Multi-channel Parallel Computing in Capsule Network and Its Application in Mechanical Fault Diagnosis

Due to complex noise environments, it is difficult to extract bearing fault features from vibration signal and it affects the accuracy of fault diagnosis. To solve the problem, this paper proposes a multi-channel parallel computing capsule network (MPCN), which simultaneously uses various scales kernels to extract the features from original signals. In MPCN, the vibration signal is directly input into the model, then various specifications kernels explore multiple aspects characteristic of signals. Finally, MPCN takes advantage of vector neurons to fuse multiple aspects characteristic. In order to verify the effectiveness of MPCN, fault diagnosis experiments were conducted under different noise conditions. The experimental results show that the accuracy of MPCN exceeds 95%, which is significantly better than traditional diagnostic methods.