Hybrid Modeling and Simulation of the Grinding and Classification Process Driven by Multi-Source Compensation

Jiawei Yang, Guobin Zou, Junwu Zhou, Qingkai Wang, Tao Song, Kang LiMinerals 2024 · DOI

TL;DR

This paper introduces a novel hybrid modeling approach for the complex grinding and classification process in mineral processing. By combining mechanistic models with data-driven compensation using a deep learning network, they achieve accurate predictions of product quality, paving the way for improved process control and efficiency in industrial settings.

What Problem Does This Paper Address?

Mineral processing, particularly the grinding and classification stage, is crucial for extracting valuable materials but is notoriously complex and difficult to control. Traditional steady-state models struggle to capture the dynamic and uncertain nature of these processes, limiting their effectiveness in real-time control systems. The variability in ore properties and operating conditions introduces significant challenges, leading to fluctuations in product quality, reduced efficiency, and increased operational costs. This paper aims to bridge this gap by developing a dynamic, hybrid model that can accurately represent and predict the grinding and classification process, even in the face of these uncertainties.

Methods

The core of this research lies in a hybrid modeling approach that integrates two key components:

  • Mechanistic Modeling: The paper starts by establishing mechanistic models for the key equipment in the grinding and classification circuit, specifically the ball mill, cyclone, and pump pool. These models are based on fundamental principles like mass balance and established empirical relationships (e.g., population balance equations for breakage and selection functions in the ball mill, and Nageswararao's model for the cyclone). This provides a foundational understanding of the physical processes involved.
  • Data-Driven Compensation with Multi-Source Signals: To address the limitations of purely mechanistic models and account for unmodeled dynamics and uncertainties, a data-driven compensation mechanism is employed.
  • Multi-Source Signal Integration: The model leverages a wide array of real-time sensor data (e.g., feed rates, material levels, currents, frequencies, pressures, concentrations) as input.
  • Nonlinear Feature Extraction (DBDAE): To handle the complex, intercorrelated nature of these signals, a Dual Blind Denoising Autoencoder (DBDAE) is used. This deep learning architecture, combining CNN and GRU encoders, effectively reduces dimensionality and extracts nonlinear feature vectors from the multi-source signals, capturing hidden patterns and correlations.
  • Dynamic Radial Basis Function (RBF) Network Compensation: The extracted nonlinear features are then fed into a dynamic Radial Basis Function (RBF) network. This network learns to compensate for the errors and uncertainties of the mechanistic model, effectively creating a "digital twin" that can predict the process behavior. The RBF network training involves normalization, initial value assignment, sensitivity analysis, and iterative refinement of weights and centers.

The entire system is integrated into a simulation platform called BGRIMM Process Simulator—Grinding (BPSim-G), which includes modules for edge computing, simulation engine, human-computer interaction, and databases.

Datasets

The paper describes the use of real-time sensor data from a gold concentrator plant for industrial validation. This data is collected from the Distributed Control System (DCS) historical database and includes measurements for:

  • Cyclone overflow particle size
  • Cyclone overflow rate
  • Cyclone overflow concentration

The paper also mentions offline training, testing, and validation of the DBDAE model using input raw signals and reference signals, with quality coefficients (RSME, R², t/Epoch) reported in Table 3.

Key Insights and Ideas

  • Synergy of Mechanistic and Data-Driven Models: The most significant insight is the successful integration of physics-based models with data-driven techniques. This hybrid approach leverages the interpretability and foundational understanding of mechanistic models while using data-driven methods to capture complex, unmodeled dynamics and uncertainties.
  • Power of Multi-Source Signal Fusion: The paper highlights the importance of utilizing a broad spectrum of sensor data. By fusing these diverse signals, the model gains a more comprehensive understanding of the process state, leading to more robust predictions.
  • DBDAE for Effective Feature Engineering: The application of DBDAE for nonlinear feature extraction is a key enabler. It demonstrates how deep learning can effectively distill complex, high-dimensional sensor data into meaningful features that improve the performance of subsequent models.
  • Dynamic Compensation for Real-World Processes: The use of a dynamic RBF network for compensation is crucial. It allows the model to adapt to the time-varying nature of the grinding and classification process, which is essential for effective real-time control.
  • Industrial Validation is Key: The paper goes beyond simulation by validating its model on an actual gold concentrator. This practical application underscores the real-world applicability and reliability of the proposed method.

Key Results

  • High Model Fit: The simulation results show a high degree of fit to the real physical system.
  • Cyclone Overflow Particle Size: The deviation between simulated and measured cyclone overflow particle size was 4.04% after excluding abnormal data.
  • Cyclone Overflow Rate: The deviation between simulated and measured cyclone overflow flow rate was 1.62%.
  • Cyclone Overflow Concentration: A mean square error of 3.40% was achieved for the cyclone overflow concentration.
  • Improved Control: The model application demonstrated that adjusting the ball mill feed rate could effectively suppress fluctuations in cyclone overflow particle size, keeping it within the target range, even when the feed particle size varied due to crusher operations.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the steady-state model is limited by its structure, which is why the hybrid approach is necessary.
  • While the DBDAE model's accuracy meets online application requirements, the specific computation time per epoch (1.568s) might be a consideration for extremely high-frequency control loops.
  • The paper focuses on a specific gold concentrator; generalizability to other ore types or processing plants might require further validation.
  • Future Work:
    • The paper suggests that the simulation system is designed for industrial application and can be used for control rule design. This implies future work could focus on implementing and testing these control strategies in real-time.
  • The BPSim-G platform is designed to be extensible, allowing for model upgrades and the integration of third-party models, indicating potential for expanding the scope of the simulation.
  • Further research could explore optimizing the DBDAE architecture or exploring alternative deep learning models for feature extraction to potentially improve accuracy or reduce computational load.

Summary updated

Abstract

The grinding process is a key link in mineral processing production and a typical complex, controlled process. The steady-state model is limited by its model structure and thus difficult to applyied in a control system. A hybrid modeling method driven by multi-source compensation is proposed in this paper based on the mechanism model using key equipment in the grinding and classification process, addressing the uncertainties which affect the stability of the control systems. This method combines the relevant multi-source signals with uncertainties by using a priori knowledge, extracts the nonlinear feature vector in the signal through an unsupervised depth network, and constructs a compensation model based on dynamic radial basis function network to realize the integration of mechanism modeling and data-driven compensation modeling. The simulation results show that the model has a high degree of fit to the real physical system; the industrial validation was conducted at a gold concentrator, the grinding product quality was predicted and controlled with the model.
BibTeX
@inproceedings{yang2024hybrid,
  author = {Jiawei Yang and Guobin Zou and Junwu Zhou and Qingkai Wang and Tao Song and Kang Li},
  title = {{Hybrid Modeling and Simulation of the Grinding and Classification Process Driven by Multi-Source Compensation}},
  booktitle = {Minerals},
  year = {2024},
  doi = {10.3390/min14101019},
}
Figure