EVFaultNet: A Cell-Decay Recurrent–Convolutional Architecture for Imbalanced Electric Vehicle Battery Fault Diagnosis Under Sparse Telemetry

Battery telemetry from electric vehicles arrives at a low sampling rate and is severely imbalanced, making fault diagnosis a long-memory, long-tailed classification problem. This paper presents EVFaultNet, a recurrent–convolutional architecture for this setting. Its key innovation is a learned resting point for the recurrent hidden state: before each gate update, the state relaxes toward a learnable per-unit baseline rather than zero, which lets stale information decay to a neutral value. A causal temporal convolution then refines local trajectory patterns, and a deferred imbalance-aware objective handles the long-tailed labels. On the CH-BatteryGen benchmark (500 LFP vehicles, four fault classes, 10-second cadence), EVFaultNet achieves the highest mean battery-level F1 at two of three window lengths (0.903 at L=64, 0.895 at L=256) and is statistically comparable to the strongest baseline at L=128 (0.892 vs. 0.911, p=0.38), while using roughly four times fewer parameters. Ablations confirm that both the resting-state decay and periodic within-window time encoding contribute to the result. Battery-level aggregation resolves ambiguity for high resistance and self-discharge faults, but low-capacity detection remains the main limitation: 7.0% of healthy vehicles are false-flagged, at 0.500 precision and 0.933 recall. Measured workstation-GPU cost ranges from 0.57 to 2.42 ms per window; automotive embedded performance is not yet characterized.