Paper: Anomalous sound detection with complex-valued hybrid neural networks considering phase variations
Author: Shota Nishiyama, Akira Tamamori
Affiliation, Organization: Aichi Institute of Technology Graduate School of Business Administration and Computer Science (AIT)
keyword: Anomalous sound detection, complex-valued neural networks, phase variations
paper info: EA2022-106,SIP2022-150,SP2022-70
date of issue: 2023-02-21 (EA, SIP, SP)
Anomalous sound detection is the task of identifying whether an incoming mechanical sound is normal or anomalous. Since anomalous sounds occur infrequently and are highly diverse, it is treated as a problem of detecting anomalous sounds from normal sounds only. The acoustic features used as input to most anomalous sound detection models are mel-spectrogram. However, the phase variation is lost when the complex-spectrogram obtained by Fourier transforming the sound waveform is converted to the mel-spectrogram. In this study, we compare anomalous sound detection methods using complex-valued neural networks and real-valued neural networks to demonstrate the usefulness of phase variation. As a result of the comparison, there existed machine sounds for which phase variation was valuable and machine sounds for which it was not valuable. In this study, we propose a complex-valued hybrid neural network that combines a complex-valued module that preserves the structure of complex values and a real-valued module that takes mel-spectrogram as input for all feature extraction operations in which complex-spectrogram can be input in order to take phase variation into account. We propose a complex-valued hybrid neural network that combines a complex-valued structure-preserving module and a real-valued module that takes the mel-spectrogram as input for all feature extraction operations. Experiments verified the effectiveness of the proposed method on anomalous sound detection for multi-channel sound in the ToyADMOS dataset. Experimental results showed that the proposed method improved the average AUC of all machine sounds by around 3% compared to both complex-valued and real-valued neural networks.
cd environments/gpu
docker compose up -d
docker compose exec complex-toyadmos bash
poetry install
poetry run bash all_features_run.shNote
plase rewrite your environment -> configures/experiments/default.yaml
@techreport{weko_224446_1,
author = "西山,翔大 and 玉森,聡",
title = "位相変動を考慮した複素数値ハイブリッドニューラルワークによる異常音検知",
year = "2023",
institution = "愛知工業大学大学院経営情報科学研究科, 愛知工業大学大学院経営情報科学研究科",
number = "49",
month = "feb"
}