Outlier Exposure-Based Anomalous Sound Detection Using Deep Learning

Abstract

Automatic Anomalous Sound Detection (ASD) systems play an important role nowadays. The objective is to automatically determine whether a machine, such as a slider, is making a normal, or an abnormal sound even if there are domain shifts such as different factory noise, or different recording conditions. The Outlier exposure approach is applied by incorporating in the training data outliers which are other machine types in order to expose the model to both samples of the target machine which is the slider, and outliers. To further improve the results of our proposed system, experiments were conducted including trying different activation functions, different loss functions, or adding a dropout layer in the Autoencoder with data augmentation. Furthermore, a Convolutional Autoencoder is implemented to replace the simple dense Autoencoder. We also tried using a pre-trained MobileNetV2 with data augmentation. Good performance has been achieved by using data augmentation techniques. The experiments' results were compared with the original autoencoder results in terms of the AUC.