Study on Multi-Model Soft Sensor Modeling Method and Its Model Optimization for the Fermentation Process of Pichia pastoris

Bo Wang, Xingyu Wang, Mengyi He, Xianglin Zhu2021 · DOI

TL;DR

This paper introduces a novel multi-model soft sensor approach for real-time monitoring of key variables in Pichia pastoris fermentation. By combining piecewise affine (PWA) modeling with an improved particle swarm optimization (ICF-PSO) and least squares support vector machines (LS-SVM), the method achieves significantly higher prediction accuracy and robustness compared to existing techniques.

What Problem Does This Paper Address?

The Pichia pastoris fermentation process is crucial for producing valuable proteins like protease K, especially in the context of COVID-19 diagnostics. However, directly measuring key biomass variables (like cell concentration and protease K concentration) in real-time is challenging and expensive. Traditional offline methods are time-consuming and can lead to inaccurate results. While soft sensors offer a cost-effective solution by estimating these variables from readily available sensor data, a single global model often struggles to capture the complex, nonlinear, and dynamic nature of fermentation processes, which typically involve distinct phases (adjustment, growth, stabilization). This paper aims to bridge this gap by developing a more accurate and robust soft sensor capable of handling these complexities.

Methods

The core of this work is a multi-model soft sensor based on Piecewise Affine (PWA) modeling. This approach breaks down the complex nonlinear system into several simpler, locally linear models, each active within a specific region of the input space.

Key methodological contributions include:

  • False Nearest Neighbor (FNN) method: Used to determine the optimal order (embedding dimension) of the PWA model, which is crucial for accurate representation.
  • Improved Compression Factor Particle Swarm Optimization (ICF-PSO): A novel optimization algorithm designed to overcome the limitations of standard PSO, particularly its tendency to get stuck in local optima. The ICF-PSO collaboratively optimizes two critical aspects:
  • The number of local models (S) required to represent the fermentation process.
  • The parameters of each local model (coefficients of the linear equations). The "improved compression factor" dynamically adjusts particle velocity, allowing for faster exploration early in the optimization and finer tuning later.

  • Least Squares Support Vector Machine (LS-SVM): Employed to determine the dividing surfaces that define the operational regions for each local PWA model. LS-SVM is chosen for its efficiency in handling small datasets and high-dimensional data.

The overall process involves determining the model order, collaboratively optimizing the number of models and their parameters using ICF-PSO, and then defining the boundaries of each local model's domain using LS-SVM.

Datasets

The study utilized real-world data collected from a microbial fermentation lab at Zhenjiang Yangzhong Weikete Bioengineering Equipment Co., Ltd., China. The experiments involved Pichia pastoris fermentation using the KM71 strain.

  • Data Collection: Environmental variables were monitored every 5 minutes, while cell concentration and protease K concentration were measured offline every 1 hour.
  • Data Preprocessing: A one-to-one correspondence between input and output variables was established through interpolation. Data was normalized to the range of -1 to 1.
  • Dataset Split: A total of 960 data points were obtained. This was divided into:
  • 576 samples for training.
  • 204 samples for verification.
  • 180 samples for testing.

Key Insights and Ideas

  • Decomposition-Synthesis for Fermentation: The paper demonstrates that breaking down the complex, multi-phase fermentation process into multiple simpler, locally linear models (PWA) is more effective than a single global model. This "decomposition-synthesis" approach allows for better capture of process dynamics.
  • ICF-PSO for Robust Optimization: The ICF-PSO algorithm is a significant improvement over standard PSO for this application. Its ability to dynamically adjust particle velocity helps avoid premature convergence and find more accurate solutions for both the number of models and their parameters simultaneously.
  • Synergy of PWA, ICF-PSO, and LS-SVM: The paper highlights the power of combining these techniques. PWA provides the modeling framework, ICF-PSO handles the complex multi-objective optimization, and LS-SVM precisely defines the boundaries between local models. This integrated approach leads to superior performance.
  • Three Models for Three Phases: The optimization process naturally identified three local models, which the authors correlate with the three distinct phases of the fermentation process (adjustment, logarithmic growth, and stabilization), underscoring the model's ability to adapt to changing process conditions.

Key Results

The proposed ICF-PSO-PWA soft sensor model demonstrated superior performance compared to a PSO-optimized Weighted Least Squares Support Vector Regression (PSO-WLSSVR) model.

  • Prediction Accuracy:
    • Cell concentration (X): Improved accuracy by 2.4642% (1.8932% vs. 4.3573%).
  • Protease K concentration (P): Improved accuracy by 6.5127% (2.5974% vs. 9.1101%).
  • Overall Mean Accuracy Improvement: 4.4884%.
  • Root Mean Square Errors (RMSEs): The ICF-PSO-PWA model consistently showed lower RMSEs across varying test dataset sizes, indicating greater stability and robustness. For instance, with 180 test samples, the RMSE for X was 0.0154 vs. 0.0335, and for P was 0.7113 vs. 2.2788.
  • Generalization Performance: The ICF-PSO-PWA model exhibited excellent generalization performance and self-adaptability, maintaining stable RMSEs even with increasing test data.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that determining the optimal order and number of models can be challenging without sufficient prior information.
  • While the ICF-PSO is an improvement, the performance of any optimization algorithm is sensitive to parameter tuning.
  • Future Work:
    • The authors express interest in studying the optimal control of PWA models.
  • They also aim to develop online monitoring systems for the entire fermentation process.
  • Further exploration of the ICF-PSO algorithm's performance with different parameter settings and test functions is suggested.

Summary updated

BibTeX
@misc{wang2021study,
  author = {Bo Wang and Xingyu Wang and Mengyi He and Xianglin Zhu},
  title = {{Study on Multi-Model Soft Sensor Modeling Method and Its Model Optimization for the Fermentation Process of Pichia pastoris}},
  year = {2021},
  doi = {10.3390/s21227635},
}
Figure