Implicit–Explicit Collaborative Cross-Stitch Network With Multitask Disentanglement Learning for Fault Diagnosis and Size Estimation

Traditional fault diagnosis methods may suffer from the following shortcomings: 1) they are designed to identify fault types, overlooking the estimation of fault size which is equally important and 2) they typically learn features using only either implicit or explicit learning without fully exploiting their collaborative potential, which deepens the black-box nature or leads to excessive reliance on rules. To address these issues, we propose the implicit–explicit collaborative cross-stitch network for simultaneous fault diagnosis and size estimation. First, to robustly mine the representational features from the original graph constructed from data, the implicit feature learning is performed on the multigranularity subviews with a selective parameter-sharing graph attention network (GAN). Then, the refined self-starting cross-stitch network (RSCS) is employed for explicit characterization and interaction, helping to fully leverage the complementarity between multiple tasks. After that, the disentanglement learning with joint training loss term is introduced to transform the individual tasks into collective representations, and enhance the cooperativity between task-specific and task-shared features for simultaneous fault diagnosis and size estimation. Finally, the experimental results demonstrate that our approach effectively performs fault diagnosis and size estimation simultaneously, which shows promising practicality.