Detecting Syrup Brix in Cane Sugar Crystallization Using an Improved Least Squares Support Vector Machine

Yanmei Meng, Jian Cheng, Guancheng Lu, Zhihong TangInternational Journal of Control and Automation 2015 · DOI

TL;DR

This paper introduces an improved Least Squares Support Vector Machine (LS-SVM) model to accurately predict syrup brix in real-time during cane sugar crystallization. By incorporating auxiliary sensor data and addressing issues with unusual samples, their method achieves high accuracy and offers a more stable and affordable alternative to traditional brix sensors.

What Problem Does This Paper Address?

The manual operation of cane sugar boiling processes in China is largely due to the difficulty in accurately measuring syrup brix online. Existing brix sensors are expensive and prone to fouling, leading to unstable and discontinuous measurements. This lack of reliable online data hinders the implementation of advanced automation and control strategies, impacting production efficiency and quality. The paper aims to bridge this gap by developing a robust and cost-effective soft sensor that can predict syrup brix using readily available auxiliary process variables.

Methods

The core of this work is an improved Least Squares Support Vector Machine (LS-SVM) model. Standard LS-SVM is used as a foundation, but the authors introduce several key enhancements:

  • Auxiliary Variables: The model predicts syrup brix using five easily measurable auxiliary variables: vacuum degree, syrup temperature, steam pressure, steam temperature, and material flow rate.
  • Handling Unusual Samples: To improve model robustness, a weighting mechanism is introduced. Empirical errors of samples are assigned weights, effectively downplaying the influence of unusual or outlier data points during training. This is achieved by modifying the optimization objective function.
  • Improved Optimization: The optimization function of the LS-SVM is refined to enhance the model's generalization capacity. This involves making the coefficient matrix positive definite, simplifying calculations and improving stability.
  • Kernel Function: A Gaussian Radial Basis Function (RBF) kernel is employed to map the non-linear relationships between input variables and syrup brix into a high-dimensional feature space, allowing for linear separation.
  • Parameter Optimization: An improved Leave-One-Out Cross-Validation (LOO-CV) method, combined with a three-step search (TSS), is used for efficient and accurate optimization of model parameters. This approach aims to simplify calculations and improve performance compared to traditional methods.

Datasets

The study utilizes data collected from a self-regulating comprehensive experimental platform designed to simulate the cane sugar process. This platform includes a refractometric brix sensor (MRP E-Scant) for measuring actual brix. The paper does not mention the use of publicly available datasets or the creation of a new benchmark dataset.

Key Insights and Ideas

  • Soft Sensors as a Viable Alternative: The research demonstrates that soft sensors, leveraging readily available process data, can effectively replace expensive and unreliable direct sensors for critical parameters like syrup brix.
  • The Importance of Outlier Handling: The paper highlights that statistical models are sensitive to unusual data. The proposed weighting scheme for empirical errors is crucial for building a robust model that performs well even with imperfect real-world data.
  • Stability through Mathematical Refinement: By ensuring the coefficient matrix is positive definite, the authors not only simplify computations but also enhance the overall stability and reliability of the LS-SVM model.
  • Online Optimization for Dynamic Processes: The integration of on-line parameter optimization is key to adapting the soft sensor to the dynamic nature of the sugar crystallization process, ensuring continued accuracy over time.

Key Results

  • The developed PLS-SVM model achieved a maximal relative error (MRE) of approximately 2.5% when predicting syrup brix.
  • When compared to other algorithms (BP, SVM, LS-SVM) using 50 testing samples, the PLS-SVM model exhibited lower MRE (2.5%) and Mean Square Error (MSE) (1.17%).
  • The PLS-SVM model also demonstrated significantly faster computation time compared to the other algorithms.
  • Experimental validation on a self-regulating platform confirmed the model's accuracy and reliability, with the soft sensor output closely matching the measured brix.

Limitations and Future Work

  • Sensor Fouling Impact: The paper acknowledges that the performance of the actual refractometric brix sensor (MRP E-Scant) used for validation is significantly affected by fouling on its probe surface. While the soft sensor remains accurate, the direct measurement's reliability is a limitation.
  • Offline Parameter Optimization: While the paper discusses on-line optimization, some initial parameter tuning (like $\gamma$ and C) might still be performed offline, which could be a limitation in fully dynamic environments.
  • Future Work: The paper implicitly suggests that further refinement of the on-line optimization techniques and potentially exploring other kernel functions or advanced LS-SVM variants could lead to even greater improvements in accuracy and robustness. The development of a complete offline and online soft sensor system in VC++ 6.0 environment is also a notable outcome.

Summary updated

Abstract

The cane sugar boiling was a process involved syrup-sugar physical transformation, which had still been operated manually in the sugar production fields across China. The major reason for this was that the syrup brix in the process had not yet been measured accurately on-line. Aimed to solve this issue, an improved LS-SVM based soft sensor model for it was proposed. This model was able to predict the syrup brix based on five auxiliary variables. Since this was a statistical model, the unusual samples in the training set had to be removed, which could be achieved by adding weighting coefficients to the empirical errors of samples. Also, the optimization function of LS-SVM was improved to ensure the generalization capacity of the model. Then, this model was verified in a self-regulating comprehensive experimental platform, which proved that the predicted brix was close to the real brix where the maximal relative error was nearly 2.5%.
BibTeX
@article{meng2015detecting,
  author = {Yanmei Meng and Jian Cheng and Guancheng Lu and Zhihong Tang},
  title = {{Detecting Syrup Brix in Cane Sugar Crystallization Using an Improved Least Squares Support Vector Machine}},
  journal = {International Journal of Control and Automation},
  year = {2015},
  doi = {10.14257/ijca.2015.8.3.19},
}
Figure