# Physics-informed self-supervised diagnosis of rotating machinery using latent ODEs and transformer encoders

## TL;DR

This paper introduces a new framework for detecting faults in rotating machinery by combining physics-based models with self-supervised learning, achieving high accuracy and reliability.

## Contribution

The novel PI-SSD framework integrates latent ODEs and Transformer encoders with physics priors for self-supervised fault diagnosis in rotating machinery.

## Key findings

- PI-SSD achieves a Macro-F1 score of 0.91 on the NASA PHM’09 dataset with strong domain transfer performance on Aalto.
- The framework maintains low Expected Calibration Error (ECE = 0.022) and shows robustness across varying speeds.
- Ablation studies confirm the effectiveness of each component in the PI-SSD architecture.

## Abstract

This paper proposes a novel Physics-Informed Self-Supervised Diagnosis (PI-SSD) framework for rotating machinery fault detection, combining physical modeling, self-supervised representation learning, and uncertainty-aware classification. The architecture integrates a multi-resolution convolutional encoder, a windowed Transformer for temporal context modeling, and a latent neural ordinary differential equation (ODE) module that embeds mechanical priors, such as Jeffcott rotor dynamics, directly into the learning process. A masked segment reconstruction objective enables self-supervised pretraining using unlabeled healthy signals, while an evidential classifier head produces fault probabilities with calibrated uncertainty. We evaluate PI-SSD on two publicly available datasets, the NASA PHM’09 Gearbox dataset and the Aalto Rotor dataset, covering 6 fault types and over 5,500 multichannel vibration recordings. Compared to seven strong baselines, PI-SSD achieves the highest Macro-F1 score (0.91) and lowest Expected Calibration Error (ECE = 0.022) on the NASA dataset, while maintaining strong domain transfer performance on Aalto (Macro-F1 = 0.81, PR-MSE = 0.067) without fine-tuning. Ablation studies confirm the contribution of each component, and physics consistency analysis demonstrates low violation rates under varying speeds. These results highlight the potential of embedding physics knowledge into self-supervised neural systems for robust, interpretable, and transferable fault diagnosis in industrial applications.