An Ensemble Machine Learning-Based Approach for Identifying New Fault Modes Beyond Known Classes

Purpose: Machine learning-based prognostics and health management (PHM) has been increasingly adopted for real-time fault diagnosis. However, models trained on normal data and a limited set of known faults often fail to recognize unseen fault modes during deployment in shipboard auxiliary rotating machinery. Methods: A 1D-CNN was used to classify three-axis vibration signals while extracting latent features and out-of-distribution (OOD) scores (MSP, Energy, and Entropy). These scores were fused and modeled using a one-class SVM (OC-SVM) to detect previously unseen fault modes. Results: Among the individual OOD scores, Entropy demonstrated the best performance, with accuracies of 0.93, 0.92, 0.87, and 0.82 for Cases 1-4, respectively. The OC-SVM ensemble incorporating latent features achieved the highest AUROC (≥ 0.92) and F1 scores across all cases for detecting unseen faults. Conclusion: By combining MSP, Energy, and Entropy scores with latent features extracted from a baseline classifier through an OC-SVM, the proposed framework enables robust detection of previously unseen fault modes in three-axis screw-pump vibration data. This approach improves the reliability of PHM systems and supports faster responses to unexpected failures.