| Presentation | 2021-08-23 A study on transfer learning in unsupervised anomalous sound detection based on deep metric learning considering variance of normal data Hiroki Narita, Akira Tamamori, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | In recent years, anomaly detection research in the field of computer vision has focused on methods based on transfer learning. Transfer learning can significantly reduce the training time of deep learning models for high-dimensional media data such as images, and has the advantage that anomaly detection models can be operated even when there is little training data. However, there have been few studies on the use of pre-trained models for acoustic anomaly detection, and the design of the model structure and loss function has not been investigated. In this study, we propose a method to improve the accuracy of anomalous sound detection by suppressing the variance of normal data extracted by the pre-trained model using distance learning. In the proposed method, acoustic features are extracted from the middle layer of ResNet, which is pre-trained with acoustic data, and are trained and inferred by the distance learning model in the later stage to detect anomlous sounds. We also applied feature extraction by transfer learning to the DCASE Challenge2020 Task2 dataset, and conducted evaluation experiments with several anomaly detection models including the proposed method. The experimental results show that the proposed method is more effective than the variational autoencoders and the Mahalanobis distance-based anomaly detection methods in achieving higher detection accuracy. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Anomalous sound detection / Unsupervised learning / Transfer learning / Deep metric Learning |
| Paper # | SIP2021-28 |
| Date of Issue | 2021-08-16 (SIP) |
| Conference Information | |
| Committee | SIP |
|---|---|
| Conference Date | 2021/8/23(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Online |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Yukihiro Bandou(NTT) |
| Vice Chair | Toshihisa Tanaka(Tokyo Univ. Agri.&Tech.) / Takayuki Nakachi(Ryukyu Univ.) |
| Secretary | Toshihisa Tanaka(Xiaomi) / Takayuki Nakachi(Takushoku Univ.) |
| Assistant | Taichi Yoshida(UEC) / Seisuke Kyochi(Univ. of Kitakyushu) |
| Paper Information | |
| Registration To | Technical Committee on Signal Processing |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | A study on transfer learning in unsupervised anomalous sound detection based on deep metric learning considering variance of normal data |
| Sub Title (in English) | |
| Keyword(1) | Anomalous sound detection |
| Keyword(2) | Unsupervised learning |
| Keyword(3) | Transfer learning |
| Keyword(4) | Deep metric Learning |
| 1st Author's Name | Hiroki Narita |
| 1st Author's Affiliation | Aichi Institute Of Technology(AIT) |
| 2nd Author's Name | Akira Tamamori |
| 2nd Author's Affiliation | Aichi Institute Of Technology(AIT) |
| Date | 2021-08-23 |
| Paper # | SIP2021-28 |
| Volume (vol) | vol.121 |
| Number (no) | SIP-144 |
| Page | pp.pp.5-10(SIP), |
| #Pages | 6 |
| Date of Issue | 2021-08-16 (SIP) |