Comprehensive feature integrated capsule network for Machinery fault diagnosis

Deep transfer learning is widely used for intelligent fault diagnosis due to its advantage of transferring knowledge already learnt. However, existing models still suffer from following issues: the failure to fully consider the effective information implied by vibration signals causes the unrepresentative feature extraction; the excessive focus on transferable features results in a reduction of feature discriminability; the selection of optimal diagnosis model totally depends on the iterative number or minimum loss. Our study proposes a comprehensive feature integrated capsule network (CFICN) to tackle these issues. Specifically, a multiscale jointing 1D-2D convolution is first presented to excavate the close dependency among different local features, in which the advantages of both 1D and 2D convolutions in feature extraction are fully utilized to obtain representative features. Then, a spectral penalization strategy is built to acquire comprehensive features by conducting an optimal balance between transferability and discriminability of representative features. Furthermore, comprehensive features are integrated into capsule network for the effective information fusion. Finally, a strategy for screening an optimal diagnosis model is constructed by defining a cumulative confidence index. Fourteen kinds of transfer tasks in two cases validate that our CFICN has an obvious advantage over five advanced methods for machinery fault diagnosis.