Developing multistep-ahead quality prediction models for early warning

Abstract

In the manufacturing industry, product quality is a critical variable to monitor and control. Chemical plants, often large-scale to meet market demands, face challenges due to inherent nonlinearities and slow dynamics. These issues can hinder conventional methods from identifying disturbances before significant deviations in product quality occur, making early warning systems essential for preventing defects. This study introduces a sophisticated two-step multi-step ahead quality prediction framework for early warning. In the first step, a multi-step nonlinear state-space model (MS-NSSM) utilizes a lower-dimensional latent space with reduced noise to capture dynamic information from past process variables for future multi-step process variable prediction. In the second step, a regression variational autoencoder (Reg-VAE) uses these predicted future process variables to establish the process-quality relationship, enabling future multi-step quality prediction through another lower-dimensional latent space. The proposed method’s effectiveness is demonstrated through numerical simulations and real-world industrial cases. Performance metrics show a significant average accuracy for five prediction steps, with the proposed method achieving an R² value of 0.81 and MAE of 0.26 in numerical cases. In the industrial cases, the proposed method achieves an R² value of 0.995 and MAE of 0.03. The proposed method outperforms the comparative methods by providing early warnings 30 time points before offline laboratory tests and shows substantial potential for improving quality monitoring in industrial applications.