On choosing decision thresholds for anomalous sound detection in machine condition monitoring

Most anomalous sound detection (ASD) systems output a score for each audio sample presented to the system. Ideally, these anomaly scores differ for normal and anomalous samples such that one can determine whether a given sample is normal or anomalous by comparing the scores to predefined thresholds. However, determining these thresholds is non-trivial, especially when no anomalous samples are provided as training data. In this work, several methods for finding such decision thresholds are evaluated and compared to each other when acoustically monitoring the condition of machines in noisy environments. To this end, the state-of-the-art in ASD for machine condition monitoring will be reviewed first. Using a state-of-the-art ASD system, experimental evaluations are conducted on the DCASE 2020 ASD dataset to evaluate differently attained decision thresholds.