基于注意力机制的无监督异常声音检测方法
Unsupervised Abnormal Sound Detection Method Based on Attention Mechanism
王超 1李敬兆 2张金伟2
作者信息
- 1. 安徽理工大学人工智能学院,安徽 淮南 232001
- 2. 安徽理工大学计算机科学与工程学院,安徽 淮南 232001
- 折叠
摘要
针对传统的基于自编码器的无监督异常声音检测方法存在特征表达能力不足的问题,提出一种基于注意力-跳跃自编码器-生成对抗网络的无监督异常声音检测方法ASAE-GAN(Atten-tional Skip-connected Auto Encoder and Generative Adversarial Network).ASAE-GAN在跳跃自编码器和生成对抗网络的基础上,引入通道间注意力机制和时间注意力机制,增强模型的特征表达能力.使用MIMII数据集中的pump声音数据进行实验,评价指标使用AUC分数.结果表明:ASAE-GAN的平均AUC分数相比较于AE、UNET和Skip-GANomaly分别提升了16.27%、14.23%和6.55%,验证了其在无监督异常声音检测方面的优越性.
Abstract
Aiming at the problem of the insufficient feature expression ability of the traditional unsupervised ab-normal sound detection method based on autoencoder,this paper proposes an unsupervised anomaly sound detec-tion method called ASAE-GAN(Attentional Skip-connected Auto Encoder and Generative Adversarial Network).Based on jump autoencoder and generative adversarial network,ASAE-GAN introduces interchannel attention mechanism and time attention mechanism to enhance the feature expression ability of the model.The experiment is conducted by using pump sound data from the MIMII dataset,and the evaluation index uses the AUC score.The results show that the average AUC score of ASAE-GAN is increased by 16.27%,14.23%and 6.55%com-pared with AE,UNET and Skip-GANomaly,respectively,which verifies its superiority in unsupervised abnormal sound detection.
关键词
自编码器/无监督/异常声音检测/生成对抗网络/注意力机制Key words
auto encoder/unsupervised/abnormal sound detection/generative adversarial networks/attention mechanism引用本文复制引用
基金项目
国家自然科学基金(51874010)
淮南市科技计划项目(2021A243)
出版年
2024
NETL
NSTL
万方数据