Hybrid physics-embedded recurrent neural networks for fault diagnosis under time-varying conditions based on multivariate proprioceptive signals

Accurate fault diagnosis for industrial robots is imperative to improve their availability. Proprioceptive signals collected by intrinsic sensors of robot joint servo drive systems provide a nonintrusive and promising way for practical in-situ diagnosis. However, they generally exhibit significant non-stationarity owing to time-varying operation conditions and limited sampling frequencies constrained by system hardware, which poses challenges in fault signature identification. Thus, a hybrid physics-embedded recurrent neural network is proposed for robot fault diagnosis under variable operation conditions based on proprioceptive signals. It embeds robot governing ordinary differential equations (ODE) as an inductive bias to account for known dynamics. Concurrently, tailored neural networks (NN) are leveraged to compensate for unmodeled dynamics residuum and unmeasurable health states, efficiently extending the hypothesis space. Hereinto, system status-represented latent space inferred from observations is comprehensively regularized by state reconstruction, fault classification, and Fisher discrimination losses to promote state representability and class distinguishability. Furthermore, a bilinear layer-based NN is constructed to statistically model intrinsic nonlinearities simplified away by physical models. Finally, the model-based and data-driven components are synergistically integrated by a differentiable ODE solver to form an end-to-end trainable framework. The superiority of the presented method is illustrated through the simulated and in-situ industrial robot datasets.