Self-Supervised Augmented Diffusion Model for Anomalous Sound Detection

Generative models have significantly enhanced the capability of unsupervised anomalous sound detection (ASD) with their strong data modeling capabilities. However, many existing ASD methods based on generative models focus solely on accurately reconstructing sound data itself, neglecting the use of metadata. This results in limited features learned by the models. Additionally, these methods suffer from issues such as low generation quality and mode collapse. To address these challenges, we propose a self-supervised augmented diffusion model (SSDM) to improve ASD performance. SSDM learns expressive embeddings through a self-supervised learning module with a dual-path time-frequency self-attention ASD framework and then uses a denoising diffusion module to learn the distribution of these embeddings as the basis for anomaly detection. The reconstruction loss, which is derived from the reconstruction of test data embeddings measured by the self-supervised learning module in the denoising diffusion module, is used as the anomaly score. Experiments on the DCASE Challenge 2023 Task 2 development dataset demonstrate the effectiveness of the proposed method.