Improving Unsupervised Anomalous Sound Detection Performance of Autoencoder and Its Variant with Pretrained Deep Belief Network

Pretrained deep belief networks have been shown to speed up the convergence of models and improve their performance on many supervised recognition tasks. However, its performance on unsupervised anomalous sound detection remains to be explored. In this paper, we initialize the parameters of the autoencoder (AE) and its variant with the parameters of pretrained deep belief network (DBN) and use them for unsupervised anomalous sound detection. We explore the effect of the number of layers initialized with pretrained parameters on the detection performance and the effect of pre-training the DBN with different data on the detection performance. Experimental results show that the appropriate number of layers initialized with pretrained parameters can substantially improve the anomaly detection performance of the AE and its variant. In addition, the experimental results also show that using data that is quite different from the anomaly detection experimental data as the pre-training data of the DBN can also improve the detection results. Specifically, the experimental results show that the proposed approach achieves a 9.6% improvement based on the standard AUC score.