Activity-Guided Industrial Anomalous Sound Detection against Interferences

TL;DR

This work tackles industrial anomaly detection under interference from neighboring machines by proposing SSAD, a framework that first performs activity-guided source separation and then applies anomaly detection using two-step masking. The approach leverages machine activity signals to both improve separation (via Informed X-UMX) and focus detection on active segments (via masked auto-encoders and a masked anomaly score). Empirical results on the MIMII dataset show SSAD can approach oracle performance using only corrupted mixtures plus activity information, with notable gains under interference and with multiple sources. The method demonstrates practical potential for robust, single-channel industrial monitoring by integrating separation and detection with activity-informed cues and masking strategies.

Abstract

We address a practical scenario of anomaly detection for industrial sound data, where the sound of a target machine is corrupted by background noise and interference from neighboring machines. Overcoming this challenge is difficult since the interference is often virtually indistinguishable from the target machine without additional information. To address the issue, we propose SSAD, a framework of source separation (SS) followed by anomaly detection (AD), which leverages machine activity information, often readily available in practical settings. SSAD consists of two components: (i) activity-informed SS, enabling effective source separation even given interference with similar timbre, and (ii) two-step masking, robustifying anomaly detection by emphasizing anomalies aligned with the machine activity. Our experiments demonstrate that SSAD achieves comparable accuracy to a baseline with full access to clean signals, while SSAD is provided only a corrupted signal and activity information. In addition, thanks to the activity-informed SS and AD with the two-step masking, SSAD outperforms standard approaches, particularly in cases with interference. It highlights the practical efficacy of SSAD in addressing the complexities of anomaly detection in industrial sound data.

Figures (8)

• Figure 1: The target situation where the machine sounds with similar acoustic features are recorded by a single-channel recording device.
• Figure 2: An illustration of SSAD framework for anomaly detection of two machines, depicting the activity signals for source separation in green and the activity signals for two-step masking in blue.
• Figure 3: Model architecture of the informed X-UMX, where the differences from the original X-UMX model is highlighted in blue. The model separates the $k$ number of sources based on the provided mixture $x$ and the activity signal $a_k$.
• Figure 4: The baseline anomaly detection model of a mixture using reconstruction error.
• Figure 5: Anomaly score distribution of valve data in normal and abnormal data. Minimum, mean, and maximum values are marked.