Generative Latent Space Modeling of Stamping Sound for Anomaly Detection and Predictive Maintenance

Abstract

Operational condition monitoring for predictive maintenance is crucial to enable early detection of abnormal behavior and prevention of costly downtime. In industry settings, normal operation data are abundant, whereas abnormal or failure data are rarely available, limiting the development of reliable data-driven diagnostic models. Generative modeling provides a promising approach to address this imbalance by learning the latent distribution of normal operation and defining the boundary of potential anomalies. This study proposes a generative latent space framework for sound-based condition monitoring of a stamping machine using an internal sound sensor (ISS). The framework incorporates MTConnect for machine context as conditional information to guide the generative process. Sound sensing offers a low-cost, non-intrusive, and high-resolution modality, while MTConnect context provides synchronized operational states. The latent representation of sound features is organized around part-level representative signals, forming structured regions that capture characteristic acoustic patterns of normal operation. This structure guides the generative process toward realistic variations of normal sound while limiting deviations toward implausible samples. The proposed framework is implemented with two alternative generative architectures, Conditional Variational Autoencoder (CVAE) and Conditional Diffusion Model (CDM), evaluated comparatively through ablation studies. Experimental results show that retraining the encoder with the synthesized data enhances separation of part-specific sound representations and improves discrimination between normal and abnormal operation states. These findings demonstrate that the proposed framework effectively leverages normal sound data to define anomaly boundaries and strengthen the robustness of sound-based predictive maintenance systems for stamping operations.