checked, and a new trajectory is calculated based on the re-trained AI model.

In addition to efficient trajectory generation, safe active learning also helps to minimize the shutdown of test benches when they operate in certain ranges that exceed the allowed physical limit of the system, such as critical temperatures or vibrations. For this purpose, we use an entropy-based active learning framework extended by a constraint-inducing safety. For the main regression task, we use Gaussian processes (GP’s) to learn the system behavior from a limited budget of incrementally sampled data points. Our exploration strategy for determining the next query points is based on the differential entropy criteria. We also employ a problem-specific GP classifier to identify safe and unsafe regions. The basic idea of this discriminative GP is to learn a decision boundary between two classes, preferably without sampling a point in the uncertain region. This contributes to an automated, efficient and unsupervised test-bench measurement and AI model training and ensures safe exploration and data generation without risking test-bench damage or time wasted due to unnecessary shutdowns.

We feel future proof with our flexible e-mobility prototype testing concept, even for new applications like data generation for artificial intelligence.