Abstract
Dynamic process modeling, system identification, and process monitoring constitute the analytical backbone of modern autonomous manufacturing. Contemporary industrial plants generate millisecond-rate sensor measurements that are highly collinear, strongly time dependent, and frequently nonlinear, yielding the observed data dimensionality far exceeds the true underlying dynamics. Accurate dynamic models are essential for optimal operation and life-cycle management, while effective process monitoring safeguards safety, quality, and profitability. The Industrial Internet of Things magnifies both the opportunities and the challenges inherent in this setting.Latent space, or reduced-dimensional modeling, offers a principled methodology by projecting measurements onto a compact set of latent variables that preserve essential dynamics. Nevertheless, conventional subspace identification and statistical latent variable methods encounter the following challenges: (i) Insufficient model parsimony: model parameterizations may not be parsimonious in high-order dynamic models. (ii) Collinearity-induced ill-conditioning: highly redundant sensing could inflate system condition numbers and degrade numerical robustness. (iii) Monolithic plant assumptions: treating the entire plant facilities as a single block could obscure directed interactions in large-scale networked plants. (iv) Linear model restrictions: pervasive nonlinearities could render linear models inadequate for modern process industries.
To address these challenges systematically, the dissertation proposes a coherent sequence of latent state-space modeling algorithms:
• Firstly, latent state space modeling with a canonical correlation analysis objective (LaSS-CCA) is proposed to incorporate stochastic subspace identification, realizing predictable latent variables, parsimonious model, and intrinsic dynamics. LaSS-CCA plays the fundamental role for subsequent algorithms.
• Secondly, latent state space identification (LaSSID) algorithm introduces causality enforcement within reduced dimension, effectively improving the condition numbers in collinear MIMO process systems.
• Thirdly, networked latent vector auto-regressive model with exogenous inputs (Net-LaVARX) and networked latent state space model (Net-LaSSIDy) recast the plant as a network of node-wise latent system modules that characterize the intra-node dynamics and inter-node interactions, enabling system identification with network topology.
• Finally, kernel latent vector auto-regressive (K-LaVAR) and kernel latent state space (KLaSS) models generalize the linear assumption to nonlinear regimes with layered structure, which achieve nonlinear representation, predictable latent variables extraction, and compact intrinsic dynamics.
The superiority of the proposed models has been validated by the following experiments: 1. The Eastman Chemical dataset; 2. The revamped Tennessee Eastman benchmark; 3. The Dow Chemical challenge problem; and 4. A multi-phase flow facility. The proposed models consistently outperform alternative methods in system identification accuracy and process monitoring effectiveness. The results advance latent-space modeling methodology and strengthen its practical relevance for modern industrial applications.
| Date of Award | 1 Sept 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Yining DONG (Supervisor) & S Joe QIN (Co-supervisor) |