An Audio-Based Framework for Anomaly Detection in Large-Scale Structural Testing

FastBlade is a research facility that tests large-scale composite and metal structures. To maximise its throughput by running experiments uninterruptedly, unmanned operation of the site is desired. In this work, we implement advanced signal processing and machine learning techniques to detect anomalous behaviour based on audio data collected on-site. First, we describe and contrast several tools for signal analysis in the time-frequency domains, acting as a kernel to extract features from the original high-dimensional signal. Among them, we select wavelet scattering transform (WST) due to its remarkable stability and low computational cost. Machine learning is incorporated into the signal-processing pipeline to classify whether an incoming sample represents normal or abnormal operation. To this end, we train a convolutional autoencoder (CAE) with a latent code of 24 dimensions on features extracted from normal operation data. Later, using a different set of normal operation data, we set a threshold for the CAE reconstruction error, which is used as a criterion to set normal and abnormal samples apart. The model correctly classifies most abnormal datasets as 100\% anomalous, while the manually chosen reconstruction error threshold determines the sensitivity for datasets with transient faults. The successful implementation of WST and CAE showcases their suitability for real-time detection of various anomaly types, thereby enhancing the efficiency and reliability of structural testing at FastBlade. Moreover, the flexibility of this method allows for deployment in systems with limited information about critical assets and easy extrapolation to other setups, underscoring its potential for broader application in engineering contexts.