Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization

Harsh Purohit; Kaori Suefusa; Ryo Tanabe; Takashi Endo; Yohei Kawaguchi; Yuki Nikaido

REVIEW 5 major objections 6 minor 1 cited by

Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization

T0 review · 5 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A deep autoencoding Gaussian mixture model with automatic hyperparameter choice detects faulty industrial fans from sound at 0.96 AUC, well above earlier unsupervised baselines.

desk verdict A sensible DAGMM-for-audio adaptation with two hyperparameter heuristics, but the claimed 20% improvement is unearned because conventional DAGMM is never compared and the selection rules are validated against the evaluation set. read the letter →

arxiv 2009.12042 v1 pith:ASSQZT2C submitted 2020-09-25 eess.AS cs.LGcs.SDstat.ML

Harsh Purohit , Ryo Tanabe , Takashi Endo , Kaori Suefusa , Yuki Nikaido , Yohei Kawaguchi This is my paper · ORCID

classification eess.AScs.LGcs.SDstat.ML

keywords acousticanomalydetectionunsupervisedlearningdeepautoencodingGaussianmixturemodelDAGMM-HOgapstatisticprincipalcomponentanalysislog-melspectrogramindustrialmachinesound

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a deep autoencoding Gaussian mixture model—an autoencoder that feeds a compact code plus reconstruction error into a mixture-density network—can spot faulty industrial fans from sound alone, without any labeled anomalies. Its contribution is to bring this model to acoustic signals for the first time and to make it practical by choosing its two most sensitive settings automatically from normal data: the number of mixture components via the gap statistic, and the autoencoder bottleneck size via the bending point of a PCA cumulative-variance curve. On recordings of six industrial fans, the method is reported to reach an AUC of 0.96 and an F1 of 0.94, while the best baseline included reaches 0.79 AUC. If this stands, unsupervised listening becomes a viable first-line screen for machine condition monitoring.

What carries the argument

The machinery is the joint training objective of a deep autoencoding Gaussian mixture model plus a tuning stage that fixes its two critical hyperparameters. The model compresses each log-mel spectrogram frame through an autoencoder, concatenates the latent code $z_c$ with a reconstruction-error feature $z_r$ into $z$, and a multi-layer estimation network predicts soft mixture memberships $\gamma$. The Gaussian mixture parameters $\hat{\phi}_k,\hat{\mu}_k,\hat{\Sigma}_k$ are estimated from batched memberships, and training minimizes $J=\frac{1}{N}\sum_i L(x_i,x'_i)+\lambda_1\frac{1}{N}\sum_i E(z_i)+\lambda_2 P(\Sigma)$, so compression and density estimation are optimized together. The tuning stage applies the same bending-point detector to two curves: the gap statistic over candidate component counts chooses $K$, and the cumulative PCA variance-ratio curve chooses the encoder output dimension $c$; both curves are computed from normal data alone.

What would settle it

Hold out a contiguous block of normal and anomalous fan recordings, choose hyperparameters only from the normal training block, and compare test-set AUC against the same model with hyperparameters chosen by an exhaustive grid search on that same training block; if the bending-point rules select settings that clearly underperform the grid-search optimum or fail to reproduce the reported margin over baselines, the central claim would be undermined.

Watch

Extended reading notes

Core claim

The paper's central claim is that applying DAGMM to acoustic signals, with hyperparameters chosen by two data-driven rules, yields an improvement in unsupervised anomaly detection for industrial fans. The method uses a compression network to produce a low-dimensional code and reconstruction error for each log-mel spectrogram frame, an estimation network to predict Gaussian mixture membership, and a joint objective that minimizes reconstruction error, sample energy, and a covariance singularity penalty. The gap statistic chooses the number of mixture components; the PCA cumulative-variance curve chooses the autoencoder bottleneck dimension. On six industrial fans, the resulting model reports an AUC of 0.96 and an F1 of 0.94, against 0.79 AUC for the best baseline it includes—a roughly 20 percent improvement on the AUC.

Load-bearing premise

The whole argument rests on treating the bending point of the gap-statistic curve and the bending point of the PCA cumulative-variance curve, both computed from normal data, as reliable proxies for the optimal number of mixture components and the optimal autoencoder bottleneck size in the anomaly-detection objective.

Editorial extensions

If this is right

If the reported 0.96 AUC holds, factories could screen fans and similar rotating machinery with a microphone rather than vibration or visual sensors.
The tuning stage removes the need for a per-machine grid search over mixture count and bottleneck size, since both settings come from normal data.
Because training needs only normal recordings, the method can be deployed before any fault has ever been observed on a particular machine.
The margin over reconstruction-only and density-only baselines suggests that jointly optimizing compression and density is what carries the gain, not either component alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same bending-point recipe could be tested on other deep density-estimation models whose performance depends on a latent dimension and a mixture count; if it transfers, the tuning stage becomes a general plug-in rather than a DAGMM-specific fix.
The hyperparameter rules are heuristics, so a natural stress test is to vary fan size, background noise, and recording position to see whether the gap-statistic and PCA choices remain near the true optimum.
If the model is run on streaming audio, the same energy score might flag the onset of a developing fault before it becomes audible; the paper does not address detection latency.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a reasonable workshop-level idea, but the central comparative claim does not survive contact with the experimental section. The authors take the known DAGMM model, bring it to acoustic anomaly detection for the first time, and add two automated hyperparameter choices: gap statistic for number of GMM components, PCA cumulative eigenvalues for the autoencoder bottleneck. That combination is new, and the motivation is sound—DAGMM really is sensitive to K and c, and hand-tuning them is a pain. The writing is clear and the evaluation on six real fans is a genuine attempt to test the method against several two-step baselines.

The soft spot is big. Table 1 lists DAE, GMM, PCA+GMM, DAE+GMM, OC-SVM variants, but not conventional DAGMM. Without that baseline, the reported 0.96 AUC versus 0.79 for DAE+GMM could come entirely from the DAGMM architecture, not from the gap-statistic/PCA selection rules that are supposed to be the contribution. You cannot isolate the effect of the hyperparameter optimization. The paper needs a DAGMM run with default or randomly chosen K and c, and ideally an ablation that tests the two heuristics against simple defaults on a validation split.

Also worrying is the validation of the selection rules. Figure 2 shows an accuracy surface over K and c with an 'optimal' point. If that surface is computed with the same test segments that produce the final AUC, then the validation of the heuristics is circular. The text says the heuristics use normal data, but it does not say whether the accuracy surface itself was produced on a held-out set. The authors should state this explicitly and, better, evaluate the selected hyperparameters on a disjoint test partition.

The smaller issues are consistent with this: no per-fan AUCs, no error bars, no code or data release, and Algorithm 1 is described too vaguely to reproduce (the threshold T_lmx is not precisely defined). These are fixable, but they matter for a performance claim.

Who is this for? People working on unsupervised factory-monitoring sound detection might find the combination of DAGMM with these two heuristics a useful starting point. As a DCASE workshop paper, it's a fine poster. As a serious journal claim, it needs the missing baseline, a proper validation split, and variance estimates.

My recommendation: send it to peer review only if the authors are asked to add the missing DAGMM baseline and a non-circular validation. Without those, the core claim is unproven. I would not cite it as evidence for the 20% improvement until then.

Referee Report

5 major / 6 minor

Summary. The manuscript proposes DAGMM-HO, an unsupervised anomaly detection method for acoustic signals that applies the deep autoencoding Gaussian mixture model (DAGMM) to log-mel-spectrograms of industrial fan sounds. It introduces two hyperparameter-selection rules: the gap statistic to set the number of GMM components K and a PCA cumulative-eigenvalue bending point to set the autoencoder bottleneck dimension c. On a dataset of six fans, the method is reported to achieve AUC 0.96 and F1 0.94, outperforming several two-step baselines with up to a 20% AUC improvement.

Significance. If the central claim were fully supported, the paper would make a useful practical contribution by showing that a density-based deep model with automated hyperparameter selection can reach high accuracy in unsupervised machine-sound anomaly detection. The DAGMM objective is presented clearly, and the two proposed heuristics are simple and potentially attractive for practitioners. However, as written, the evidence does not establish the claimed improvement: the conventional DAGMM that the method extends is never evaluated, the hyperparameter-selection validation appears to use the test data, and no variance or per-fan results are reported. The potential contribution is therefore not yet substantiated.

major comments (5)

[Table 1] The proposed method is compared with DAE, GMM, PCA+GMM, DAE+GMM, OC-SVM, PCA+OC-SVM, and DAE+OC-SVM, but the conventional DAGMM [14] that DAGMM-HO extends is absent. Because the only new components in the title and claims are the hyperparameter-optimization rules, the reported 0.96 AUC versus 0.79 AUC for DAE+GMM cannot be attributed to those rules; it may be entirely due to the end-to-end DAGMM architecture. Please add a DAGMM baseline with default/typical hyperparameters and, ideally, DAGMM with randomly chosen K and c, to isolate the effect of the proposed selection.
[Section 3 and Figure 2] Figure 2 shows anomaly-detection accuracy as a function of K and c and marks an optimal point. If K and c are chosen using this accuracy landscape computed on the same test set that produces the final Table 1 results, then the validation of the proposed selection rules is circular. Please clarify whether the selection in Algorithm 1 and the PCA bending point uses only normal training data, and if Figure 2 is only illustrative, state so explicitly; otherwise re-run the evaluation with hyperparameters chosen on a training/validation split and tested on held-out data.
[Sections 3.1 and 3.2] The gap statistic and PCA cumulative-eigenvalue rules are heuristic proxies, but no ablation is provided showing that they lead to better anomaly detection than simpler alternatives (e.g., a coarse grid of K and c values, or random search) on a validation split. Moreover, the selected values of K and c are not reported, making it impossible to assess whether the proposed rules yield nontrivial or stable hyperparameter choices.
[Section 4] All metrics in Table 1 are single numbers with no variance or per-fan breakdown, and the train/test split is described in one sentence. The claim of a 'significant improvement' is not supported by any statistical test or confidence interval. Please report per-fan AUC and F1 scores and appropriate variability measures.
[Algorithm 1] The bending-point detection procedure is defined operationally but its threshold T_lmx is not given a precise formula, and its sensitivity to the smoothing spline is not analyzed. Since this algorithm is load-bearing for both K and c, its behavior should be demonstrated on the actual gap-statistic and cumulative-variance curves from the fans data.

minor comments (6)

[Section 1] 'principle component analysis' should be 'principal component analysis'.
[Equation (13)] The within-cluster dispersion D(k) is written with a summation over k=1..K inside, which is inconsistent with the definition of D(k) for a single cluster; please correct the index notation.
[Section 4] The hyperparameters λ1 and λ2 in Eq. (11) and the anomaly threshold η are never specified; please report the values used.
[Figures] The figures (especially Figure 2) are reproduced at low resolution in the manuscript and the axes for Figure 2 are not described; please ensure all labels are readable and explain what is plotted.
[Reproducibility] The paper does not mention whether code or data will be released; for reproducibility, this should be stated.
[Throughout] There is a typo in the abstract header and repeated elsewhere: 'melspectrogram' and 'mel' should be 'Mel-spectrogram' or 'Mel' when referring to the filter bank.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the hyper-parameter heuristics are external to the target metric, and the reported gains are compared against external baselines.

full rationale

The paper's derivation chain is self-contained in the relevant sense. It takes the conventional DAGMM model from Zong et al. [14], applies it to log-melspectrogram features, and selects two hyper-parameters (GMM component count K and autoencoder bottleneck dimension c) using the gap statistic [15] and the bending point of PCA cumulative eigenvalues. These selection rules are computed from normal training data (Section 3 states the methods are applied 'with the use of normal data'), not from anomaly labels or from the test-set AUC that is later reported. The final comparison in Table 1 is against external baselines (DAE, GMM, PCA+GMM, DAE+GMM, OC-SVM variants), so the headline improvement is not equivalent by construction to any fitted parameter. The absence of a conventional DAGMM baseline is a real experimental omission, and Figure 2's accuracy landscape could raise concerns about test-set-aware hyper-parameter inspection, but neither amounts to circularity under the definition used here: the paper does not state that the reported K and c were chosen by optimizing Figure 2's accuracy surface, nor does it rename a fitted quantity as a prediction. The only self-citation ([8], a general reconstruction-based anomaly-detection reference) is peripheral and not load-bearing. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its load-bearing assumptions are the heuristic links from gap statistic and PCA curves to DAGMM hyperparameters, and the implicit validity of the unlabeled anomaly test data. These are ad hoc modeling choices rather than established theory.

free parameters (6)

Number of GMM components K = not explicitly reported in text; selected by gap statistic
Key hyperparameter of DAGMM; selection method is the paper's main contribution.
Reduced dimension c = not explicitly reported in text; selected by PCA cumulative eigenvalues
Key hyperparameter of the autoencoder bottleneck.
Regularization weights lambda1, lambda2 = not reported
Appear in objective function Eq. 11; affect the trade-off between reconstruction, energy, and singularity penalty.
Mel-spectrogram parameters = frame 1024, hop 512, 64 mel filters
Feature extraction settings fixed by hand without sensitivity analysis.
Network layer sizes = 60,30,10,c and c+1,10,k
Architecture chosen by hand; no ablation is provided.
Anomaly threshold eta = not reported
Final decision threshold for anomaly detection; its selection procedure is not described.

assumptions (4)

ad hoc to paper Gap statistic computed on normal data selects a GMM component count that is near-optimal for DAGMM anomaly detection.
Section 3.1 motivates gap statistic for clustering, but DAGMM uses a mixture density network, not hard clustering; the link is heuristic.
ad hoc to paper PCA cumulative eigenvalue curve gives the optimal autoencoder bottleneck dimension.
Section 3.2 states PCA 'can give a good estimate' because both minimize reconstruction error, but no proof is offered.
ad hoc to paper The bending point detection algorithm (Algorithm 1) reliably identifies the elbow of a curve.
Algorithm 1 is under-specified: threshold Tlmx and local maxima logic are not defined precisely enough to be reproduced.
domain assumption Anomalous segments in the fan dataset are valid ground truth and independent of normal segments.
Section 4 reports 'anomalous segments' without describing how they were generated or labeled.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization." pith.science (2026). https://pith.science/paper/ASSQZT2C

@misc{pith2026200912042,
  author       = {Pith},
  title        = {Pith review of: Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ASSQZT2C}},
  note         = {Machine review of arXiv:2009.12042}
}

read the original abstract

Failures or breakdowns in factory machinery can be costly to companies, so there is an increasing demand for automatic machine inspection. Existing approaches to acoustic signal-based unsupervised anomaly detection, such as those using a deep autoencoder (DA) or Gaussian mixture model (GMM), have poor anomaly-detection performance. In this work, we propose a new method based on a deep autoencoding Gaussian mixture model with hyper-parameter optimization (DAGMM-HO). In our method, the DAGMM-HO applies the conventional DAGMM to the audio domain for the first time, with the idea that its total optimization on reduction of dimensions and statistical modelling will improve the anomaly-detection performance. In addition, the DAGMM-HO solves the hyper-parameter sensitivity problem of the conventional DAGMM by performing hyper-parameter optimization based on the gap statistic and the cumulative eigenvalues. Our evaluation of the proposed method with experimental data of the industrial fans showed that it significantly outperforms previous approaches and achieves up to a 20% improvement based on the standard AUC score.

Figures

Figures reproduced from arXiv: 2009.12042 by the authors.

**Figure 1.** Overview of DAGMM-HO-based acoustic anomaly de [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** Illustration of curve bending point detection. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: 2-D representation of audio signals of six industrial fans. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Contrastive Self-Supervised Network Intrusion Detection using Augmented Negative Pairs
cs.LG 2025-09 conditional novelty 5.0 of 10

CLAN clusters genuine benign network flows while repelling augmented copies, then classifies new flows by distance to the cluster centroid; on Lycos2017 it reports the highest mean AUROC among compared SSL and anomaly...

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages · cited by 1 Pith paper

[14]

How can we detect anomalies from subsampled audio signals?

Y . Kawaguchi and T. Endo, “How can we detect anomalies from subsampled audio signals?” in Proceedings of the IEEE 27th International Workshop on Machine Learning for Signal Processing (MLSP), 2017, pp. 1–6

work page 2017
[1]

To avoid this issue, many anomaly detection techniques are applied for the smart factory maintenance

INTRODUCTION Anomaly events can decrease the quality of manufactured products and deteriorate the reliability of the industrial processes. To avoid this issue, many anomaly detection techniques are applied for the smart factory maintenance. These techniques are mainly based on sensor data parameters, environment variables, quality metrics of the industria...

work page
[2]

Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization

DAGMM-HO-BASED ACOUSTIC ANOMALY DETECTION MODEL The proposed acoustic anomaly detection model with the deep autoencoding Gaussian mixture model with hyper-parameter opti- mization (DAGMM-HO) is composed of two parts, as shown in Fig1. In the ﬁrst part, hyper-parameter tuning has been performed to determine the correct number of GMM components and the opti...

work page Pith review arXiv 2009
[3]

Fig 2 depicts the change in accuracy of anomaly detection with respect to the number of GMM components and re- duced dimension in the DA

HYPER-PARAMETER OPTIMIZA TION In order to develop an acoustic anomaly detection method based on DAGMM, we found that two important hyper-parameters should be tuned accurately. Fig 2 depicts the change in accuracy of anomaly detection with respect to the number of GMM components and re- duced dimension in the DA. The black point represents the opti- mal va...

work page 2020
[4]

We are interested in the point of maximum change (represented as the red dot) on the curve because it indicates the best cluster count

(13) The curve in Fig 3 represents gap statistic valuesGk for differ- ent cluster counts k. We are interested in the point of maximum change (represented as the red dot) on the curve because it indicates the best cluster count. For example, in Fig 4, six different clusters are present but the best minimum number of components to repre- sent the overall da...

work page 2020
[5]

We collected sound data (∼ 10 min each) of six industrial fans with different sizes and manufacturers

EXPERIMENTA TION To evaluate the effectiveness of the proposed approach, we applied it using real experimental data. We collected sound data (∼ 10 min each) of six industrial fans with different sizes and manufacturers. A microphone was placed at 50 cm away from each fan to collect the audio signal. Saturation of the recorded signal was avoided during dat...

work page
[6]

We proposed an automated procedure to determine the optimal number of GMM components and reduce the dimension in the DA

CONCLUSIONS We developed an acoustic anomaly detection method using a deep autoencoding Gaussian mixture model with hyper-parameter opti- mization. We proposed an automated procedure to determine the optimal number of GMM components and reduce the dimension in the DA. Signiﬁcant improvement in accuracy over conventional methods was achieved during an expe...

work page
[7]

Predictive maintenance management using sensor-based degradation models,

K. A. Kaiser and N. Z. Gebraeel, “Predictive maintenance management using sensor-based degradation models,” IEEE Transactions on Systems, Man, and Cybernetics-Part A: Sys- tems and Humans, vol. 39, no. 4, pp. 840–849, 2009. Detection and Classiﬁcation of Acoustic Scenes and Events 2020 2–3 November 2020, Tokyo, Japan

work page 2009

Show all 22 references

[8]

Enhanced industrial machinery condition monitoring methodology based on nov- elty detection and multi-modal analysis,

J. A. Carino, M. Delgado-Prieto, D. Zurita, M. Millan, J. A. O. Redondo, and R. Romero-Troncoso, “Enhanced industrial machinery condition monitoring methodology based on nov- elty detection and multi-modal analysis,”IEEE Access, vol. 4, pp. 7594–7604, 2016

2016
[9]

Thermal image based fault diagnosis for rotating machinery,

O. Janssens, R. Schulz, V . Slavkovikj, K. Stockman, M. Loc- cuﬁer, R. Van de Walle, and S. Van Hoecke, “Thermal image based fault diagnosis for rotating machinery,”Infrared Physics & Technology, vol. 73, pp. 78–87, 2015

2015
[10]

Rare sound event detection us- ing 1d convolutional recurrent neural networks,

H. Lim, J. Park, and Y . Han, “Rare sound event detection us- ing 1d convolutional recurrent neural networks,” in Proceed- ings of the Detection and Classiﬁcation of Acoustic Scenes and Events (DCASE) Workshop, 2017, pp. 80–84

2017
[11]

Unsupervised detection of anomalous sound based on deep learning and the Neyman–Pearson lemma,

Y . Koizumi, S. Saito, H. Uematsu, Y . Kawachi, and N. Harada, “Unsupervised detection of anomalous sound based on deep learning and the Neyman–Pearson lemma,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 27, no. 1, pp. 212–224, 2018

2018
[12]

Structured denoising autoencoder for fault detection and analysis,

T. Tagawa, Y . Tadokoro, and T. Yairi, “Structured denoising autoencoder for fault detection and analysis,” in Proceedings of the Asian Conference on Machine Learning (ACML), 2015, pp. 96–111

2015
[13]

A novel approach for automatic acoustic novelty detection using a denoising autoencoder with bidirectional LSTM neural networks,

E. Marchi, F. Vesperini, F. Eyben, S. Squartini, and B. Schuller, “A novel approach for automatic acoustic novelty detection using a denoising autoencoder with bidirectional LSTM neural networks,” in Proceedings of the IEEE Interna- tional Conference on Acoustics, Speech and S...

2015
[15]

Residual error based anomaly detection using auto-encoder in smd machine sound,

D. Oh and I. Yun, “Residual error based anomaly detection using auto-encoder in smd machine sound,” Sensors, vol. 18, no. 5, p. 1308, 2018

2018
[16]

Deep boltzmann machines,

R. Salakhutdinov and G. Hinton, “Deep boltzmann machines,” in Artiﬁcial intelligence and statistics, 2009, pp. 448–455

2009
[17]

C. M. Bishop, Pattern recognition and machine learning . Springer, 2006

2006
[18]

A comparative evaluation of unsupervised anomaly detection algorithms for multivariate data,

M. Goldstein and S. Uchida, “A comparative evaluation of unsupervised anomaly detection algorithms for multivariate data,”PloS one, vol. 11, no. 4, p. 173, 2016

2016
[19]

One-class svm based approach for detecting anomalous audio events,

F. Aurino, M. Folla, F. Gargiulo, V . Moscato, A. Picariello, and C. Sansone, “One-class svm based approach for detecting anomalous audio events,” in 2014 International Conference on Intelligent Networking and Collaborative Systems. IEEE, 2014, pp. 145–151

2014
[20]

Deep Autoencoding Gaussian Mix- ture Model for Unsupervised Anomaly Detection,

B. Zong, Q. Song, M. R. Min, W. Cheng, C. Lumezanu, D. ki Cho, and H. Chen, “Deep Autoencoding Gaussian Mix- ture Model for Unsupervised Anomaly Detection,” inInterna- tional Conference on Learning Representations (ICLR), 2018

2018
[21]

Estimating the num- ber of clusters in a data set via the gap statistic,

R. Tibshirani, G. Walther, and T. Hastie, “Estimating the num- ber of clusters in a data set via the gap statistic,”Journal of the Royal Statistical Society: Series B (Statistical Methodology) , vol. 63, no. 2, pp. 411–423, 2001

2001
[22]

Principal component analy- sis,

H. Abdi and L. J. Williams, “Principal component analy- sis,”Wiley interdisciplinary reviews: computational statistics, vol. 2, no. 4, pp. 433–459, 2010

2010

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [14]

How can we detect anomalies from subsampled audio signals?

Y . Kawaguchi and T. Endo, “How can we detect anomalies from subsampled audio signals?” in Proceedings of the IEEE 27th International Workshop on Machine Learning for Signal Processing (MLSP), 2017, pp. 1–6

work page 2017

[2] [1]

To avoid this issue, many anomaly detection techniques are applied for the smart factory maintenance

INTRODUCTION Anomaly events can decrease the quality of manufactured products and deteriorate the reliability of the industrial processes. To avoid this issue, many anomaly detection techniques are applied for the smart factory maintenance. These techniques are mainly based on sensor data parameters, environment variables, quality metrics of the industria...

work page

[3] [2]

Deep Autoencoding GMM-based Unsupervised Anomaly Detection in Acoustic Signals and its Hyper-parameter Optimization

DAGMM-HO-BASED ACOUSTIC ANOMALY DETECTION MODEL The proposed acoustic anomaly detection model with the deep autoencoding Gaussian mixture model with hyper-parameter opti- mization (DAGMM-HO) is composed of two parts, as shown in Fig1. In the ﬁrst part, hyper-parameter tuning has been performed to determine the correct number of GMM components and the opti...

work page Pith review arXiv 2009

[4] [3]

Fig 2 depicts the change in accuracy of anomaly detection with respect to the number of GMM components and re- duced dimension in the DA

HYPER-PARAMETER OPTIMIZA TION In order to develop an acoustic anomaly detection method based on DAGMM, we found that two important hyper-parameters should be tuned accurately. Fig 2 depicts the change in accuracy of anomaly detection with respect to the number of GMM components and re- duced dimension in the DA. The black point represents the opti- mal va...

work page 2020

[5] [4]

We are interested in the point of maximum change (represented as the red dot) on the curve because it indicates the best cluster count

(13) The curve in Fig 3 represents gap statistic valuesGk for differ- ent cluster counts k. We are interested in the point of maximum change (represented as the red dot) on the curve because it indicates the best cluster count. For example, in Fig 4, six different clusters are present but the best minimum number of components to repre- sent the overall da...

work page 2020

[6] [5]

We collected sound data (∼ 10 min each) of six industrial fans with different sizes and manufacturers

EXPERIMENTA TION To evaluate the effectiveness of the proposed approach, we applied it using real experimental data. We collected sound data (∼ 10 min each) of six industrial fans with different sizes and manufacturers. A microphone was placed at 50 cm away from each fan to collect the audio signal. Saturation of the recorded signal was avoided during dat...

work page

[7] [6]

We proposed an automated procedure to determine the optimal number of GMM components and reduce the dimension in the DA

CONCLUSIONS We developed an acoustic anomaly detection method using a deep autoencoding Gaussian mixture model with hyper-parameter opti- mization. We proposed an automated procedure to determine the optimal number of GMM components and reduce the dimension in the DA. Signiﬁcant improvement in accuracy over conventional methods was achieved during an expe...

work page

[8] [7]

Predictive maintenance management using sensor-based degradation models,

K. A. Kaiser and N. Z. Gebraeel, “Predictive maintenance management using sensor-based degradation models,” IEEE Transactions on Systems, Man, and Cybernetics-Part A: Sys- tems and Humans, vol. 39, no. 4, pp. 840–849, 2009. Detection and Classiﬁcation of Acoustic Scenes and Events 2020 2–3 November 2020, Tokyo, Japan

work page 2009