Multirate Autoregressive Dynamic Latent Variable Model for Soft Sensing in Dynamic Processes

Using data-driven technology for soft sensing can significantly reduce the difficulty of obtaining key quality variables in industrial processes. Under real industrial conditions with diverse sensors, the measured data inevitably contain complex multi-rate and dynamic characteristics. To improve accuracy in such multi-rate dynamic processes, a multi-rate autoregressive dynamic latent variable model is proposed for dynamic soft sensing. The model is built using the expectation-maximization algorithm, introducing two sampling indicators to extract high-order auto-correlation and cross-correlation between multi-rate variables. The model effectiveness has been preliminarily validated through a numerical case, and it has also demonstrated excellent performance in two real cases: a petrochemical process and a papermaking wastewater treatment process.