.png){kind=icon}
.png){kind=icon}
Return
Prototype-Driven Diffusion Transformer for Multimode Industrial Soft Sensing
.png){kind=icon}
PRE
AI
DOI：10.1109/tase.2026.3725874.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Soft sensors play a crucial role in modern process industries by providing virtual measurements of key quality variables from readily available process data. However, real industrial processes typically exhibit multiple operating conditions, strong nonlinearity, and pronounced time variation, while reliable mode labels are rarely available, making it difficult for conventional deep soft sensors to maintain accuracy and robustness. To address these challenges, this paper proposes a prototype-driven diffusion Transformer, termed PD-DiffFormer, for soft sensing in multimode industrial processes. PD-DiffFormer employs a Transformer-based encoder to extract temporal representations, constructs a set of learnable prototypes to form a discriminative multimodal latent space, uses a prototype-gated mixture-of-experts regressor to realize mode-adaptive prediction, and incorporates a mode-aware diffusion branch that performs multi-scale denoising of latent trajectories and imposes a structured temporal prior. By tightly coupling prototype-based multimodal representation learning with diffusion-based temporal regularization, the proposed framework achieves improved predictive accuracy and robustness under multimode operating conditions. Case studies on a power-plant gas turbine NOx emission dataset and a multiphase flow process dataset show that PD-DiffFormer consistently outperforms several representative baselines in terms of prediction accuracy and generalization. Note to Practitioners—Soft sensors are widely used in process industries for online estimation of quality variables, but their performance often degrades in practice due to multimode operation, strong nonlinearity, and the lack of reliable mode labels in historical data. In such settings, a single global model or a manually segmented per-mode model is difficult to maintain and may be highly sensitive to distribution shifts between training and field conditions. This paper proposes PD-DiffFormer, a prototype-driven diffusion Transformer that is designed for direct deployment in multimode industrial processes without requiring explicit mode labels. A Transformer encoder extracts temporal features, a learnable prototype space automatically discovers latent operating modes, and a prototype-gated mixture-of-experts regressor performs mode-adaptive prediction. A mode-aware diffusion branch further regularizes temporal dynamics, improving robustness under mode transitions. Case studies on a gas turbine NOx emission dataset and a multiphase flow process show that PD-DiffFormer achieves significantly lower prediction errors than common baselines, indicating that it can serve as a practical option for engineers who need stable, mode-adaptive soft sensors in complex plants.
Keywords:
Soft sensing
multimode industrial processes
diffusion models
transformer networks
industrial time-series modeling
Journal
IF:
6.4
Papers:
4.9K
Citations:
1.6W