Munko, M.J. orcid.org/0009-0009-2040-5025, Cuthill, F. orcid.org/0000-0002-8674-1656, Camacho, M.A.V. et al. (2 more authors) (2025) An audio-based framework for anomaly detection in large-scale structural testing. Engineering Applications of Artificial Intelligence, 142. 109889. ISSN 0952-1976
Abstract
FastBlade is a research facility that tests large-scale composite and metal structures. To maximise its throughput by uninterrupted running of experiments, unmanned operation of the site is desired. One of its key enablers is anomaly detection, where microphones are used as a non-specific, affordable, and well-established sensing method. The dataset collected during the operation of the system consists of both normal and anomalous samples, which we need to classify. The problems associated with the dataset involve significant intraclass variability of the normal operation samples, as well as the scarcity of anomalous data, increasing the complexity of the classification problem. In this work, we evaluate the performance of several tools for time–frequency signal analysis, which are used to extract features from the original high-dimensional signal. We choose to apply the wavelet scattering transform (WST) due to its remarkable performance. Based on the findings from the literature review, we first rely on the reconstruction error of the processed WST images to detect anomalous samples. However, due to the nature of the dataset, both the convolutional autoencoder (CAE) and the principal component analysis (PCA) transform turn out to be unsuccessful. We then investigate the hidden layers of the CAE in search of features that can be used to separate normal and anomalous samples. Having identified the most suitable candidates, we discover that applying the normalised cross-correlation (NCC) to measure the similarity of the generic features generated and our dataset results in satisfactory separation. We train a number of classifiers and test the method on unseen data. The model’s accuracy is 99.58%, with a recall of 100% and 92% on normal and anomalous operation samples, respectively. The model’s accuracy and low latency prove the WST’s suitability for robust, real-time detection of different anomaly types. Therefore, the method can be deployed in systems with limited information about the critical assets and can be easily extrapolated to other setups.
Metadata
Item Type: | Article |
---|---|
Authors/Creators: |
|
Copyright, Publisher and Additional Information: | © 2024 The Authors. This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. |
Keywords: | Anomaly detection; Fatigue testing; Artificial neural network; Convolutional autoencoder; Wavelet scattering |
Dates: |
|
Institution: | The University of Sheffield |
Academic Units: | The University of Sheffield > Faculty of Engineering (Sheffield) |
Depositing User: | Symplectic Sheffield |
Date Deposited: | 02 Jan 2025 11:38 |
Last Modified: | 02 Jan 2025 11:38 |
Published Version: | https://doi.org/10.1016/j.engappai.2024.109889 |
Status: | Published |
Publisher: | Elsevier BV |
Refereed: | Yes |
Identification Number: | 10.1016/j.engappai.2024.109889 |
Open Archives Initiative ID (OAI ID): | oai:eprints.whiterose.ac.uk:221142 |
Download
Filename: An audio-based framework for anomaly detection.pdf
Licence: CC-BY 4.0