Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Impact of Frequency Bands on Acoustic Anomaly Detection of Machines using Deep Learning Based Model

Tin Nguyen, Lam Pham, Phat Lam, Dat Ngo, Hieu Tang, Alexander Schindler

TL;DR

This work tackles Acoustic Anomaly Detection of Machines (AADoM) under domain shift by emphasizing spectrogram-based feature extraction and the localization of discriminative information within narrow frequency bands. The authors introduce a baseline system (ProBaseline) using a Mobile-FaceNet–style classifier and a Gamma-distribution–based anomaly score derived from train embeddings, with distance defined as $d_e = ||\mathbf{x}-\mathbf{m}||^2_2$. They then propose a suite of improvements—pseudo audio samples, segment-level processing, online data augmentations, and explicit focus on narrow frequency bands—paired with Mahalanobis distance and Gamma-threshold tuning, achieving significant gains on the DCASE 2022 Task 2 Development set. The results demonstrate that frequency-band locality is a strong predictor of fault-related spectral content across machine types, with machine-dependent optimal bands and robust improvements over the baseline, suggesting new directions for frequency-aware network architectures in AADoM.

Abstract

In this paper, we propose a deep learning based model for Acoustic Anomaly Detection of Machines, the task for detecting abnormal machines by analysing the machine sound. By conducting extensive experiments, we indicate that multiple techniques of pseudo audios, audio segment, data augmentation, Mahalanobis distance, and narrow frequency bands, which mainly focus on feature engineering, are effective to enhance the system performance. Among the evaluating techniques, the narrow frequency bands presents a significant impact. Indeed, our proposed model, which focuses on the narrow frequency bands, outperforms the DCASE baseline on the benchmark dataset of DCASE 2022 Task 2 Development set. The important role of the narrow frequency bands indicated in this paper inspires the research community on the task of Acoustic Anomaly Detection of Machines to further investigate and propose novel network architectures focusing on the frequency bands.

The Impact of Frequency Bands on Acoustic Anomaly Detection of Machines using Deep Learning Based Model

TL;DR

This work tackles Acoustic Anomaly Detection of Machines (AADoM) under domain shift by emphasizing spectrogram-based feature extraction and the localization of discriminative information within narrow frequency bands. The authors introduce a baseline system (ProBaseline) using a Mobile-FaceNet–style classifier and a Gamma-distribution–based anomaly score derived from train embeddings, with distance defined as . They then propose a suite of improvements—pseudo audio samples, segment-level processing, online data augmentations, and explicit focus on narrow frequency bands—paired with Mahalanobis distance and Gamma-threshold tuning, achieving significant gains on the DCASE 2022 Task 2 Development set. The results demonstrate that frequency-band locality is a strong predictor of fault-related spectral content across machine types, with machine-dependent optimal bands and robust improvements over the baseline, suggesting new directions for frequency-aware network architectures in AADoM.

Abstract

In this paper, we propose a deep learning based model for Acoustic Anomaly Detection of Machines, the task for detecting abnormal machines by analysing the machine sound. By conducting extensive experiments, we indicate that multiple techniques of pseudo audios, audio segment, data augmentation, Mahalanobis distance, and narrow frequency bands, which mainly focus on feature engineering, are effective to enhance the system performance. Among the evaluating techniques, the narrow frequency bands presents a significant impact. Indeed, our proposed model, which focuses on the narrow frequency bands, outperforms the DCASE baseline on the benchmark dataset of DCASE 2022 Task 2 Development set. The important role of the narrow frequency bands indicated in this paper inspires the research community on the task of Acoustic Anomaly Detection of Machines to further investigate and propose novel network architectures focusing on the frequency bands.
Paper Structure (15 sections, 2 equations, 7 figures, 2 tables)

This paper contains 15 sections, 2 equations, 7 figures, 2 tables.

Table of Contents

  1. INTRODUCTION
  2. The Proposed Baseline System
  3. Further Improve the Proposed Baseline
  4. Generate Pseudo Audio Samples
  5. Working on Audio Segment Level
  6. Apply Online Data Augmentations
  7. Evaluate Narrow Frequency Bands
  8. Apply Mahalanobis Distance
  9. Evaluate Different Thresholds of Gamma Distribution
  10. Experiment And Results
  11. Datasets
  12. Evaluation Metric
  13. Experimental Settings
  14. Experimental Results and Discussion
  15. Conclusion

Figures (7)

  • Figure 1: The proposed baseline system for AADoM
  • Figure 2: Performance comparison (AUC) on 'source' test domain among the proposed baseline (ProBaselline) and the proposed baseline with a certain improvement (ProBaseline-Mah, ProBaseline-Mix, ProBaseline-Pse, ProBaseline-Seg, and ProBaseline-Imp)
  • Figure 3: Performance comparison (AUC) on 'target' test domain among the proposed baseline (ProBaselline) and the proposed baseline with a certain improvement (ProBaseline-Mah, ProBaseline-Mix, ProBaseline-Pse, ProBaseline-Seg, and ProBaseline-Imp)
  • Figure 4: The effect of frequency bands on 'source' test domain
  • Figure 5: The effect of frequency bands on 'target' test domain
  • ...and 2 more figures