Abstract

Shipboard lithium-ion battery systems experience continuous mechanical vibration, yet model-based and data-driven internal short-circuit (ISC) detectors have not been compared under such conditions. We present, to our knowledge, the first unified vibration-aware ISC benchmark: a model-based Extended Kalman Filter (EKF) ΔSOC rule and three convolutional detectors—ModernTCN, LITE, and NPU-Conv2D—are evaluated on a simulated NCR18650PF module under quiescent, MIL-STD-810H-derived, and head-sea vibration, with vibration coupled to cell resistance through a phenomenological assumption ΔR=kR|a|. The EKF observes the terminal voltage alone, whereas the data-driven detectors additionally observe cell temperature, so the comparison couples detector class with input observability. Under this assumed envelope and the swept coupling range, the dual-channel data-driven configurations pass 135/135 deadline-scored outcomes against 129/135 for the EKF, with zero pre-onset false alarms versus 15, and their advantage lies in detection-delay dispersion rather than in mean latency. Deployed on an STM32N6 microcontroller, the INT8 NPU-Conv2D completes one inference in 0.752 ms, 179× lower latency than the EKF firmware. Calibration-free robustness emerges as the practical advantage of the evaluated dual-channel detectors.

IPC Classification

Keywords