In recent years, lithium-ion battery applications have undergone substantial growth, yet concerns regarding safety persist. Despite the relatively low failure rate, any cell fault can impede battery application and, in the context of increasing system capacities for mobility or grid stability applications, frequently result in the destruction of substantial portions or even the entire energy storage system. Given the inherent limitations of total prevention, reliable detection of faults, such as internal short circuits (ISC), is paramount for ensuring the safe implementation of lithium-ion battery systems.
However, the selection of an appropriate detection algorithm based on published studies is hindered by the absence of standardized conditions for test boundaries, fault emulation, and evaluation aspects. This study employs a Monte-Carlo-like approach, covering a broad range of ISC faults (resistance, time, duration) under the influence of various cell-to-cell variations (CtCV) within a module. This approach enables more detailed aggregated analysis and overcomes the limitations of often-published binary detection/miss results. The consideration of a 12s1p-module allows for advanced comparison between cells, eliminating the need for complex self-parameterizing models as a reference.
In this study, a range of recently published detection methods based on the voltage signal commonly monitored are implemented and evaluated comparably. These methods span from basic statistics (deviation from the mean), correlations, and outlier detection to information entropy (e.g., sample entropy). The methods are tested against the same test data that incorporates faults and fault-free samples. By systematically sweeping both CtCV and algorithm parameters, the corresponding sensitivity is analysed, and the best detection performance is identified. The study reveals that the achievable detection performance differs significantly between methods, and that the selected algorithm parameters, such as the moving window width, have a substantial influence on the detectability of ISC faults. Therefore, the utilization of robust methods with fewer parameters is recommended, particularly since no overall superior detection performance could be identified for the more complex methods. Within these methods, only correlation-based algorithms such as Pearson Correlation and certain variants of intra-class correlation (ICC) could compensate the detection disturbance of impedance and voltage fluctuations expected within real-world applications. In addition, the reduced complexity of these methods leads to a simplification of the application, as they require less prior parameterization and lower computational complexity, enabling online monitoring with reduced hardware requirements.
The results of this study provide the side-by-side comparison of many established detection methods and reveal significant differences in performance and resilience for disturbances by imperfection of measurement and cell homogeneity. Thus, the results further confirm the requirement for such investigations based on a controlled and large dataset.