Author(s): Pengcheng Li; Cheng Lv; Meng Cui; Zhibing Liu; Jun Wu; Gang Wu
Linked Author(s):
Keywords: Polar ships; Fault diagnosis; Deep belief network (DBN); Explainable AI (XAI); Thermal dynamics
Abstract: The extreme polar marine environment, characterized by ultra-low temperatures and sea spray icing, severely compromises the operational reliability of deck machinery. While intelligent fault diagnosis is critical for navigational safety, traditional multi-sensor frameworks lack robustness in freezing conditions, and "black-box" deep learning algorithms fail to earn seafarers' trust due to a lack of physical interpretability. To address these challenges, this study proposes a novel, interpretable state recognition framework utilizing pure temperature time-series data and a Knowledge-Enhanced Deep Belief Network (DBN). Based on rigorous environmental chamber experiments (down to -60 ℃) on electro-hydraulic remote control valves and LED lighting fixtures, dynamic thermal features—such as temporal temperature gradients and spatial thermal differentials—were extracted to proxy mechanical load variations and icing insulation mechanisms. Furthermore, a knowledge extraction algorithm was developed to translate hidden neural activations into explicit, thermodynamically sound symbolic rules. The model achieved a diagnostic accuracy exceeding 96% in identifying complex degradation states, including hydraulic sluggishness at -40 ℃ and heater failure at -20 ℃. By successfully integrating physical mechanism features into the AI model, this research not only ensures diagnostic robustness in harsh environments but also provides transparent decision logic, offering vital theoretical guidance and practical reference for the winterization design and intelligent maintenance of polar ships.
Year: 2026