Data-Driven Fault Diagnosis in Battery Systems Through Cross-Cell Monitoring

Fault diagnosis is a central task of Battery Management Systems (BMS) of electric vehicle batteries. The effective implementation of fault diagnosis in the BMS can prevent costly and catastrophic consequences such as thermal runaway of battery cells. As fire incidents of electric vehicles show, the early detection of faults in the latent phase before a thermal runaway is still a problem. The goal is therefore to develop methods with high sensitivity and robustness that detect abnormalities in the battery system even under dynamic load profiles and sensor noise. This work presents a novel data-driven approach to fault diagnosis based on a comparison of single cell voltages. Faults are detected and localized by a statistical evaluation based on a Principal Component Analysis (PCA) of the data. To increase the sensitivity and robustness of the Cross-Cell Monitoring (CCM) method, an outlier robust sample studentization and a new method for selecting the number of principle components is proposed. The CCM data model is recursively updated, to handle non-stationarities caused by cell parameter changes. An application to the data of a large battery system consisting of 432 Lithium-ion cells shows the fault detection and isolation capability. The ability to learn and generalize is shown by an artificial parameter change and cross-validation.

Automated summary

Fault diagnosis in Battery Management Systems is crucial to prevent catastrophic consequences such as thermal runaway of battery cells. This study presents a novel data-driven approach based on comparing single cell voltages to detect faults and localize them using Principal Component Analysis. The method shows sensitivity and robustness in detecting abnormalities even under dynamic load profiles and sensor noise, demonstrating fault detection and isolation capabilities in a large battery system.