Adaptive temporal diffusion-based reconstruction model for industrial dynamic uncertain process monitoring

Abstract

A key property of industrial processes is that they are often related to the dynamic uncertainty behaviors of the measurement data (e.g., sensor performance degradation or environmental changes), which poses significant challenges for traditional uncertainty-based monitoring research that typically assumes the measurement data exhibits invariant uncertainty. This study addresses this challenge by enriching the corrupted state of the data. We propose a novel diffusion-based method called the Dynamic Uncertainty Process Monitoring (DUPM) method. DUPM consists of a temporal diffusion convolutional network (TDCN) module and an adaptive diffusion reconstruction (ADR) module. First, in TDCN module, the diffusion process enhances the pattern coverage by gradually corrupting the original data, enabling the model to cover data under different uncertainties. Then an unsupervised backbone is designed to extract the latent temporal features of the input data and remove the noise in generation process, in which a nonlinear autoencoder is equipped with a one-dimensional convolution operation. The monitoring threshold is determined based on the reconstruction error after the diffusion process and generation process. Finally, the ADR module determines the number of steps to add noise in the online stage by calculating the similarity between the online data and the historical diffusion states. In this way, the reconstruction error can be used as a monitoring score. Experiments conducted on numerical simulations and the real-world MetroPT-3 dataset show that DUPM reduces the false alarm rate by at least 3% compared to the comparison method while maintaining fault detection rates. The verification indicates that the proposed method has potential in monitoring dynamic uncertain industrial processes.