Sound-based anomaly detection using a locally constrained capsule network

In this study, we propose a new technique for abnormal acoustic diagnosis based on a locally constrained capsule network architecture. The newly designed anomaly diagnosis capsule network simplifies the learning network by removing auxiliary layers, and reduces computation by minimizing the operation between capsule layers. When compared with those of other network models utilized for abnormality diagnosis, the accuracy, area under the curve, and F1-score of the proposed model were sufficient.