[Paper Review] Description and Discussion on DCASE 2021 Challenge Task 2: Unsupervised Anomalous Sound Detection for Machine Condition Monitoring under Domain Shifted Conditions

Yohei Kawaguchi, Keisuke Imoto+7 more|arXiv (Cornell University)|Jun 8, 2021

Anomaly Detection Techniques and ApplicationsComputer Science15 references34 citations

TL;DR

This paper analyzes DCASE 2021 Task 2 on unsupervised anomalous sound detection under domain shift, surveys top approaches (OE+IM ensembles and IM-based feature detection), and discusses domain adaptation challenges.

ABSTRACT

We present the task description and discussion on the results of the DCASE 2021 Challenge Task 2. In 2020, we organized an unsupervised anomalous sound detection (ASD) task, identifying whether a given sound was normal or anomalous without anomalous training data. In 2021, we organized an advanced unsupervised ASD task under domain-shift conditions, which focuses on the inevitable problem of the practical use of ASD systems. The main challenge of this task is to detect unknown anomalous sounds where the acoustic characteristics of the training and testing samples are different, i.e., domain-shifted. This problem frequently occurs due to changes in seasons, manufactured products, and/or environmental noise. We received 75 submissions from 26 teams, and several novel approaches have been developed in this challenge. On the basis of the analysis of the evaluation results, we found that there are two types of remarkable approaches that TOP-5 winning teams adopted: 1) ensemble approaches of ``outlier exposure'' (OE)-based detectors and ``inlier modeling'' (IM)-based detectors and 2) approaches based on IM-based detection for features learned in a machine-identification task.

Motivation & Objective

• Motivate the need for unsupervised anomalous sound detection in real-world machines under domain drift (e.g., speed, noise, load).
• Describe task setup, datasets, and evaluation metrics for domain-shift ASD.
• Analyze submitted approaches and identify promising strategies for robust detection across domains.

Proposed method

• Define anomaly scoring framework with A_theta(x) and threshold phi for decisions.
• Summarize two main high-performing approaches: parallel-type ensembles of outlier exposure (OE) detectors and inlier modeling (IM) detectors, and serial-type IM detection on features learned in machine identification tasks.
• Discuss baseline systems: autoencoder-based anomaly scoring and OE-based MobileNetV2 with section-conditioned objectives.
• Analyze top submissions and their use of domain adaptation, ensemble strategies, and feature learning.
• Provide evaluation metrics including AUC and pAUC over source and target domains with harmonic official scores.

Experimental results

Research questions

• RQ1 How do domain shifts between training (source/target normal conditions) and testing (target anomalous conditions) affect ASD performance?
• RQ2 What combination of OE and IM techniques yields robust performance under domain shift?
• RQ3 Can IM-based detection on machine-identification features provide robust anomaly detection under drift?
• RQ4 What is the role of domain adaptation in improving ASD under domain-shift conditions?

Key findings

• Top-5 teams largely used two strategies: parallel-type hybrids (OE ensembles plus IM detectors) and serial-type hybrids (IM on features from machine-identification tasks).
• Ensembling OE and IM can improve robustness to domain shifts despite potential increases in model complexity.
• Domain adaptation applied to IM-based detectors, rather than feature extractors, can be effective when target-domain data are scarce.
• The approach without explicit domain adaptation can still yield strong target-domain performance in some cases, suggesting ensemble robustness.
• Two primary directions for future work: compact/domain-adaptable parallel hybrids and broader-domain adaptation for serial hybrids.