On the Choice of Score Aggregation Functions for Anomalous Sound Detection

Abstract

The standard autoencoder (AE) baseline for anomalous sound detection aggregates frame-level reconstruction errors by their mean, which can dilute localized anomalies in non-stationary machine sounds. We systematically compare four aggregation functions - mean, max, percentile-95, and variance - across all four machine types in the MIMII dataset at 6 dB SNR. Variance aggregation achieves the highest AUC for both non-stationary types (valve: 0.854, slider: 0.923), while mean aggregation remains best for both stationary types (fan: 0.918, pump: 0.842). The ranking of aggregation functions approximately reverses between the two groups, with variance and mean swapping positions at both ends. We further show that detection performance is more sensitive to the choice of evaluation-time aggregation than to the training loss function, suggesting that a standard mean-trained AE combined with signal-dependent aggregation is a practical and effective strategy.