PaperlayerView exact revision on arXiv

Source-linked AI summary

Few-Shot Adaptation for Anomalous Sound Detection at Low False-Positive Rate

Simone Lugani, Christof Pichler, Hannes Wegleiter, Christoph Feilmayr, Rodolfo Zunino, Edoardo Ragusa

arXiv:2610.19893v1cs.SDcs.LGeess.AS

TL;DR

Reliable industrial acoustic fault detection is constrained by false alarms, while complex sound conditions make detection difficult. The paper combines a regularized unsupervised detector trained on normal audio with few-shot adaptation after a fault; adaptation improves detection at very low false-positive rates.

  • Problem

    Industrial ASD must operate at very low false-positive rates, but complex acoustic conditions and remaining false alarms constrain reliable deployment.

    Source: §I · p. 1

  • Method

    The system trains a CVAE–GMM model on normal audio, then fine-tunes the CVAE with a few anomalous samples while keeping the GMM fixed as the normal reference.

    Source: §I · p. 1

  • Results

    At FPR = 0.1%, the adapted system detects more than 1 out of 2 anomalous samples, while UNS R detects on average fewer than 1 out of 10.

    Source: D. Performance at Low… · p. 4

  • Takeaways & Limitations

    Covariance regularization improves the unsupervised baseline’s average performance and stability, while few-shot adaptation substantially improves detection in the low-FPR region.

    Source: C. Comparison with State-of-the-Art · p. 3, §IV · p. 4

    Evaluation will be extended to additional fault types, longer deployments, and hyperparameter sensitivity analysis.

    Source: §IV · p. 4

Abstract

from arXiv · show

Reliable fault detection is crucial for industrial machinery in large-scale facilities, where unexpected failures can cause costly downtime and production interruptions. Although acoustic monitoring provides a scalable solution for machine condition monitoring, its deployment is often constrained by high false-positive rates. We propose a two-step approach to automatic anomalous sound detection: first, when only normal operating audio samples are available, a regularized unsupervised model is deployed; then, after the first machine fault, a few-shot adaptation strategy exploits the anomalous samples acquired during the fault to improve the system. This adaptation increases the area under the precision-recall curve (AUPR) from 23.0 to 64.0 and improves detection in the low false-positive-rate region, with the partial area under the receiver operating characteristic curve (pAUROC) increasing from 39.6 to 72.8. We also show that covariance regularization improves both prediction accuracy and training stability in the unsupervised model. These results provide a practical strategy for acoustic anomaly detection systems, enabling improved accuracy and reduced false alarms.

I. INTRODUCTION

The paper targets reliable acoustic fault detection in complex industrial settings, where false alarms constrain deployment. It proposes unsupervised detection followed by adaptation using a few fault samples.

  • Acoustic monitoring uses microphones, but concurrent screens, changing background noise, and other disturbances can mask weak fault signatures.Vibration monitoring is well established, but sensor effort and installation cost can limit its use in large-scale facilities.Source: §I · p. 1
  • Unsupervised ASD learns normal machine behavior and identifies deviations, avoiding the need for large labeled fault datasets.Recent DCASE tasks have also explored first-shot and domain-generalization settings [11]–.Source: §I · p. 1
  • The framework compresses log-mel spectrograms with a CVAE before a GMM computes anomaly scores against the learned normal representation.Source: Figure 1 · p. 1
  • In large facilities, even a small number of false alarms can limit system usability and acceptance, while complex sound environments make reliable detection difficult.Source: §I · p. 1
  • After a first observed fault, a few anomalous samples can be used to refine the unsupervised model, targeting better normal–fault separation at low FPR.The framework also investigates covariance regularization for robustness in the unsupervised setting and is evaluated on real-world blast-furnace stock-house audio.Source: §I · p. 1
  • The proposed framework outperforms baselines in overall AUPR and detection under stringent false-alarm requirements.Source: §I · p. 2

II. METHODS

The system begins with unsupervised learning from normal operating data and later incorporates fault data through few-shot adaptation. Its CVAE–GMM framework scores deviations from learned normal audio representations.

II. METHODS

  • The two-stage paradigm trains and deploys an unsupervised detector on normal data, then adapts it using sound recorded after the first machine fault.Source: §II · p. 2

A. Anomalous Sound Detection Framework

1) Preprocessing:

The method converts audio segments into normalized log-mel spectrograms, encodes them as CVAE latent statistics, and uses a GMM likelihood to produce anomaly scores.

1) Preprocessing:

  • Each audio segment is represented by a log-mel spectrogram normalized to the range [0,1].Source: 1) Preprocessing · p. 2

2) Convolutional Variational Autoencoder:

3) Gaussian Mixture Model Anomaly Scoring:

B. Unsupervised Training

Unsupervised training uses only normal recordings: a CVAE learns a compact reconstruction with a variational objective, then a regularized GMM models its latent representations.

B. Unsupervised Training

  • The CVAE is trained on normal log-mel spectrograms with reconstruction and KL-regularization terms, the latter promoting a smoother, more stable latent space.Source: B. Unsupervised Training · p. 2
  • The loss uses the input and reconstructed spectrograms, latent means and variances, and β as the KL-term weight.Source: B. Unsupervised Training · p. 2
  • A GMM is fitted to latent representations from dedicated normal recordings, using covariance regularization for a more robust distribution estimate.Its encoder–GMM combination forms the unsupervised stage and starting point for subsequent learning.Source: B. Unsupervised Training · p. 2

C. Few-shot Adaptation with Single Machine Fault

After the first fault, the CVAE is fine-tuned on a small set of anomalous samples to improve separation of normal and faulty sounds without specializing to one machine. The adaptation keeps the unsupervised GMM fixed and combines latent-space separation with normal-data reconstruction regularization.

C. Few-shot Adaptation with Single Machine Fault

  • Few-shot adaptation fine-tunes the unsupervised CVAE on a small set of samples from one fault event to improve normal–fault separation across the facility.The stated objective is not to specialize the model to a specific machine.Source: C. Few-shot Adaptation with Single… · p. 2
  • The unsupervised GMM remains fixed as the normal latent-distribution reference while the CVAE is optimized on normal and augmented anomalous samples.Augmentations use minor temporal shifts, temporal cropping, amplitude scaling, and spectral coloration.Source: C. Few-shot Adaptation with Single… · p. 2
  • The adaptation loss separates normal and anomalous latent representations while retaining normal-data reconstruction behavior as a regularizer.The normal-consistency and unsupervised VAE terms use normal data, while the anomalous-separation term uses anomalous data; λ controls the unsupervised term.Source: C. Few-shot Adaptation with Single… · p. 2
  • The normal and anomalous latent-space terms use margins selected manually from pre-adaptation normal anomaly scores, without anomalous validation data.No independent fault-validation set was available, and using test faults for margin selection would introduce data leakage.Source: C. Few-shot Adaptation with Single… · p. 2

A. Data Acquisition and Datasets

The study collected a custom real-world acoustic dataset at an operating blast furnace stock house under industrial conditions. The setup includes a vibrating screen line and recordings of normal operation amid concurrent sound sources and disturbances.

A. Data Acquisition and Datasets

  • The measurement system was deployed in an operating blast furnace stock house to collect a custom real-world dataset.Recordings were made with a microphone approximately 5 m from the vibrating screen line, in an environment around 95 dBSPL with 13 screens and additional disturbances.Source: A. Data Acquisition and Datasets · pp. 2–3
  • Figure 2 depicts the facility layout, microphone and fault locations, vibrating screening machine, and physical reconstruction of typical mechanical failures.The recreated fault case includes a physical reconstruction of a mechanical failure scenario.Source: Figure 2 · p. 3
  • The normal dataset provides 60,000 s for CVAE training, 60,000 s for GMM fitting, and 71,200 s of independent normal test data.Recordings were acquired over several months at 48 kHz and intentionally used without preprocessing.Source: A. Data Acquisition and Datasets · p. 3

B. Experimental setup

The experimental setup evaluates unsupervised and adapted ASD using normal data, three fault folds, and repeated training runs. Figure 3 compares AUPR across the unsupervised baseline, its regularized version, and the few-shot model.

B. Experimental setup

  • The CVAE uses 64 × 128 log-mel inputs and a 16-dimensional latent space, while the GMM uses 50 full-covariance components with covariance regularization 10^-5.The GMM was fitted using 10 random initializations.Source: B. Experimental setup · p. 3
  • Figure 3 reports a drastic AUPR improvement after few-shot adaptation compared with the unsupervised methods.The figure compares UNS, UNS R, and FS.Source: Figure 3 · p. 3
  • Each model was trained 10 times with different random seeds, and metrics are reported as mean and standard deviation.Source: B. Experimental setup · p. 3
  • Fault recordings from three machines were split into folds, with one machine's fold used for few-shot training and the other two for testing in each configuration.The recordings totaled 120 s, 220 s, and 60 s across the three machines; global metrics average the three fold configurations.Source: B. Experimental setup · p. 3

C. Comparison with State-of-the-Art

Few-shot adaptation raises AUPR after covariance regularization improves the unsupervised baseline. The adapted model also improves detection in the low-FPR region, with fold results consistent across machines but sensitive in absolute value to the train–test split.

C. Comparison with State-of-the-Art

D. Performance at Low False-Positive-Rate

  • Table I reports low-FPR model performance; pAUROC covers FPR ≤1%, while TPR@FPR 1% and 0.1% use thresholds fixed at those rates.Source: D. Performance at Low… · p. 3, Table I · p. 4
  • At FPR = 0.1%, the adapted system detects more than 1 out of 2 anomalous samples, while the regularized baseline averages fewer than 1 out of 10.The regularized baseline remains unreliable at FPR = 0.1%.Source: D. Performance at Low… · p. 4
  • Across folds, adaptation using anomalies from one machine improves detection on the other two, though absolute values vary with the split when samples are very few.Table II reports pAUROC by fold.Source: D. Performance at Low… · p. 4, Table II · p. 4

IV. CONCLUSION

Covariance regularization improves the unsupervised baseline, while few-shot adaptation substantially boosts detection in the low-FPR region. The evaluation is positioned as promising for industrial ASD, with broader fault types and longer deployments left for future work.

  • Few-shot adaptation substantially boosts detection, especially in the low-FPR region relevant to industrial deployment.It yields higher pAUROC and TPR at very low FPR values.Source: §IV · p. 4
  • Covariance regularization improves the unsupervised GMM baseline in both average performance and stability.Source: §IV · p. 4
  • Future work will extend evaluation to additional fault types, longer deployments, and hyperparameter sensitivity.Source: §IV · p. 4
Loading 2610.19893v1…