Vibration-Based Anomaly Detection in Industrial Machines: A Comparison of Autoencoders and Latent Spaces

In industrial settings, machinery components inevitably wear and degrade due to friction between moving parts. To address this, various maintenance strategies, including corrective, preventive, and predictive maintenance, are commonly employed. This paper focuses on predictive maintenance through vibration analysis, utilizing data-driven models. This study explores the application of unsupervised learning methods, particularly Convolutional Autoencoders (CAEs) and variational Autoencoders (VAEs), for anomaly detection (AD) in vibration signals. By transforming vibration signals into images using the Synchrosqueezing Transform (SST), this research leverages the strengths of convolutional neural networks (CNNs) in image processing, which have proven effective in AD, especially at the pixel level. The methodology involves training CAEs and VAEs on data from machinery in healthy condition and testing them on new data samples representing different levels of system degradation. The results indicate that models with spatial latent spaces outperform those with dense latent spaces in terms of reconstruction accuracy and AD capabilities. However, VAEs did not yield satisfactory results, likely because reconstruction-based metrics are not entirely useful for AD purposes in such models. This study also highlights the potential of ReLU residuals in enhancing the visibility of anomalies. The data used in this study are openly available.