Toward Faithful Explanations in Acoustic Anomaly Detection

TL;DR

The paper tackles the need for faithful explanations in acoustic anomaly detection for industrial monitoring. It compares a standard autoencoder (AE) and a masked autoencoder (MAE) on spectrogram-based anomaly detection, evaluating multiple attribution methods to assess explanation quality. Using real wood planer data, the authors introduce a time-local F-score and a faithfulness metric based on replacing highlighted regions with reconstructions, and show that MAE yields more temporally precise and faithful explanations, with error maps being particularly effective. The results suggest that MAE-based training improves interpretability with minimal impact on detection performance, making it well-suited for safety-critical industrial monitoring where transparent decision-making is essential.

Abstract

Interpretability is essential for user trust in real-world anomaly detection applications. However, deep learning models, despite their strong performance, often lack transparency. In this work, we study the interpretability of autoencoder-based models for audio anomaly detection, by comparing a standard autoencoder (AE) with a mask autoencoder (MAE) in terms of detection performance and interpretability. We applied several attribution methods, including error maps, saliency maps, SmoothGrad, Integrated Gradients, GradSHAP, and Grad-CAM. Although MAE shows a slightly lower detection, it consistently provides more faithful and temporally precise explanations, suggesting a better alignment with true anomalies. To assess the relevance of the regions highlighted by the explanation method, we propose a perturbation-based faithfulness metric that replaces them with their reconstructions to simulate normal input. Our findings, based on experiments in a real industrial scenario, highlight the importance of incorporating interpretability into anomaly detection pipelines and show that masked training improves explanation quality without compromising performance.

Figures (5)

• Figure 1: Example of a human annotation for a broken board. The blue shaded region indicates the labeled anomaly interval.
• Figure 2: Segment-based input replacement. Gray areas indicate regions selected by the MAE error map at the 98th percentile.
• Figure 3: Qualitative comparison of interpretability methods on a broken board anomaly. (a) shows the input spectrogram, with the input overlaid by human-annotated anomaly intervals. (b) and (c) display the raw 2D attribution maps from six methods, following their corresponding binarized masks at the 98th percentile, and the last row presents 1D temporal attribution signals collapsed over frequency at the 98th percentile, with red dots marking detected peaks and shaded regions indicating annotated anomalies, for the AE and MAE, respectively.
• Figure 4: F-score across different percentile thresholds (90th–99th) for all interpretability methods applied to AE and MAE models. The highest score for each method is marked with an “x”.
• Figure 5: Faithfulness scores across percentile thresholds (90th–99th).