A Multi-Scale Feature-Enhanced XGBoost Model for Fault Prediction in Shipboard Computing Systems

For intelligent modern vessels, the reliability of shipboard computing equipment is critical to maritime mission safety. To tackle fault prediction from massive multi - source heterogeneous monitoring data, this paper proposes an XGBoost - based fault prediction method for such equipment. After analyzing multi-dimensional monitoring data, 6 key performance indicators (KPIs) are identified. A feature engineering framework integrates 9-scale sliding window statistics and three-phase degradation modeling, yielding 630 features. An XGBoost classification model is constructed, employing a dual-level balancing strategy with noise-injected data augmentation and dynamically weighted loss to address class imbalance. On the Google cluster dataset, it achieves an AUC of 0.997, an accuracy of 96.1%, a recall of 100%, and a false positive rate of 1.7%, effectively supporting predictive maintenance of such equipment.