Files
website/_site/research/sound-anomaly-transfer/index.html

779 lines
21 KiB
HTML
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
<!doctype html>
<!--
Minimal Mistakes Jekyll Theme 4.27.1 by Michael Rose
Copyright 2013-2025 Michael Rose - mademistakes.com | @mmistakes
Free for personal and commercial use under the MIT license
https://github.com/mmistakes/minimal-mistakes/blob/master/LICENSE
---
Modified by Steffen Illium 2023 - 2024
https://steffenillium.de/about
-->
<html lang="en-US" class="no-js">
<head>
<meta charset="utf-8">
<!-- begin _includes/seo.html --><title>Sound Anomaly Transfer — Steffen Illium</title>
<meta name="description" content="Image nets detect acoustic anomalies in machinery via spectrograms.">
<meta name="author" content="Steffen Illium">
<meta property="article:author" content="Steffen Illium">
<meta property="og:type" content="article">
<meta property="og:locale" content="en_US">
<meta property="og:site_name" content="Steffen Illium">
<meta property="og:title" content="Sound Anomaly Transfer">
<meta property="og:url" content="http://localhost:4000/research/sound-anomaly-transfer/">
<meta property="og:description" content="Image nets detect acoustic anomalies in machinery via spectrograms.">
<meta property="og:image" content="http://localhost:4000/assets/figures/9_image_transfer_sound_teaser.jpg">
<meta property="article:published_time" content="2021-03-02T00:00:00+01:00">
<link rel="canonical" href="http://localhost:4000/research/sound-anomaly-transfer/">
<meta name="google-site-verification" content="ZNT7_sXtFdgqBLCPLavTHWcviZRk__BNvClY8I-sFjU" />
<!-- end _includes/seo.html -->
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<script>
document.documentElement.className = document.documentElement.className.replace(/\bno-js\b/g, '') + ' js ';
</script>
<!-- For all browsers -->
<link rel="stylesheet" href="/assets/css/main.css">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/@fortawesome/fontawesome-free@latest/css/all.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<noscript><link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/@fortawesome/fontawesome-free@latest/css/all.min.css"></noscript>
<!-- Academicons -->
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<!-- start custom head snippets -->
<!-- insert favicons. use https://realfavicongenerator.net/ -->
<!-- end custom head snippets -->
</head>
<body class="layout--single" dir="ltr">
<nav class="skip-links">
<ul>
<li><a href="#site-nav" class="screen-reader-shortcut">Skip to primary navigation</a></li>
<li><a href="#main" class="screen-reader-shortcut">Skip to content</a></li>
<li><a href="#footer" class="screen-reader-shortcut">Skip to footer</a></li>
</ul>
</nav>
<div class="masthead">
<div class="masthead__inner-wrap">
<div class="masthead__menu">
<nav id="site-nav" class="greedy-nav">
<a class="site-title" href="/">
portfolio
<span class="site-subtitle"></span>
</a>
<ul class="visible-links"><li class="masthead__menu-item">
<a
href="/publications"
>publications</a>
</li><li class="masthead__menu-item">
<a
href="/research"
>research</a>
</li><li class="masthead__menu-item">
<a
href="/projects"
>projects</a>
</li><li class="masthead__menu-item">
<a
href="/teaching"
>teaching</a>
</li><li class="masthead__menu-item">
<a
href="/about"
>about me</a>
</li></ul>
<button class="greedy-nav__toggle hidden" type="button">
<span class="visually-hidden">Toggle menu</span>
<div class="navicon"></div>
</button>
<ul class="hidden-links hidden"></ul>
</nav>
</div>
</div>
</div>
<div class="initial-content">
<div id="main" role="main">
<div class="sidebar sticky">
<div itemscope itemtype="https://schema.org/Person" class="h-card">
<div class="author__avatar">
<a href="http://localhost:4000/">
<img src="/assets/images/newshot_2.jpg" alt="Steffen Illium" itemprop="image" class="u-photo">
</a>
</div>
<div class="author__content">
<h3 class="author__name p-name" itemprop="name">
<a class="u-url" rel="me" href="http://localhost:4000/" itemprop="url">Steffen Illium</a>
</h3>
<div class="author__bio p-note" itemprop="description">
<p><a href="https://www.mobile.ifi.lmu.de/team/steffen-illium/">PhD. in Comp. Science</a> <br /><a href="/research/">AI Consultant &amp; Researcher</a></p>
</div>
</div>
<div class="author__urls-wrapper">
<button class="btn btn--inverse">Follow</button>
<ul class="author__urls social-icons">
<li itemprop="homeLocation" itemscope itemtype="https://schema.org/Place">
<i class="fas fa-fw fa-map-marker-alt" aria-hidden="true"></i> <span itemprop="name" class="p-locality">Augsburg</span>
</li>
<li><a href="https://www.mobile.ifi.lmu.de/team/steffen-illium/" rel="nofollow noopener noreferrer me" itemprop="sameAs"><i class="fas fa-link" aria-hidden="true"></i><span class="label">LMU-Munich</span></a></li>
<li><a href="https://scholar.google.de/citations?hl=en&pli=1&user=NODAd94AAAAJ" rel="nofollow noopener noreferrer me" itemprop="sameAs"><i class="fab fa-google-scholar" aria-hidden="true"></i><span class="label">Scholar</span></a></li>
<li><a href="https://arxiv.org/a/illium_s_1.html" rel="nofollow noopener noreferrer me" itemprop="sameAs"><i class="ai ai-arxiv" aria-hidden="true"></i><span class="label">Arxive</span></a></li>
<li><a href="https://www.researchgate.net/profile/Steffen-Illium" rel="nofollow noopener noreferrer me" itemprop="sameAs"><i class="fab fa-researchgate" aria-hidden="true"></i><span class="label">Researchgate</span></a></li>
<!--
<li>
<a href="http://link-to-whatever-social-network.com/user/" itemprop="sameAs" rel="nofollow noopener noreferrer me">
<i class="fas fa-fw" aria-hidden="true"></i> Custom Social Profile Link
</a>
</li>
-->
</ul>
</div>
</div>
</div>
<article class="page" itemscope itemtype="https://schema.org/CreativeWork">
<meta itemprop="headline" content="Sound Anomaly Transfer">
<meta itemprop="description" content="Image nets detect acoustic anomalies in machinery via spectrograms.">
<meta itemprop="datePublished" content="2021-03-02T00:00:00+01:00">
<div class="page__inner-wrap">
<header>
<h1 id="page-title" class="page__title" itemprop="headline">
<a href="http://localhost:4000/research/sound-anomaly-transfer/" itemprop="url">Sound Anomaly Transfer
</a>
</h1>
<p class="page__meta">
<span class="page__meta-readtime">
<i class="far fa-clock" aria-hidden="true"></i>
1 minute read
</span>
</p>
</header>
<br>
<section class="page__content" itemprop="text">
<hr>
<header><h4 class="nav__title"><i class="fas fa-book"></i> Reference</h4></header>
<ol class="bibliography"><li><span id="muller2020acoustic"><span style="font-variant: small-caps">Müller, R., Ritz, F., Illium, S., and Linnhoff-Popien, C.</span> 2020. Acoustic anomaly detection for machine sounds based on image transfer learning. <i>arXiv preprint arXiv:2006.03429</i>.</span>
<div class="publication-actions" style="margin-top: 0.5em;">
<a href="/assets/publications/muller2020acoustic.pdf" style="text-decoration: none;">
<button class="btn btnId btnPub--action btnPub--pdf-download"
data-umami-event="PDF Download"
data-umami-event-bibkey="muller2020acoustic"
data-umami-event-slug="sound-anomaly-transfer"
style="outline:none; position:relative; white-space: normal; margin-right: 5px; vertical-align: middle;">
<i class="fas fa-download"></i> PDF
</button>
</a>
<button type="button" class="btn btnId btnPub--action btnPub--bibtex-download-direct"
data-umami-event="BibTeX File Download"
data-umami-event-bibkey="muller2020acoustic"
data-umami-event-slug="sound-anomaly-transfer"
data-bibtex-json='"@article{muller2020acoustic,\n title = {Acoustic anomaly detection for machine sounds based on image transfer learning},\n author = {Müller, Robert and Ritz, Fabian and Illium, Steffen and Linnhoff-Popien, Claudia},\n journal = {arXiv preprint arXiv:2006.03429},\n year = {2020}\n}\n"'
data-bibtex-filename='muller2020acoustic.bib'
style="outline:none; position:relative; white-space: normal; margin-right: 5px; vertical-align: middle;">
<i class="fas fa-download"></i> BibTeX
</button>
<button type="button" class="btn btnId btnPub--action btnPub--bibtex-copy-direct"
data-umami-event="BibTeX Copy"
data-umami-event-bibkey="muller2020acoustic"
data-umami-event-slug="sound-anomaly-transfer"
data-bibtex-json='"@article{muller2020acoustic,\n title = {Acoustic anomaly detection for machine sounds based on image transfer learning},\n author = {Müller, Robert and Ritz, Fabian and Illium, Steffen and Linnhoff-Popien, Claudia},\n journal = {arXiv preprint arXiv:2006.03429},\n year = {2020}\n}\n"'
style="outline:none; position:relative; white-space: normal; vertical-align: middle;">
<i class="fas fa-clipboard"></i> Copy
</button>
</div></li></ol>
<hr>
<p style="display:block; width:45%" class="align-right"><img src="\assets\figures\9_image_transfer_sound_workflow.jpg" alt="Workflow diagram showing mel-spectrogram input, feature extraction via image network, and anomaly detection model" /></p>
<p>This study investigates an effective approach for <strong>acoustic anomaly detection</strong> in industrial machinery, focusing on identifying malfunctions through sound analysis. The core methodology leverages <strong>transfer learning</strong> by repurposing deep neural networks originally trained for large-scale <strong>image classification</strong> (e.g., on ImageNet) as powerful feature extractors for audio data represented as <strong>mel-spectrograms</strong>.</p>
<p>The process involves:</p>
<ol>
<li>Converting audio signals from machinery into mel-spectrogram images.</li>
<li>Feeding these spectrograms into various pretrained image classification networks (specifically comparing <strong>ResNet architectures</strong> against <strong>AlexNet</strong> and <strong>SqueezeNet</strong>) to extract deep feature representations.</li>
<li>Training standard anomaly detection models particularly <strong>Gaussian Mixture Models (GMMs)</strong> and <strong>One-Class Support Vector Machines (OC-SVMs)</strong> on the features extracted from normal operation sounds.</li>
<li>Classifying new sounds as anomalous if their extracted features deviate significantly from the learned normality model.</li>
</ol>
<p>Key findings from the experiments, conducted across different machine types and noise conditions, include:</p>
<ul>
<li>The proposed transfer learning approach significantly <strong>outperforms baseline methods like traditional convolutional autoencoders</strong>, especially in the presence of background noise.</li>
<li>Features extracted using <strong>ResNet architectures consistently yielded superior anomaly detection performance</strong> compared to those from AlexNet and SqueezeNet.</li>
<li><strong>GMMs and OC-SVMs proved highly effective</strong> as anomaly detection classifiers when applied to these transferred features.</li>
</ul>
<div style="clear: both;"></div>
<center>
<img src="/assets/figures/9_image_transfer_sound_mels.jpg" alt="Examples of mel-spectrograms from normal and anomalous machine sounds" style="display:block; width:85%" />
<figcaption>Mel-spectrogram examples of normal vs. anomalous machine sounds.</figcaption>
</center>
<p>This work demonstrates the surprising effectiveness of transferring knowledge from the visual domain to the acoustic domain for anomaly detection, offering a robust and readily implementable method for monitoring industrial equipment. <a class="citation" href="/publications#muller2020acoustic">[Müller et al. 2020]</a></p>
</section>
<footer class="page__meta">
<p class="page__taxonomy">
<strong><i class="fas fa-fw fa-tags" aria-hidden="true"></i> Tags: </strong>
<span itemprop="keywords">
<a href="/tags/anomaly-detection" class="page__taxonomy-item p-category" rel="tag">anomaly-detection</a><span class="sep">, </span>
<a href="/tags/audio-classification" class="page__taxonomy-item p-category" rel="tag">audio-classification</a><span class="sep">, </span>
<a href="/tags/computer-vision" class="page__taxonomy-item p-category" rel="tag">computer-vision</a><span class="sep">, </span>
<a href="/tags/deep-learning" class="page__taxonomy-item p-category" rel="tag">deep-learning</a><span class="sep">, </span>
<a href="/tags/feature-extraction" class="page__taxonomy-item p-category" rel="tag">feature-extraction</a><span class="sep">, </span>
<a href="/tags/transfer-learning" class="page__taxonomy-item p-category" rel="tag">transfer-learning</a>
</span>
</p>
<p class="page__date"><strong><i class="fas fa-fw fa-calendar-alt" aria-hidden="true"></i> Updated:</strong> <time class="dt-published" datetime="2021-03-02T00:00:00+01:00">March 2, 2021</time></p>
</footer>
<!--
<nav class="pagination">
<a href="/research/anomalous-sound-features/" class="pagination--pager" title="Anomalous Sound Features">Previous</a>
<a href="/research/acoustic-leak-detection/" class="pagination--pager" title="Acoustic Leak Detection">Next</a>
</nav>
-->
</div>
</article>
<div class="page__related">
<h2 class="page__related-title">Related posts</h2>
<div class="grid__wrapper">
<div class="list__item">
<article class="archive__item" itemscope itemtype="https://schema.org/CreativeWork"
style="display: flex; align-items: flex-start;"
>
<div class="archive__item-teaser-list" style="margin-right: 1em; flex-shrink: 0;">
<img src="/assets/figures/21_coins_teaser.png" alt="" style="display: block;">
</div>
<div style="flex-grow: 1;" >
<h2 class="archive__item-title no_toc" itemprop="headline" style="text-align: left !important;">
<a href="/research/mas-emergence-safety/" rel="permalink">MAS Emergence Safety
</a>
</h2>
<p class="page__meta">
<span class="page__meta-readtime">
<i class="far fa-clock" aria-hidden="true"></i>
1 minute read
</span>
</p>
<p class="archive__item-excerpt" itemprop="description" style="text-align: left;">
Formalized MAS emergence misalignment; proposed safety mitigation strategies.
</p>
</div>
</article>
</div>
<div class="list__item">
<article class="archive__item" itemscope itemtype="https://schema.org/CreativeWork"
style="display: flex; align-items: flex-start;"
>
<div class="archive__item-teaser-list" style="margin-right: 1em; flex-shrink: 0;">
<img src="/assets/figures/20_aquarium.png" alt="" style="display: block;">
</div>
<div style="flex-grow: 1;" >
<h2 class="archive__item-title no_toc" itemprop="headline" style="text-align: left !important;">
<a href="/research/aquarium-marl-environment/" rel="permalink">Aquarium MARL Environment
</a>
</h2>
<p class="page__meta">
<span class="page__meta-readtime">
<i class="far fa-clock" aria-hidden="true"></i>
1 minute read
</span>
</p>
<p class="archive__item-excerpt" itemprop="description" style="text-align: left;">
Aquarium: Open-source MARL environment for predator-prey studies.
</p>
</div>
</article>
</div>
<div class="list__item">
<article class="archive__item" itemscope itemtype="https://schema.org/CreativeWork"
style="display: flex; align-items: flex-start;"
>
<div class="archive__item-teaser-list" style="margin-right: 1em; flex-shrink: 0;">
<img src="/assets/images/projects/arch.png" alt="" style="display: block;">
</div>
<div style="flex-grow: 1;" >
<h2 class="archive__item-title no_toc" itemprop="headline" style="text-align: left !important;">
<a href="/projects/server-administration/" rel="permalink">LMU DevOps Admin
</a>
</h2>
<p class="page__meta">
<span class="page__meta-readtime">
<i class="far fa-clock" aria-hidden="true"></i>
2 minute read
</span>
</p>
<p class="archive__item-excerpt" itemprop="description" style="text-align: left;">
Managed LMU chair IT: Kubernetes, CI/CD, automation (2018-2023).
</p>
</div>
</article>
</div>
<div class="list__item">
<article class="archive__item" itemscope itemtype="https://schema.org/CreativeWork"
style="display: flex; align-items: flex-start;"
>
<div class="archive__item-teaser-list" style="margin-right: 1em; flex-shrink: 0;">
<img src="/assets/figures/19_binary_primates_teaser.jpg" alt="" style="display: block;">
</div>
<div style="flex-grow: 1;" >
<h2 class="archive__item-title no_toc" itemprop="headline" style="text-align: left !important;">
<a href="/research/primate-subsegment-sorting/" rel="permalink">Primate Subsegment Sorting
</a>
</h2>
<p class="page__meta">
<span class="page__meta-readtime">
<i class="far fa-clock" aria-hidden="true"></i>
1 minute read
</span>
</p>
<p class="archive__item-excerpt" itemprop="description" style="text-align: left;">
Binary subsegment presorting improves noisy primate sound classification.
</p>
</div>
</article>
</div>
</div>
</div>
</div>
</div>
<div id="footer" class="page__footer">
<footer>
<!-- start custom footer snippets -->
<!-- end custom footer snippets -->
<div class="page__footer-follow">
<ul class="social-icons">
<li><strong>Contact:</strong></li>
<li><a href="mailto:steffen.illium@ifi.lmu.de" rel="nofollow noopener noreferrer"><i class="fas fa-fw fa-envelope" aria-hidden="true"></i> Email</a></li>
<li><a href="https://www.linkedin.com/in/steffen-illium/" rel="nofollow noopener noreferrer"><i class="fab fa-fw fa-linkedin" aria-hidden="true"></i> LinkedIn</a></li>
<li><a href="https://gitea.steffenillium.de/steffen" rel="nofollow noopener noreferrer"><i class="fab fa-git-alt" aria-hidden="true"></i> Gitea</a></li>
<li><a href="/status/" rel="nofollow noopener noreferrer"><img src="https://uptime.steffenillium.de/api/badge/2/status" alt="System Status"></a></li>
</ul>
</div>
<div class="page__footer-copyright">&copy; 2025 <a href="http://localhost:4000">Steffen Illium</a>. Powered by <a href="https://jekyllrb.com" rel="nofollow">Jekyll</a> &amp; <a href="https://mademistakes.com/work/jekyll-themes/minimal-mistakes/" rel="nofollow">Minimal Mistakes</a>.</div>
</footer>
</div>
<script src="/assets/js/main.min.js"></script>
<script src="/assets/js/custom-bibtex.js"></script>
</body>
</html>