An Effective Anomalous Sound Detection Method Based on Global and Local Attribute Mining

Recent anomalous sound detection (ASD) methods have achieved impressive performance by leveraging attribute information. However, domain shift and imbalance issues caused by inconsistent attributes between source and target domains remain a key challenge. In this paper, we propose a novel global and local attribute mining method for ASD (GLAM-ASD) with a Transformer based encoder consisting of teacher and student branches to exploit semantic-rich information. The student mines contextual information via masked patch reconstruction while the teacher obtains weights from the student via exponential moving average. Global and local attributes are then derived in an unsupervised way. The student branch is further optimized by global self-distillation and pseudo-label prediction via learned local attributes. Evaluations on the DCASE2023 Task2 benchmark demonstrate the superiority of GLAM-ASD, especially in its target domain generalization performance.