An interpretable capsule network for intelligent fault detection of industrial robots

Deep learning exhibits notable advantages in intelligent fault detection tasks due to its robust end-to-end mapping capabilities. However, challenges associated with interpreting the internal mechanisms of such models limit their practical application in industrial settings. To address this issue, an interpretable capsule network (ICaps-Net) is proposed in this study for intelligent fault detection in industrial robots. Initially, the model employs convolutional layers to extract multi-dimensional features from monitored vibration signals, capitalizing on their strong feature extraction capabilities. Subsequently, a two-layer stacked capsule architecture, utilizing a dynamic routing algorithm, is applied to further analyze fault-related features and perform classification to generate detection results. Furthermore, an interpretability analysis is conducted based on the coupling matrix obtained during the dynamic routing process. This analysis uses a backward tracking strategy to identify the most discriminative features, thereby elucidating the underlying rationale of the model's decision-making process. The performance of the proposed ICaps-Net is evaluated using a three-joint fault experiment on an industrial robot. Experimental results demonstrate that the method achieves high-precision fault detection, effectively identifies key discriminative features, enhances the interpretability of the decision-making process, and improves the overall reliability of the diagnostic outcomes.