基于生成对抗单分类网络的异常声音检测

针对正常和异常声音可能具有较大的相似性, 有时无法利用自编码器重构误差大小区分的问题, 提出一种生成对抗单分类网络方法进行异常声音检测, 通过多次训练, 该方法学习正常样本的分布特征. 在测试过程中, 测试正常样本能以极小的误差进行重构, 而异常样本重构效果较差, 在某些频率段会发生畸变, 从而给出判别分类结果. 实验采用UrbanSound8K公开数据集和实测电机声音数据集进行了测试, 获得该方法的准确率分别为86.3%和98.1%, 比卷积自动编码器等主要深度学习方法分别提高了5.0%和3.0%.

Abnormal Sound Detection Based on Generative Adversarial Single Classification Network

Aiming at the problem that normal and abnormal sounds might have great similarity and sometimes could not distinguish normal and abnormal sounds by the size of reconstruction error of autoencoder, we proposed a generative adversarial single classification network method for abnormal sound detection. Through multiple training, the method learned the distribution characteristics of normal samples. In the test process, the normal sample could be reconstructed with minimal error, while the abnormal sample had poor reconstruction effect, and distortion occurred in some frequency bands, so as to give the discriminant classification results. In the experiment, UrbanSound8K public data set and measured motor sound data set were used for testing, and the accuracy of this method is 86.3% and 98.1%, respectively, which is 5.0% and 3.0% higher than main deep learning methods such as convolutional autoencoder.