부산대학교 도서관

To solve the low frequency, wide variety and changing problems in machine abnormal sound detection, we propose an unsupervised machine abnormal sound detection method based on residual autoencoder technology. This method can automatically learn and capture the characteristics of abnormal sounds without the need to mark a large number of abnormal samples in advance. First, Mel-frequency cepstral coefficients(MFCC) is used to construct the acoustic features of the normal sound, then residual autoencoder is used to reconstruct the acoustic features, and finally the anomaly fraction between the acoustic features reconstructed by the audio features to be measured and the acoustic features reconstructed by the normal sound is calculated. This method performs well in correctly identifying the abnormal sound.