UltraModel: A Modeling Paradigm for Industrial Objects

Haoran Yang, Yinan Zhang, Qunshan He, Yuqi Ye, Jing Zhao, Wenhai WangIJCAI 2025 · OpenReview

Abstract

As Industrial 4.0 unfolds and digital twin technology rapidly advances, modeling techniques that can abstract real-world industrial objects into accurate and robust models, referred to as modeling for industrial objects (MIO) tasks, have become increasingly crucial. However, existing works still face two major limitations. First, each of these works primarily focuses on modeling a specific industrial object. When the industrial objects change, the proposed methods often struggle to adapt. Second, they fail to fully consider latent relationships within industrial data, limiting the model’s ability to leverage the data and resulting in suboptimal performance. To address these issues, we propose a novel modeling paradigm tailored for MIO tasks, named UltraModel. Specifically, a twin model graph module is designed to construct a customized graph based on the mechanisms of industrial objects and employ graph convolution to generate high-dimensional representations. Then, a multi-scale feature abstraction module and a spatial attention-based feature fusion module are proposed to complement each other in performing multi-scale feature abstraction and fusion on high-dimensional representations. Finally, the outputs are obtained by processing the fused representations through a feedforward network. Experiments on two different industrial objects demonstrate our UltraModel outperforms existing methods, offering a novel perspective for addressing industrial modeling challenges.

Introduction

In the era of Industry 4.0, the "Digital Twin"—a virtual replica of a physical asset—is essential for operational efficiency. However, Modeling for Industrial Objects (MIO) is notoriously difficult. Traditional approaches often force a choice between rigid, computationally expensive physics-based "mechanistic" models and "black-box" machine learning models that lack generalizability and fail to capture the complex physical logic of industrial systems.

Industrial modeling faces two primary challenges: the Generalization Gap, where models are too bespoke for specific equipment, and the Latent Relationship Problem, where standard deep learning models treat interconnected sensor data as independent inputs. UltraModel addresses these by using the physical mechanism of the object to guide the architecture of the AI, creating a "lightweight" yet accurate data-driven model informed by domain knowledge.

Methodology

The UltraModel architecture is a novel framework that combines Graph Neural Networks (GNNs) with multi-scale feature abstraction and spatial attention. It consists of three primary modules:

  1. Twin Model Graph (TwG): This module treats each sensor variable as a node in a graph. Relationships are categorized as fully clear (known physics), partially clear, or unclear. Graph Convolutional Networks (GCNs) enable these nodes to communicate, generating representations that respect the system's physical topology.
  2. Multi-Scale Feature Abstraction (MSFA): To capture industrial processes occurring at different temporal scales, this module utilizes parallel CNN-based blocks with varying kernel sizes and residual dense blocks.
  3. Spatial Attention-based Feature Fusion (SAFF): This module employs a spatial attention mechanism (using Global Average and Max Pooling) to intelligently weight and fuse features, ensuring the most relevant data informs the final prediction.
The UltraModel architecture overview, featuring the TwG, MSFA, and SAFF modules.

Experimental Evaluation

UltraModel was validated on two distinct datasets: DIS-COL (Distillation Column), representing complex industrial equipment, and ACE-FAC (Acentric Factor), representing abstract thermodynamic properties.

The results demonstrate that UltraModel consistently outperforms a wide array of baselines, including XGBoost, LightGBM, Transformers, and various GNN architectures like GAT and RADA. On the DIS-COL dataset, UltraModel achieved a 38.04% performance improvement over the Transformer model, the strongest baseline. On the ACE-FAC dataset, it achieved a 95.09% Model Improvement Rate (MIR) over the traditional Lee-Kesler mechanistic benchmark.

Comparison of Mean Absolute Error (MAE) across different models on the DIS-COL dataset.

Discussion and Conclusion

The success of UltraModel lies in its Physics-Informed Topology, which injects domain expertise into the neural network structure without the rigidity of pure equations. By defaulting to fully connected graphs for "unclear" systems, it maintains flexibility. Ablation studies confirmed that the synergy between graph-based communication and attention-based fusion is critical to its performance.

Future work will focus on reducing the computational complexity of GCNs for edge deployment, automating graph construction from raw data, and adapting the paradigm for highly volatile time-series data in dynamic systems. UltraModel offers a robust and generalizable path forward for the development of industrial digital twins.

Summary updated

Abstract

As Industrial 4.0 unfolds and digital twin technology rapidly advances, modeling techniques that can abstract real-world industrial objects into accurate and robust models, referred to modeling for industrial objects (MIO) tasks, have become increasingly crucial. However, existing works still face two major limitations. First, each of these works primarily focuses on modeling a specific industrial object. When the industrial objects change, the proposed methods often struggle to adapt. Second, they fail to fully consider latent relationships within industrial data, limiting the model’s ability to leverage the data and resulting in suboptimal performance. To address these issues, we propose a novel modeling paradigm tailored for MIO tasks, named UltraModel. Specifically, a twin model graph module is designed to construct a customized graph based on the mechanisms of industrial objects and employ graph convolution to generate high-dimensional representations. Then, a multi-scale feature abstraction module and a spatial attention-based feature fusion module are proposed to complement each other in performing multi-scale feature abstraction and fusion on high-dimensional representations. Finally, the outputs are obtained by processing the fused representations through a feedforward network. Experiments on two different industrial objects demonstrate our UltraModel outperforms existing methods, offering a novel perspective for addressing industrial modeling challenges.
BibTeX
@inproceedings{yang2025ultramodel,
  author = {Haoran Yang and Yinan Zhang and Qunshan He and Yuqi Ye and Jing Zhao and Wenhai Wang},
  title = {{UltraModel: A Modeling Paradigm for Industrial Objects}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=Ju6WRFNyRn},
}
Figure