Fig 1.
Architecture and training workflow of the Physics-Informed Self-Supervised Diagnosis (PI-SSD) framework.
Table 1.
PI-SSD architecture, training configuration, and hyperparameter summary.
Table 2.
Summary of datasets used for training and evaluation.
Table 3.
Performance comparison on NASA Gearbox and Aalto Rotor Datasets (mean ± 95% CI).
Table 4.
Effect of hyperparameter variations on PI-SSD performance (mean ± 95% CI).
Table 5.
Cross-domain generalization results (Train on NASA, Test on Aalto Rotor, No Fine-Tuning).
Table 6.
Ablation study: Full Model vs. Component Variants (mean ± 95% CI).
Table 7.
Physics consistency breakdown by operating speed.
Table 8.
Calibration metrics on NASA Gearbox and Aalto Rotor datasets.
Fig 2.
Violin plots of physics residuals (PR-MSE) for healthy vs. faulty windows comparing Full PI-SSD and –Physics variant.
Fig 3.
Training curves (loss, PR-MSE, AUROC) for benchmark models over 50 epochs.
PI-SSD consistently achieves the best balance between constraint satisfaction and classification.