Fast Robust Capsule Network With Dynamic Pruning and Multiscale Mutual Information Maximization for Compound-Fault Diagnosis

Rotating machinery, such as ventilators and water pumps, are crucial components in modern industry, of which safety monitoring requires intelligent fault diagnosis. Feature representation learning is essential in the intelligent fault diagnosis of rotating machinery. In this study, a fast robust capsule network augmented with a dynamic pruning technique and a mutual information loss is proposed. The capsule layer overcomes limitations in pooling layers and scale-invariant feature transformation by learning tensor representations of features. The dynamic pruning method employs a dropout-like strategy to prevent repeated calculations and reduce the scale of parameters to simplify the network topology while increasing robustness. The enhanced agreement function limits the similarity of capsules in the same layer to avoid homogeneous features. The local and global discriminators are designed to learn and obtain mutual information in two aspects. The resulting multiscale mutual information loss for the proposed model successfully increases the model's representation learning capacity by integrating local and global information. The performance of the proposed method is successfully verified on several datasets with various noise levels obtained from a simulation platform.