Diffusion Model-Driven Data Augmentation for Multivariate Time Series Process Anomaly Detection in Continuous Reactors

Abstract

The manufacturing of chemicals, pharmaceuticals, and energy materials requires continuous reactors, but early anomaly identification in rare-fault situations has proven challenging because of their nonlinear kinetics and high coupling of heat and mass transport. A data-augmented diffusion model for process anomaly identification in continuous reactors is proposed in this research. The technique creates physically realistic rare-anomaly samples for training an anomaly-sensitive temporal discriminator, models multivariate reactor trajectories, and learns a restricted denoising process from normal and abnormal operating windows. In comparison to LSTM, Transformer, variational autoencoder, GAN augmentation, and oversampling baselines, experiments on continuous reactor process data demonstrate that the suggested method increases macro-F1 from 89.2% to 95.1%, improves detection accuracy from 91.4% to 96.7%, and decreases missed detection for minority fault modes by 31.8%. The macro-F1 remains above 93.0% in the presence of sensor noise and missing variables, according to robustness testing. The results demonstrate that diffusion-based sample augmentation may extend the boundaries of a rare abnormality without altering the usual process dynamics; thus, a workable data-centric approach for dependable continuous reactor monitoring has been offered.