Graph-guided fault detection for multi-type lithium-ion batteries in realistic electric vehicles optimized by ensemble learning

Abstract

Accurately evaluating the safety status of lithium-ion battery systems in electric vehicles is imperative due to the challenges in effectively predicting potential battery failure risks under stochastic profiles. Complex battery fault mechanisms and limited poor-quality data collection impede fault detection for battery systems under real-world conditions. This paper proposes a novel graph-guided fault detection method designed to recognize concealed anomalies in realistic data. Graphs guided by physical relationships are constructed for learning the dynamic evolution of physical quantities under normal conditions and their potential change characteristics in fault scenarios. An ensemble Graph Sample and Aggregate Network model are developed to tackle sample distribution imbalances and non-uniformity battery system specifications across vehicles. Failure risk probabilities for diverse battery charging and discharging segments are derived. An ablation study verifies the necessity of ensemble learning in addressing imbalanced datasets. Analysis of 102,095 segments across 86 vehicles with different battery material systems, battery capacities, and numbers of cells and temperature sensors confirms the robustness and generalization of the proposed method, yielding a recall of 98.37%. By introducing the graph, spatio-temporal global fault characteristics of battery systems are automatically extracted. The coupling relationship and evolution of physical quantities under both normal and faulty states are established, effectively uncovering fault information hidden in collected battery data without observable anomalies. The safety state of battery systems is reflected in terms of failure risk probability, providing reliable data support for battery system maintenance.