Unsupervised Anomaly Sound Analysis Method Using StyleGAN2 to Estimate the Degree and Type of Abnormalities

This study proposes a pioneering approach for anomaly detection in smart factories through the analysis of equipment acoustic data, leveraging the StyleGAN2-ADA deep generative model, originally developed for facial image generation, to analyze acoustic data. This method shows the potential to accurately estimate the severity and type of unusual sounds, allowing for early identification of equipment faults or wear and tear, and subsequently estimating their operational lifespan. Specifically, acoustic data from small motors are converted into Mel spectrograms, which are treated as images. By training the StyleGAN2-ADA model with these spectrograms and analyzing the resulting latent vectors, the technique classifies the nature and extent of anomalies in previously unseen acoustic data. Using unsupervised learning to analyze anomalous sounds holds significant potential for improving industrial monitoring and maintenance planning.