Speaker
Description
Anomalous sound detection (ASD) is a key task in industrial applications, where reliable detection of machine faults can prevent costly downtime and ensure operational safety. Although single-channel recordings simplify data acquisition and model design, they often do not capture spatial information that can be critical for distinguishing between normal and anomalous acoustic events in complex industrial environments. To address this, we propose a multichannel approach leveraging inter-channel correlations. Two autoencoder-based models are presented: in both, one channel serves as a reference, and differences with the remaining channels are computed. Two encoders process the reference and the remaining channels, respectively, while decoder designs vary—one model uses separate decoders for reference and remaining channels, and the other employs a single decoder to reconstruct all channels jointly. Our approach captures richer machine behavior representations, improving robustness and detection performance in real-world industrial settings.