Methodology for the quantitative assessment and comparison of fault detection approaches at lithium-ion batteries with special consideration of internal short circuits

Methodology for the quantitative assessment and comparison of fault detection approaches at lithium-ion batteries with special consideration of internal short circuits

With the strong growth of lithium-ion battery applications in recent years, the corresponding safety problems have also come into focus. Therefore, numerous papers have been published in the past with approaches for the early detection of battery faults, e.g. internal short circuits (ISC), which deliver promising results based on test data sets – simulations, abuse tests or also field data. In order to select a suitable method for the specific application and requirements, a well-founded evaluation with regard to qualitative and quantitative properties is necessary. Here, for example, the acceptable false positive rate or maximum detection time can represent a criterion.

However, due to the multitude of boundary conditions and the heterogeneity of the test data, such a comparison is currently very limited. Among other things, the strongly varying consideration of disturbance variables such as measurement uncertainty can be observed. Furthermore, the data basis of many studies is relatively small, which limits the statistical significance of the results, while real fleet data used is usually not accessible and the tests are accordingly not reproducible.

The method presented here combines the procedure of a Monte Carlo simulation with simulative ISC replication of a 12s module to generate a reliable broad data basis for the validation of fault detection methods. Here, not only a wide variety of fault characteristics in terms of resistance, timing and duration are investigated, but also the unavoidable deviations between individual cell voltage measurements.

The test data generated this way are tested for the presence of an ISC on the basis of both the z-score and the deviation from the module mean. For this purpose, error-free data are first used for the deterministic definition of a trigger limit. The erroneous data are then classified and the quality achieved is compared. In addition to the variation of detection parameters (trigger threshold, filter), the simulative data generation also allows the investigation of different influencing variables, e.g. the level of the measurement noise.

The results of the outlined workflow show that ISC errors can be detected – however, with a small fault (high resistance, short duration), reliable detection is no longer possible. Furthermore, it becomes clear that the achieved detection quality on the same data is significantly influenced by the method parameters (threshold and filter) and that when varying the measurement noise, the basic behaviour remains constant but the quality varies significantly.

The sensitivities identified underline the absolute necessity to validate detection methods on identical test data to allow comparison. The number of individual test data shown (N&gt“;“1000) could only be realised experimentally with great effort and without the possibility of varying influencing variables. Thus, the Monte Carlo approach presented is a helpful building block for the validation of detection methods.