Machine learning-based fault diagnosis for lithium-ion battery systems

Abstract

Accurate and reliable fault detection and isolation (FDI) in lithium-ion battery (LIB) systems is essential for ensuring safety, optimal performance, and lifespan, particularly as LIBs power critical applications such as electric vehicles and renewable energy storage. Traditional model-based diagnostics face limitations in handling sudden and compound faults due to modeling uncertainties and transient dynamics. This study proposes an experimental/synthetic hybrid, data-driven FDI framework that leverages supervised machine learning (ML) integrated with an experimentally validated second-order electro-thermal battery model to generate a mixed experimental–synthetic dataset covering healthy operation and diverse fault conditions, including cell, connection, and sensor faults. Time-domain statistical features are extracted from current, voltage, and temperature measurements, and nature-inspired wrapper-based algorithms are employed for feature selection, while multiple ML classifiers—Random Forest (RF), Decision Trees (DTs), K-Nearest Neighbors (KNN), and Naive Bayes (NB)—are comparatively evaluated on a six-class diagnosis problem. The best-performing configurations achieve near-perfect multi-class accuracy of approximately 99.80–99.87% with low variance across repeated train–test splits, demonstrating high robustness and stability overall. Rather than identifying a single universally optimal model, the results indicate a practical trade-off between diagnostic performance and implementation cost. Within this group, the RF classifier combined with Particle Swarm Optimization (PSO) for feature selection (RF–PSO) attains slightly higher accuracy and stability than the other top combinations, while DT–PSO and KNN–MPA (Marine Predators Algorithm) yield much smaller models, shorter inference times, and more compact feature subsets, which are attractive for constrained embedded implementations and large pack deployments. The proposed framework thus enables fast, real-time diagnosis of complex multi-fault scenarios at the cell or module level in series–parallel LIB pack architectures and allows practitioners to select the most suitable classifier–optimizer pair according to application-specific priorities on accuracy, robustness, computational cost, and diagnostic granularity.