Source-linked AI summary
Description and Discussion on DCASE 2022 Challenge Task 2: Unsupervised Anomalous Sound Detection for Machine Condition Monitoring Applying Domain Generalization Techniques
Kota Dohi, Keisuke Imoto, Noboru Harada, Daisuke Niizumi, Yuma Koizumi, Tomoya Nishida, Harsh Purohit, Takashi Endo, Masaaki Yamamoto, Yohei Kawaguchi
TL;DR
Domain shifts can degrade anomalous sound detection, while domain identification and frequent adaptation may be unavailable or costly. DCASE 2022 Task 2 evaluates domain generalization without test-time domain labels, and analysis of 81 submissions identified two broad approaches.
Problem
Domain shifts can degrade ASD, while real-world samples may lack usable domain labels and frequent model adaptation can be costly.
Source: §1 · p. 1Method
DCASE 2022 Task 2 evaluates domain-generalization techniques without test-time domain labels, providing shift-causing attributes in training data.
Source: §1 · p. 1Results
Analysis of 81 submissions from 31 teams identified domain-mixing and domain-classification approaches to domain generalization.
Source: §1 · p. 1Takeaways & Limitations
The task analysis identifies both approaches as useful for domain generalization, while leaving better synthetic representations and completely unseen target types for future study.
Source: §5 · p. 4Domain-classification approaches assume training target data includes all target-data types in testing, and may degrade when test data contains too many unseen types.
Source: §4.3 · p. 4, §4.1 · p. 4
Abstract
from arXiv · showhide
We present the task description and discussion on the results of the DCASE 2022 Challenge Task 2: ``Unsupervised anomalous sound detection (ASD) for machine condition monitoring applying domain generalization techniques''. Domain shifts are a critical problem for the application of ASD systems. Because domain shifts can change the acoustic characteristics of data, a model trained in a source domain performs poorly for a target domain. In DCASE 2021 Challenge Task 2, we organized an ASD task for handling domain shifts. In this task, it was assumed that the occurrences of domain shifts are known. However, in practice, the domain of each sample may not be given, and the domain shifts can occur implicitly. In 2022 Task 2, we focus on domain generalization techniques that detects anomalies regardless of the domain shifts. Specifically, the domain of each sample is not given in the test data and only one threshold is allowed for all domains. Analysis of 81 submissions from 31 teams revealed two remarkable types of domain generalization techniques: 1) domain-mixing-based approach that obtains generalized representations and 2) domain-classification-based approach that explicitly or implicitly classifies different domains to improve detection performance for each domain.
1. INTRODUCTION
Unsupervised ASD must contend with scarce anomalous training examples and domain shifts that can degrade detection when acoustic conditions change. DCASE 2022 Task 2 focuses on domain generalization without requiring test-time domain identification, and analysis of its submissions revealed two broad approaches.
1. INTRODUCTION
- Unsupervised ASD uses only normal sounds for training, addressing the scarcity of anomalous samples in both number and type.The paper describes this as a challenge to the application scope of ASD systems.Source: §1 · p. 1
- Domain shifts are acoustic differences between source and target data that can degrade detection by changing conditions unrelated to anomalies.The differences can arise from machine operating conditions or environmental noise.Source: §1 · p. 1
- The 2021 task assumed domain shifts were known in advance and infrequent, but real-world changes can be hard to identify and costly to adapt to repeatedly.Changes in background sound may have unclear causes, while machine operating conditions can change quickly.Source: §1 · p. 1
- DCASE 2022 Task 2 targets domain generalization without test-time domain detection or model adaptation, with shift-causing attributes provided in training data.The test data does not provide the domain of each sample.Source: §1 · p. 1
- Analysis of the challenge submissions identified domain-mixing, which seeks generalized representations, and domain-classification, which distinguishes domains to specialize detection.The paper reports 81 submissions from 31 teams.Source: §1 · p. 1
2. UNSUPERVISED ANOMALOUS SOUND DETECTION APPLYING DOMAIN GENERALIZATION TECHNIQUES
The section frames ASD as anomaly scoring with a threshold, then describes why domain shifts complicate detection and adaptation. DCASE 2022 Task 2 evaluates domain generalization with an undisclosed test domain and a single decision rule across source and target conditions.
2. UNSUPERVISED ANOMALOUS SOUND DETECTION APPLYING DOMAIN GENERALIZATION TECHNIQUES
- ASD determines machine state from an anomaly score, classifying a machine as anomalous when the score exceeds a predefined threshold.The score is calculated from an audio clip by a parameterized anomaly-score calculator.Source: §2 · p. 2
- Training only on normal sounds is a primary difficulty because anomalous sounds are rarely obtained in practice.Source: §2 · p. 2
- Domain shifts change data distributions across machine or environmental conditions, with the target domain having zero or few training clips.The paper distinguishes source and target distributions for both normal and anomalous conditions.Source: §2 · p. 2
- The 2021 task used domain adaptation for target-domain detection but assumed each sample's domain was known and shifts were not too frequent.Frequent shifts can make repeated model adaptation too costly, and domain detection may be difficult in practice.Source: §2 · p. 2
- Real-world shift scenarios include changing machine conditions, environmental noise, maintenance or parts replacement, and different recording devices.Adapting for each location or microphone can be costly.Source: §2 · p. 2
- Domain generalization aims to detect anomalies across domains with one threshold, without test-time shift detection or model adaptation.The paper presents it as a possible solution when shifts are difficult to detect or costly to adapt to.Source: §2 · p. 2
- In DCASE 2022 Task 2, test samples have no domain labels, and one anomaly score and threshold distinguish pooled normal and anomalous source- and target-domain data.The decision is designed not to be affected by differences between domains.Source: §2 · p. 2
3. TASK SETUP
The task setup specifies the dataset, evaluation metrics, and two baseline systems. It defines how scores are aggregated and how baseline anomaly decisions are made.
3.1. Dataset
- The dataset combines laboratory-recorded machine sounds with environmental noise recorded in real-world factories.Recordings are single-channel, 10 seconds long, and sampled at 16 kHz.Source: §3.1 · p. 2
- The development, additional-training, and evaluation datasets divide data into sections, with evaluation clips lacking condition labels and domain information.Development sections include source- and target-domain training clips and labeled test clips; evaluation sections contain 200 test clips each.Source: §3.1 · p. 2
3.2. Evaluation metrics
- AUC is calculated for each domain, while pAUC is calculated per section because detection must use the same threshold across domains.The pAUC uses the low-FPR range [0, 0.1], and the official score is the harmonic mean of AUC and pAUC over machine types and sections.Source: §3.2 · p. 3
3.3. Baseline systems and results
- The AE baseline uses reconstruction error, while MobileNetV2 predicts the section; both set the anomaly threshold using the 90th percentile of a fitted gamma distribution.The organizers report AUC and pAUC for both baselines, with averages and standard deviations over five independent trials.Source: §3.3 · p. 3, Table 1 · p. 3, Table 2 · p. 3
4. CHALLENGE RESULTS
Across the evaluation, domain-generalization submissions often beat the MobileNetV2 baseline, and target-domain performance was more closely associated with ranking than source-domain performance. The analysis distinguishes domain mixing from domain classification and notes a test-domain coverage condition for classification methods.
4.1. Results for evaluation dataset
- 22 of 81 teams outperformed the MobileNetV2 baseline in the official score, and target-domain AUC tracked top-team rank more closely than source-domain AUC.Among top teams, the reported rank correlations were −0.862 for target-domain AUC and −0.033 for source-domain AUC.Source: §4.1 · p. 3, Figure 1 · p. 4
4.2. Domain-mixing-based approach
- Domain mixing combines cross-domain batch data, Mixup, or augmentation to learn common representations, but these approaches were outperformed by domain classification.The authors suggest that synthesized data may not represent target-domain data meaningfully and identify improved synthetic representations, potentially using attribute information, as future work.Source: §4.2 · p. 4, Figure 2 · p. 4
4.3. Domain-classification-based approach
- Domain-classification methods distinguish domains explicitly or implicitly to specialize detection, and the analysis reports that they outperformed domain mixing.Their performance may degrade if test target-domain data contains too many types absent from training, so handling completely unseen target data needs further investigation.Source: §4.3 · p. 4
5. CONCLUSION
The paper summarizes DCASE 2022 Task 2 and the submitted solutions for domain generalization under implicit domain shifts. It identifies two approaches and distinct directions for future investigation.
5. CONCLUSION
- The challenge analysis identified domain mixing and domain classification as useful approaches, with future work aimed at meaningful synthetic representations and completely unseen target-domain types.For mixing, the authors point to meaningful synthetic representations; for classification, they call for analysis on target-domain types absent from training.Source: §5 · p. 4