Physics-embedded variational autoencoder with diffusion-based uncertainty estimation for trustworthy fault diagnosis

Effective fault diagnosis for industrial robots is imperative to improve their reliability and availability in safety-critical applications. Proprioceptive signals from servo drive systems provide a nonintrusive and promising way for in-situ diagnosis. However, their exhibited pronounced nonstationarity induced by time-varying operating conditions poses challenges in fault signature identification. Concurrently, existing methods generally struggle with integrating prior knowledge and yielding calibrated diagnostic results with associated uncertainty estimates. Thereupon, a physics-embedded variational autoencoder with diffusion-based uncertainty estimation is proposed for trustworthy fault diagnosis based on robot proprioceptive signals under nonstationary conditions. Specifically, the intrinsic system health status is identified and encoded into the latent space to decouple from the nonstationary exogenous inference. The physics-embedded decoder incorporates the robot governing ordinary differential equation as the inductive bias and leverages expressive networks to compensate for unmodeled dynamics. It further regularizes the latent space via modeling the inherent interaction among latent representations and system observations. The diffusion-based probabilistic classifier is built on the latent space for implicit predictive distribution recovery and calibration improvement via uncertainty quantification. The derived variational evidence lower bound provides theoretical ground for synergetic model optimization. The superiority of the presented method is validated on an in-situ industrial robot dataset.