Improving reliability of acoustic fault detection in harsh industrial environments via statistical modeling and temporal integration

Acoustic condition monitoring provides a scalable alternative to vibration sensing in large industrial facilities, but the use of microphones is challenged by disturbing sounds from co-located machinery, safety signals, and human activity, which often lead to high false-positive rates. In this work, we propose an anomalous sound detection (ASD) framework based on probabilistic anomaly scoring with Gaussian mixture models (GMM) and investigate three different feature extraction strategies: an interpretable audio feature vector, a Mel-spectrogram representation, and unsupervised features obtained from a convolutional variational autoencoder. To improve robustness against disturbing sounds, we introduce a preprocessing step that filters non-stationary disturbing signals and a post-processing step that incorporates temporal persistence of fault states using binary integration. The system is evaluated on both real recordings from a blast-furnace stock house and synthetic mixtures with controlled signal-to-noise ratios. Results demonstrate that the proposed preprocessing and temporal integration effectively suppress false detections and substantially improve robustness under realistic industrial conditions. For the best-performing feature representations, temporal integration yields false-positive rates (FPR) well below 1 %, while achieving very high discrimination performance with area under curve (AUC) values approaching 1.0 and F1 scores up to 1.0. These results demonstrate that explicitly modeling normal-state statistics and exploiting temporal persistence enables stable and controllable detection performance despite dynamically varying acoustic conditions.