DMS-SVDD: Dynamic Multiscale State-Space Support Vector Data Description for Lithium-Ion Battery Fault Detection from Electric Vehicle Charging Segments

Fault detection from electric vehicle charging segments is challenging because real-world records are noisy, verified fault labels are scarce, and weak signatures may evolve gradually across time and unevenly across a vehicle’s charging history. This study proposes an unsupervised dynamic multiscale state-space support vector data description framework for vehicle-level battery fault detection. A gated diagonal state-space encoder preserves long-range charging patterns while retaining a direct input path for transient changes. A progressive cross-scale fusion module then combines short-term fluctuations with accumulated deviations in the learned hidden representation. Finally, a two-stage hypersphere optimisation strategy first estimates the normal centre and then refines the boundary around that fixed centre. This coordinated design avoids sequence reconstruction and directly scores charging segments by their distance from normal behaviour before robust vehicle-level aggregation. On the two EVBattery subsets that permit statistically reliable evaluation, the proposed framework achieved vehicle-level areas under the receiver operating characteristic curves of 0.8849 and 0.8438. These results exceed those of the best-performing baseline on the corresponding subsets by 0.0889 and 0.0417, respectively. The results show that coordinating long-range encoding, cross-scale fusion, and staged boundary learning improves threshold-independent vehicle-level fault ranking in real charging data.