PaperlayerView exact revision on arXiv

Source-linked AI summary

MIMII DG: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection for Domain Generalization Task

Kota Dohi, Tomoya Nishida, Harsh Purohit, Ryo Tanabe, Takashi Endo, Masaaki Yamamoto, Yuki Nikaido, Yohei Kawaguchi

arXiv:2205.13879v2cs.SDcs.AIcs.LGeess.AS

TL;DR

Domain shifts can degrade anomalous sound detection, and some shifts are difficult to detect or adapt to. MIMII DG is a five-machine dataset built to benchmark domain-generalization methods across varied shift conditions. Baseline experiments show lower target-domain AUC on average, suggesting that the dataset reproduces domain shifts and is useful for benchmarking.

  • Problem

    Domain shifts in normal machine sounds can degrade ASD performance, and shifts such as background-noise changes may be difficult to detect or adapt to.

    Source: §1 · p. 1

  • Method

    MIMII DG is a dataset for ASD domain-generalization benchmarking, with five machine types and three shift scenarios per type.

    Source: §1 · p. 1, §2 · p. 2

  • Results

    Target-domain AUC was lower on average than source-domain AUC by 13.2% with the autoencoder-based system and 12.3% with the MobileNetV2-based system, suggesting that the dataset reproduces domain shifts.

    Source: §4.3 · p. 4

  • Takeaways & Limitations

    The authors conclude that MIMII DG is useful for benchmarking domain-generalization techniques for ASD.

    Source: §4.3 · p. 4, §5 · p. 4

Abstract

from arXiv · show

We present a machine sound dataset to benchmark domain generalization techniques for anomalous sound detection (ASD). Domain shifts are differences in data distributions that can degrade the detection performance, and handling them is a major issue for the application of ASD systems. While currently available datasets for ASD tasks assume that occurrences of domain shifts are known, in practice, they can be difficult to detect. To handle such domain shifts, domain generalization techniques that perform well regardless of the domains should be investigated. In this paper, we present the first ASD dataset for the domain generalization techniques, called MIMII DG. The dataset consists of five machine types and three domain shift scenarios for each machine type. The dataset is dedicated to the domain generalization task with features such as multiple different values for parameters that cause domain shifts and introduction of domain shifts that can be difficult to detect, such as shifts in the background noise. Experimental results using two baseline systems indicate that the dataset reproduces domain shift scenarios and is useful for benchmarking domain generalization techniques.

1. INTRODUCTION

Anomalous sound detection faces domain shifts that can degrade performance, while shifts may be difficult to detect or too frequent for adaptation. MIMII DG is introduced as a dataset for benchmarking domain-generalization methods on these challenges.

1. INTRODUCTION

  • Domain shifts in normal machine sounds can degrade ASD performance and arise from machine-state changes or environmental changes such as background noise.Existing domain-adaptation datasets assume an unsupervised scenario, but domain generalization can suit shifts that are too frequent or difficult to detect for adaptation.Source: §1 · p. 1
  • Domain generalization is motivated for shifts that are too frequent to adapt to or too difficult to detect during operation.Frequent shifts can leave little adaptation data and make repeated adaptation costly; generalization aims to detect anomalies regardless of domain without operational adaptation.Source: §1 · p. 1
  • A dedicated benchmark is needed because domain generalization may require more sets recorded under different conditions and shifts that are difficult to detect.Source: §1 · p. 1
  • MIMII DG provides five machine types and three shift sections per type, each with source data for generalization and target data for evaluation.Source domains include at least two sets of values for parameters that cause shifts, and the dataset includes shifts intended to be handled with domain generalization.Source: §1 · p. 1

2. RECORDING ENVIRONMENT AND SETUP

MIMII DG records normal and anomalous sounds across five industrial machine types, then organizes them into source and target domains for three shift scenarios per type. The setup also mixes factory noise and provides training and test samples for each section.

2. RECORDING ENVIRONMENT AND SETUP

  • The dataset covers fan, gearbox, bearing, slide rail, and valve, with three factory-noise types and three shift scenarios per machine type.Shift scenarios were chosen based on experience building ASD systems for commercial applications; Table 1 describes each section’s shift type and parameter values.Source: §2 · p. 2
  • Machine recordings use 16-bit, 16-kHz, 10-second clips from one microphone channel, captured in a sound-proof room or anechoic chamber.Anomalies were recorded using deliberately damaged machines or incorrect operation; fan and valve recordings used a sound-proof room, while the other machines used an anechoic chamber.Source: §2 · p. 2
  • The five machine setups induce shifts through different operational or environmental parameters, including operating conditions, microphone location, and surrounding panels.Fan conditions stay the same across domains; gearbox varies voltage and load, bearing varies rotation speed and microphone location, slide rail varies velocity and acceleration, and valve varies operating pattern and panel location.Source: §2 · p. 2
  • Machine clips are mixed with prerecorded factory noise A, B, or C to simulate real-world environments.For selected sections, source domains use factory noise A and B while target domains use C; other sections use A and B in both domains, and Fan section 00 also includes pump sounds from MIMII DUE.Source: §2 · p. 2
  • Each machine type has three sections with source and target samples; training data include 990 source and 10 target samples per section, while test data include 50 normal and 50 anomalous samples per domain.The ten target training samples allow users to use a small amount of target data if generalization is difficult.Source: §2 · p. 2
  • Table 1 lists each section’s shift type, parameter values, and SNR, with selected parameters representing mixed-machine sounds, noise levels, or microphone locations.Source: §2 · p. 2, Table 1 · p. 3

3. RELATION TO MIMII DUE AND TOYADMOS2

MIMII DG differs from MIMII DUE and ToyADMOS2 by targeting domain generalization and providing more varied shift conditions and accessible shift metadata. Its design supports testing generalization across parameter values, including shifts that can be difficult to detect.

3. RELATION TO MIMII DUE AND TOYADMOS2

  • MIMII DG increases each domain-shift parameter to at least three values, because generalization may require multiple data sets spanning different values.For slide-rail velocity, the values increase from four in MIMII DUE to 13 in MIMII DG, allowing users to adjust task difficulty.Source: §3 · p. 3
  • The dataset introduces shifts that can be difficult to detect, including differences in the operating states of background machines.Source: §3 · p. 3
  • Shift parameters are specified in file names and attribute files for both domains, making them accessible for analysis.Target-domain parameter values let users evaluate detection performance separately for each value.Source: §3 · p. 3

4. EXPERIMENT

The experiment benchmarks domain generalization on MIMII DG using two ASD baselines and evaluates AUC across source and target domains. Target-domain AUC is lower on average, indicating that the dataset reproduces domain shifts.

4. EXPERIMENT

  • The experiment benchmarks domain-generalization performance on MIMII DG with two baseline systems.The experiment description introduces the benchmark evaluation and its two systems.Source: §4 · p. 4, §4.1 · p. 4

4.1. Baseline systems

  • The benchmark uses an autoencoder-based ASD system and a MobileNetV2-based system, the DCASE 2022 Challenge Task 2 baselines.The autoencoder uses reconstruction error, while the MobileNetV2 system uses section-ID classification as an auxiliary task.Source: §4.1 · p. 4

4.2. Metric

  • AUC is computed for each domain, section, and machine type using both source- and target-domain data, with the same threshold expected across domains.The authors motivate the shared threshold by the possibility that domain shifts are difficult to detect.Source: §4.2 · p. 4

4.3. Results

  • Target-domain AUC was lower than source-domain AUC by 13.2% with the autoencoder-based system and 12.3% with the MobileNetV2-based system.The authors report that this pattern indicates a significant difference between source- and target-domain data and that the domain-shift scenarios were reproduced.Source: §4.3 · p. 4, Table 3 · p. 4

5. CONCLUSION

MIMII DG is a dataset for benchmarking domain generalization techniques in anomalous sound detection. It contains operating sounds from five industrial-machine types with domain shifts, and two ASD-system experiments show lower detection performance on target domains.

  • MIMII DG was developed to benchmark domain generalization techniques for anomalous sound detection.Source: §5 · p. 4
  • The dataset contains normal and anomalous operating sounds from five industrial-machine types with domain shifts.Source: §5 · p. 4
  • Experiments using two ASD systems show that detection performance significantly degrades for target-domain data.Source: §5 · p. 4
Loading 2205.13879v2…