[1Yin-A] Poster 1

[1Yin-A-31] Score Refinement for Memory-bank-based Anomalous Sound Detection Using a Few Anomalous

Takahiro Sawatari¹, 〇Yuto Watanabe¹, Tatsuya Maehashi¹ (1. Suzuki Motor Corporation)

Keywords:

Anomaly sound detection,Mechanical noise,Anomaly detection,manufacturing industry,mobility

Memory-bank-based approaches are widely adopted in anomalous sound detection due to their high detection performance and low computational cost. However, most existing methods typically model only the distribution of normal data and lack a mechanism to effectively leverage the limited anomalous samples available during operation. While some prior works utilize few-shot anomalous data, they often require model retraining, incurring additional computational costs and complex loss designs. To address these limitations, we propose a novel, training-free score refinement method. We incorporate an additional anomalous memory bank, built from a few anomalous samples in a pre-trained feature space, into memory-bank-based models with only normal data. During inference, the final anomaly score is derived by subtracting the distance to the anomalous memory bank from the distance to the normal memory bank. This scoring mechanism amplifies the scores of samples similar to known anomalies while suppressing those of rare normal patterns, thereby reducing false positives. Experimental results on the DCASE2025 Challenge Task2 dataset demonstrate that the proposed method improves detection accuracy compared to baselines that rely solely on normal data.