Acoustic Feature based Abnormal Diagnosis Techniques with Capsule Networks
Nac‐Woo Kim, Hyun-Yong Lee, Sangjun Park, Jungi Lee, Byung‐Tak Lee · 2020
This paper proposes a new method of converting acoustical data into mel-frequency cepstral coefficient feature vector and performing detection of abnormal acoustics using newly designed capsule network. Characteristic performance tests for extracting optimal acoustic characteristics through mel-frequency cepstral coefficient are first performed, and abnormal acoustic feature extraction model performance using capsule network is tested. Capsule network requires relatively fewer training sets compared to the convolutional neural network model, but it has equivariance properties for learning data and has strong characteristics for affine transformations. Our abnormal detection-based capsule network model shows superior performance compared to other abnormal detection models in terms of accuracy.