First-Shot Machine Anomalous Sound Detection: Exploring the Impact of Limited Target Dataset Size on Model Performance

Anomalous Sound Detection (ASD) refers to the process of identifying unusual or abnormal sounds in a system or environment, typically to detect potential faults or irregularities in machines or processes. This paper explores the challenges of ASD in machines, where collecting extensive data from every individual machine is often impractical, costly, and inefficient. First-shot ASD reduces the need for extensive data collection by leveraging comprehensive data from a single machine type and a limited number of samples from additional target machines to improve detection. While existing literature explores various models for anomalous sound detection using a target size of 10, there is little research examining the impact of varying this target size on model performance. We conducted novel experiments to evaluate the impact of varying dataset sizes for target machines. Four state-of-the-art ASD models were tested using target dataset sizes of 2, 5, and 10, leveraging the DCASE 2023 Task 2 dataset. Our results revealed that the area under curve (AUC) on the target set increased by up to 16% when the target size was expanded from 2 to 10, underscoring the importance of models capable of improved generalisation. This finding highlights the critical need for ASD models that can effectively learn from limited data.