ScaloAdaptAlert, a novel framework for supervised anomaly detection in industrial acoustic data, integrating power scalograms, adaptive filter banks, and convolutional neural networks — A case study
M.A. Zakeri Harandi, Tzu-Yuan Lin, Chen Li, Sigurd Villumsen, Maani Ghaffari, Ole Madsen
Abstract
Acoustic data, as a modality for building data-driven industrial monitoring systems, is particularly notable for its comprehensive insights into both operational and machinery states of a process. However, the effectiveness of existing time–frequency representation (TFR)-based frameworks remains limited in industrial contexts. Originally designed for analyzing human speech and music signals, these frameworks often struggle with the complex, non-stationary, and non-harmonic nature of manufacturing sound data. Addressing these challenges, this paper introduces ‘ScaloAdaptAlert’ (SAdAlert), a novel, domain-agnostic framework for deriving time–frequency representations from industrial acoustic data. SAdAlert employs wavelet transform to capture both local and global spectral characteristics, uses Gaussian filter banks in an adaptive fashion to identify spectral features at both low and high frequencies, and applies max-pooling to reduce temporal dimensionality. The presented framework effectively preserves dominant information of the acoustic data while isolating its relevant features in noisy settings and addressing class imbalance. Our method, validated on a real-world anomaly detection dataset from a robotic screwing process, demonstrates superior performance compared to state-of-the-art deep learning models and conventional TFR methods. This validation underscores SAdAlert's potential to advance industrial acoustic monitoring by providing a robust, efficient, and highly adaptable tool for analyzing complex industrial acoustic data.