A health monitoring and early fault detection method of rotating machines based on latent variables of diffusion model

Abstract

Deep generative models have gained prominence in the intelligent condition monitoring of rotating machines, primarily utilizing reconstruction errors as health indicators (HIs) to represent actual health conditions. This approach to constructions of HIs significantly influences the robustness and effectiveness of health monitoring and early fault detection. The incorporation of latent variables (LVs) is posited to alleviate these challenges. Nonetheless, the limited modeling capabilities of existing deep generative models constrain their ability to capture intricate patterns. In response, this paper introduces a novel health monitoring and fault prediction methodology leveraging the latent variables derived from diffusion models. A network architecture based on a multi-head self-attention mechanism (MHSA) is designed to effectively map time series monitoring data into the latent space. The diffusion model is initially trained using healthy monitoring samples. Subsequently, for each monitoring sample frame, the distributional differences in the latent space between the monitoring and healthy samples are analyzed to construct the HIs. A comprehensive quantitative and qualitative comparison of our proposed method is conducted against baseline models across multiple scenarios. The results underscore the superior robustness and effectiveness of our approach in the condition monitoring and fault prediction of rotating machines.