Multi-Attention Enhanced Discriminator for GAN-Based Anomalous Sound Detection

Generative adversarial networks (GAN) have been regarded as promising for anomalous sound detection (ASD) by training an unsupervised one-class classifier to pick out the anomalous sample. Existing GAN-based anomaly detection models usually focus on the generator to reduce the reconstruction error. The generator even reconstructs anomalous samples effectively without learning their features, which increases the burden on the discriminator to differentiate between original and reconstructed samples. In this paper, we propose a multi-attention enhanced discriminator for GAN-based ASD named EDGAN. It integrates attention mechanisms from multiple dimensions into the discriminator to make it more effective. Besides, we incorporate the reconstruction error to devise a new abnormal score for efficient anomaly assessment. We conducted extensive experiments on the MIMII dataset, which illustrates the average AUC 8.29% improvement compared to the state-of-the-art methods.