Fault diagnosis of polar buoy energy systems using an Arrhenius-corrected second-order RC model and deep learning

Abstract Polar ice–sea amphibious ecological observation buoys are prone to energy system faults under harsh operating conditions, including low temperatures, low-irradiance conditions, and long-term unattended operation. Faults such as ohmic resistance ( R 0 ) increase, capacity degradation, leakage fault, and polarization-resistance increase can reduce sensor uptime, interrupt data transmission, and cause missing observation data. In this study, we develop a temperature-dependent second-order resistor–capacitor (RC) model for a 4S40P LiFePO 4 battery pack using Arrhenius-based temperature correction to address the scarcity of field fault samples and labeled fault data. Representative load profiles are then introduced into the developed model, and fault injection is used to generate time-series data under different operating states. On this basis, a 1D-CNN-BiGRU fault diagnosis model is constructed using terminal voltage ( V ), terminal current ( I ) and their first-order differences ( dV ) and ( dI ) as four-channel time-series inputs. Simulation results from five independent training runs show that the 1D-CNN-BiGRU achieves an accuracy of 0.9593 ± 0.0076, a Macro- F 1 score of 0.9594 ± 0.0076, and a weighted- F 1 score of 0.9593 ± 0.0076. These results indicate consistently high classification performance across the five operating states—normal, R 0 increase, capacity degradation, leakage fault, and polarization resistance increase. Robustness analysis further shows that uncertainty-aware training with bounded temperature-dependent voltage perturbations substantially improves classification robustness under moderate battery-model uncertainty, although severe low-temperature uncertainty remains challenging. Overall, this study provides a simulation-based framework for fault diagnosis and robustness evaluation in unattended polar buoy energy systems.