Data preprocessing for fault detection using principal component analysis: A case study on faults within battery systems

An essential objective of modern battery management systems is to ensure safe battery operation. The system shall issue a warning when it detects a fault condition—the earlier, the better. Reliable fault detection requires a sensitive but equally robust detection method. In battery systems, a conspicuous change in the cell-to-cell variation can be an adequate fault indicator. To detect anomalies like this, a recurring method in the scientific literature is principal component analysis (PCA) or one of its variants. Its effectiveness depends heavily on input data preprocessing, though, a topic that is often disregarded. To raise awareness of proper preprocessing, we investigate common techniques in terms of their impact on PCA-based fault detection. We take thermal faults and internal short circuits in battery systems as prototypical examples for our investigation. Our results highlight the importance of selecting appropriate preprocessing for the problem at hand. Well-matched preprocessing shapes the input data such that the theoretical requirements of PCA are met to a greater extent, which enhances overall detection performance. Based on this understanding, we propose a combination of outlier-robust sample studentization and Pareto scaling that improves fault detection over alternatives from the literature in our exemplary application. The detection sensitivity and robustness are greatly enhanced, particularly for minor faults, without adding significant computational complexity.