An engineering control design approach based on high-dimensional data analysis in closed-loop systems for industrial networks

Shunping JiJournal of Combinatorial Mathematics and Combinatorial Computing 2025 · DOI

TL;DR

This paper introduces a novel engineering control approach for complex industrial networks by leveraging high-dimensional data analysis. It combines HOPLS-SVM for data processing and BP neural networks for prediction, enhanced by linear regression for error correction, to achieve more accurate and adaptive control in industrial processes, demonstrated on a chemical decomposition unit.

What Problem Does This Paper Address?

Modern industrial processes are increasingly large-scale, automated, and networked. This complexity generates vast amounts of high-dimensional data, posing significant challenges for traditional control systems. These challenges include managing the sheer volume of data, the computational complexity of analysis, and the need for accurate real-time prediction to ensure smooth, safe, and efficient operations. Existing methods struggle to effectively handle this high-dimensional data, leading to potential inefficiencies and safety risks.

Methods

The core of this approach is a multi-stage process:

  • High-Dimensional Data Analysis (HOPLS-SVM): The paper utilizes a combination of Higher-Order Partial Least Squares (HOPLS) and Support Vector Machines (SVMs). HOPLS is an extension of PLS designed for higher-order tensors (multidimensional arrays), allowing for more robust analysis of complex data structures. HOSVD (Higher Order Singular Value Decomposition) is employed within HOPLS for efficient computation. SVMs are then used for their powerful pattern recognition capabilities. This HOPLS-SVM framework is designed to analyze and process the high-dimensional data from industrial networks.
  • Real-time Prediction (BP Neural Network): The processed data from HOPLS-SVM serves as input for a Backpropagation (BP) neural network. This network is configured with specific transfer functions: a bipolar S-type (tansig) for the hidden layer and a unipolar S-type (logsig) for the output layer. The data is normalized to [-1, 1] before training and inverse normalized after prediction.
  • Error Correction (Linear Regression): To mitigate the potential for BP neural networks to fall into local optima and to further refine predictions, a linear regression model is employed. This model is trained using the prediction errors generated by the BP network. The final prediction is a synthesis of the BP network's prediction and the corrected error.

Datasets

The study uses actual industrial data from a chemical industry network, specifically from a decomposition unit of isopropyl benzene peroxide (CHP).

  • Training Data: 125 sets of field data samples were used for system model training.
  • Calibration Data: An additional 125 sets of field data samples were used for model calibration.
  • Evaluation Data: 20 sets of field data samples were used for analysis and verification.

The paper does not introduce a new benchmark dataset but rather applies its methodology to real-world industrial data.

Key Insights and Ideas

  • Synergy of HOPLS-SVM and BPNN: The combination of HOPLS-SVM for robust high-dimensional data processing and BP neural networks for prediction proves effective. HOPLS-SVM handles the complexity of multidimensional data, while the BPNN learns intricate patterns for forecasting.
  • Importance of Error Correction: Incorporating linear regression for error correction significantly enhances the accuracy of the BP neural network's predictions, addressing a common limitation of neural network models.
  • Adaptability through Data Updating: The paper suggests that methods incorporating sample updating (as opposed to static models like LWPLS-1) are crucial for maintaining prediction performance when engineering conditions change, highlighting the need for dynamic models in industrial settings.
  • Balancing Optimization Targets: The case study demonstrates the system's ability to optimize multiple, potentially conflicting, objectives (e.g., crude product temperature vs. flow rate) to ensure safety and efficiency.

Key Results

  • Prediction Accuracy: The proposed model achieved the smallest prediction error values among the compared methods, with a Mean Absolute Error (MAE) of 0.0347, Mean Squared Error (MSE) of 0.0176, and Mean Relative Error (MRE) of 0.0627.
  • Refining Industry Case Study: In a refining industry simulation, 35% of the data showed an error within ±5°C, and 70% within ±10°C, indicating good forecast temperature accuracy.
  • CHP Decomposition Unit Optimization: The engineering control system successfully regulated and optimized the crude product temperature and flow rate. It ensured the crude product flow rate stayed within the upper limit (25 m³/h) while keeping the temperature at a low level (93.37-93.65°C), balancing safety and output.

Limitations and Future Work

  • Limitations:
    • The paper notes that the choice of the number of hidden layer nodes in a BP neural network lacks a strict theoretical basis and is often determined empirically.
  • While the model shows good performance, the paper does not explicitly detail the computational cost or scalability for extremely large-scale, real-time industrial applications beyond the presented case study.
  • Future Work:
    • The authors suggest that their method can provide solutions to engineering control problems in complex industrial processes, implying potential for broader application.
  • The study's success in optimizing multiple parameters in the CHP decomposition unit suggests further exploration of multi-objective optimization within this framework.

Summary updated

Abstract

Industrial processes are constantly developing towards large-scale and automation, and the smooth, safe, high-quality, and efficient operation of industrial processes has become a hot spot of concern, and higher requirements have been put forward for the control of production processes. This study analyzes the high-dimensional data in the closed-loop system of industrial network based on the HOPLS-SVM algorithm with higher-order singular value decomposition method. A BP neural network model is constructed with the processed data as the input set to realize the real-time prediction of the main data in the project, and the error of the prediction model is corrected by using linear regression method, and then the project prediction control system is constructed by piggybacking on the model. The results show that the prediction performance of the model in this paper is better than that of the comparison model, and the average absolute error is only 0.0347. At the same time, it is found that the control system in the decomposition project of CHP is able to regulate and optimize the temperature and flow rate of the crude product, which ensures the balance between the product temperature and yield, and the safety of the project operation. The engineering control method designed in this study has strong adaptability and effectiveness, and can provide solutions to engineering control problems in complex industrial processes.
BibTeX
@article{ji2025engineering,
  author = {Shunping Ji},
  title = {{An engineering control design approach based on high-dimensional data analysis in closed-loop systems for industrial networks}},
  journal = {Journal of Combinatorial Mathematics and Combinatorial Computing},
  year = {2025},
  doi = {10.61091/jcmcc127a-045},
}
Figure