Diffusion Augmentation Sub-center Modeling for Unsupervised Anomalous Sound Detection with Partially Attribute-Unavailable Conditions

Abstract

Current state-of-the-art unsupervised anomalous sound detection (ASD) methods typically rely on manually annotated attribute information as labels, employing auxiliary classification tasks to learn an embedding space for normal sounds, which helps detect anomalies deviating from this space. However, attribute information is often unavailable for certain machine types, making it difficult to learn the complex intra-class data distribution features of the same machine type. Additionally, limited sample diversity in the target domain further hinders learning robust discriminative features. To address these challenges, we propose a diffusion augmentation sub-center modeling (DASM) approach for embedding learning. This method employs iterative training of sub-center modeling, adaptive diffusion augmentation, and discriminative feature learning, utilizing a min-max optimization approach to maximize intra-class diversity and minimize intra-class distance, resulting in more expressive embeddings. Experimental results on the DCASE 2024 Challenge Task 2 dataset demonstrate that the proposed method significantly improves ASD performance.