Unsupervised anomalous sound detection method based on Gammatone spectrogram and adversarial autoencoder with attention mechanism

Abstract

Anomalous sound detection (ASD) is an important technology in the fourth industrial revolution, which can monitor the abnormal state of machine by identifying whether the sound of the machine is normal or not. However, in practical applications where there are few anomalous sound samples from machines, achieving effective ASD is still a challenge. In this paper, an unsupervised ASD algorithm based on adversarial autoencoder with attention mechanism is proposed. Different from the traditional reconstruction-based ASD model, in order to make the features learned by the model more representative, complex sound timing signals are converted into Gammatone spectrogram with richer features through filtering. Then the spectrogram is used as the input of the convolutional autoencoder. At the same time, the attention mechanism is introduced in the encoder to enhance adaptive learning of the normal patterns. Then the discriminator is used in the generative adversarial network to perform adversarial learning with the improved convolutional autoencoder to improve the reconstruction ability of the model for normal samples. Experimental results demonstrate that the proposed algorithm significantly outperforms commonly used industry methods for anomaly detection and exhibits advantages over other deep learning approaches in terms of system complexity.