Learning Disentangled Representations and Temporal-Correlation Dynamics for Robotic Anomaly Diagnosis

Anomalous diagnosis is valuable for reducing potential damages in long-term autonomy robot manipulation tasks, especially in Human-robot collaboration scenarios. Deep learning-based methods have been widely investigated for robot anomaly diagnosis, which can effectively encode complex dynamics from multi-modal sensory data. However, the lacking of enough anomalous samples and the fusion of high-dimensional and modality correlation as well as time-dependent is still a challenging problem. In this paper, a novel framework is introduced to generate synthetic anomaly samples for data augmentation by learning the disentangled representation with sequential disentangled variational autoencoder (sDVAE), and a temporal-correlation VAE (tcVAE) model for robot anomaly diagnosis by learning the temporal correlation features of multimodal anomalies. To evaluate the proposed methods, 115 original anomalous samples from 7 representative anomalies that are first recorded on a self-developed human-robot kitting task. Results indicate that the proposed methods show the best performance of the highest precision (97%), f1-score (95%), and accuracy (93%) with synthetic samples across all baseline methods.