Machine Anomalous Sound Detection Based on Feature Fusion and Gaussian Mixture Model

Anomalous sound detection (ASD) is a technique used for audio monitoring of machines in factories, aiming to identify machine failures by analyzing the sound produced during machine operation. However, it is challenging to extract efficient input features for classifiers and determine suitable decision boundary, which can result in poor detection performance. This paper proposes a self-supervised method based on feature fusion and Gaussian mixture models (GMM). The feature fusion method is proposed to combine the temporal and frequency domain information. Moreover, the GMM-based method is proposed to calculate anomaly scores, which can determines more complex decision boundary. We conducted experiments on the datasets provided by DCASE 2020 Challenge Task 2. The results demonstrate that our model outperforms the baseline autoencoder model. Based on the official evaluation metrics, our model achieved an average AUC improvement of over 20.32 \(\%\) and an average pAUC improvement of over 27.29 \(\%\) .