Problem
Due to the increasing competitiveness of electric cars compared to vehicles with conventional combustion engines and legal regulations regarding the reduction of CO2 emissions, electric vehicles are recording a significant increase in the number of registrations in Germany. As the number of registrations increases, so does the variety of high-voltage batteries used in electric vehicles. Starting at cell level, these differ in terms of the cell format used, the cell size and the cell chemistry. At module and pack level, different assembly variants and integration concepts (such as "Cell2Body" or "Cell2Pack") also increase the number of high-voltage battery variants used in practice.
In order to ensure safety, performance and high guarantees on the part of the manufacturer, a large number of different tests must be carried out. In connection with the existing and future traction battery variants, this represents an enormous testing effort. Due to the increasing number of variants, this increases with increasing integration level and test duration. For this reason, service life tests at system level are particularly time-consuming and cost-intensive, but nevertheless essential in the design and validation of high-voltage batteries.
In a previous research project (KIBaTest - Artificial Intelligence for battery testing), it was shown that measurement errors have a significant impact on data quality and therefore on AI-based prediction. In addition, purely AI-based prediction is not possible at system level, as the amount of accessible data is insufficient. This has led to the identification of two major problems with lengthy service life tests:
- The increased test duration and the complex implementation in long physical tests entail an increased probability of occurrence of possible errors in the measurement, which reduces the data quality and thus also the model quality. If errors are detected too late, serious errors can lead to the test being aborted or, in the worst case, to damage to the test device or test specimen.
- The large number of battery variants leads to a very high testing requirement. However, the possible test implementation is restricted due to the limited availability of test equipment and high costs and should therefore be reduced as far as possible. To this end, it is desirable that simulations accompany the testing process so that it becomes more robust and at the same time easier to plan. Furthermore, it is not possible to couple different tests with current methods, which means that they have to be considered in isolation. At higher assembly levels, there is also insufficient ageing data available to create AI models.
Goal
The project aims to improve the robustness, transparency, sustainability and cost-effectiveness of battery testing. This is essential to meet the need for faster development of high-performance batteries.
- Robustness: A digital twin of the battery system will accompany the testing process and enable online anomaly and fault detection. After a battery test, it is often difficult to identify whether atypical behavior is due to actual cell behavior or measurement errors. AI-based methods will be used to detect errors such as swapped cells, sensor drift, errors in the sequence of test plans or unplanned downtimes. In addition, abnormal cells that behave atypically compared to cells with similar ageing conditions are to be identified and sorted out at an early stage. This ultimately provides usable ageing series with valid cells and test conditions for all defined measuring points in the test plan.
- Transparency: The use of a digital twin enables a deeper insight into the battery system in addition to the direct measured variables. The test engineer or client receives a transparent display of the test procedure in real time and retrospectively and can use their expertise to assess the quality of the test results and, if necessary, initiate an early restart or a repetition or modification of individual tests.
- Sustainability & cost-effectiveness: The focus should be on lifetime tests, as these offer the greatest leverage, as shown in the KIBaTest project. The test plan can be optimized using AI-based methods and test points can be saved or partially virtualized. Furthermore, the great effort involved in tests at system level is to be reduced by combining single-cell tests, AI-based ageing models and physical system models, with comparable reliability of the results at system level.
Implementation
The objectives described are to be achieved through the use of digital twins. These represent a virtual image of the real battery cells. The physical tests are to be coupled with a digital twin, allowing anomalies to be detected at an early stage and the associated test abort to be prevented. Fully comprehensive and automated error detection allows the test engineer to be made aware of errors and rectify them immediately, thereby improving data quality. The improved data quality not only serves the customer, but also the continuous training (online learning) of the digital twins in order to be able to create AI-based prediction models for ageing behavior. Physics-based AI models enable efficient coupling of tests at cell level to entire battery systems from module to pack level. The changed electrical and thermal boundary conditions must be correctly mapped at system level. Thermo-electrical simulation models are set up for this purpose, as the thermal and electrical parameters influence each other and these change over the service life. The models at cell and system level are therefore not static, but evolve dynamically over the lifetime test by adapting both the digital twin at cell level and the test parameters of the individual cells and thus also flowing back into the system model. The prediction of ageing at system level is to be carried out using an AI model that has been trained on cell data and a physical system model. This is intended to simulatively generate test points of the test process that are easy to predict. In order to increase the robustness of the prediction, physical correlations are integrated into the model, as simple neural networks are very accurate but require a large amount of data that is not available at system level. There is also a risk of "overfitting", whereby correlations are learned in the data that do not exist or are physically implausible. Pure physical models, on the other hand, are complex to parameterize and offer no advantage in terms of shortening test times and reducing costs. A combination of the two approaches therefore appears promising in order to link the universal applicability of neural networks with the physical behavior of the battery.
