Uncertainty-aware fault diagnosis of unknown faults using ensemble-based NODE residuals

This paper presents an uncertainty-aware approach to fault diagnosis using ensembles of neural ordinary differential equations for residual generation. The proposed method leverages physical insights in the design of the NODE models and employs bootstrap ensembles to quantify prediction uncertainty and adapt detection thresholds. The adaptive thresholds enable robust detection and isolation of faults in scenarios with limited training data. The methodology is validated on real-world data from an internal combustion engine test bench, providing insights into how training data quality and residual design impact uncertainty-aware fault diagnosis.