Physical validity of VAE-generated anomaly sequences

Determine whether Variational Autoencoder-generated anomaly sequences fail to represent physically compliant kinematic anomalies and thereby cause the TSPulse Teacher to propagate erroneous confidence scores that corrupt the MiniRocket Student’s learned decision boundary.

Background

The paper compares four augmentation regimes for distilling anomaly scores from the TSPulse time-series foundation model into a MiniRocket regressor: no augmentation, statistical fault injection, Variational Autoencoder generation, and a hybrid method. Although the injection method performs best empirically, the VAE and hybrid methods do not meaningfully improve Student performance.

The authors do not establish the causal explanation for this result. They hypothesize that VAE-generated sequences may violate the physical constraints of robotic kinematics, causing TSPulse to assign unreliable anomaly scores and consequently degrading the Student’s learned decision boundary. Validating this explanation would require assessing the physical realism of generated anomalies and the reliability of the Teacher’s labels on them.