An online intelligent detection method for slurry density in concept drift data streams based on collaborative computing

Lanhao Wang, Hao Wang, Tao Wei, Wei Dai, Hongyan WangPeerJ 2025 · DOI

TL;DR

This paper introduces a novel online method for accurately detecting slurry density in industrial settings, even when the underlying data patterns change over time (concept drift). By combining a mechanistic model with a data-driven approach and leveraging collaborative edge-cloud computing, the method significantly improves detection accuracy and adaptability in dynamic environments.

What Problem Does This Paper Address?

In industrial processes like mineral processing, accurately monitoring slurry density is crucial for efficiency and product quality. However, the data used to train detection models often changes over time due to factors like raw material variations, equipment aging, or operational adjustments. This phenomenon, known as "concept drift," causes traditional models to degrade in performance, leading to inaccurate readings and potential production issues. Existing methods struggle to adapt in real-time to these dynamic changes, necessitating a more robust and adaptive approach.

Methods

The proposed solution combines two key modeling strategies:

  • Mechanistic Model (Linear Component): This part uses Gaussian Process Regression (GPR) with a sliding window mechanism. This allows the model to continuously update its parameters based on recent data, effectively reducing the influence of older, potentially irrelevant information.
  • Data-Driven Model (Nonlinear Component): This part employs a Forgetting Mechanism Regularized Stochastic Configuration (FMRSC) algorithm. This algorithm is an advancement of Stochastic Configuration Networks (SCN) and incorporates a forgetting mechanism to prioritize recent data for updating output weights. It also uses network pruning and stochastic configuration to dynamically adjust the model's structure, making it more adaptable to changing data characteristics.

These two models work in tandem, with their outputs combined to provide the final slurry density estimate. The entire system is deployed using a collaborative edge-cloud computing framework, where edge devices handle real-time data acquisition and initial processing, while the cloud platform manages model training, updates, and deployment.

Datasets

The study utilized industrial process data collected from the grinding and classification stages of a mineral processing operation. This data included high-pressure and low-pressure sensor readings, motor current, and motor voltage. The dataset was preprocessed using Min-Max Normalization and was split into an initial training set and a streaming dataset for evaluation. The paper does not introduce a new benchmark dataset but uses real-world industrial data.

Key Insights and Ideas

  • Concept Drift is a Major Hurdle: The paper clearly demonstrates how concept drift significantly degrades the performance of static models in industrial settings.
  • Hybrid Modeling is Powerful: Combining a mechanistic model (physics-based) with a data-driven model (learning-based) offers a synergistic approach to capture both underlying physical principles and evolving data patterns.
  • Online Adaptation is Crucial: The success of the proposed method hinges on its ability to continuously update and adapt its parameters and structure in real-time, rather than relying on periodic offline retraining.
  • Edge-Cloud Collaboration Enhances Efficiency: Distributing processing between edge devices and the cloud allows for real-time monitoring and analysis while leveraging the cloud's computational power for robust model management.
  • Dynamic Structure Adjustment Matters: Beyond parameter updates, the ability to dynamically adjust the model's structure (e.g., through pruning) is key to maintaining performance when data characteristics change drastically.

Key Results

The proposed OGPRSWM-FMRSC method demonstrated superior performance across various metrics compared to other evaluated models.

  • Accuracy: Achieved the lowest Mean Absolute Error (MAE) of 6.11 and Root Mean Squared Error (RMSE) of 7.56.
  • Fit: Reached an R² value of 99.40%, indicating a high degree of agreement between model estimates and actual data.
  • Robustness: Showed a high qualification rate (above 95%) with relative errors less than 2% over 5 months of operation.
  • Concept Drift Handling: The sliding window mechanism and forgetting mechanism were effective in handling both sudden and gradual concept drift, outperforming models without these features.
  • Error Characteristics: The error distribution was centered around zero with a unimodal peak, and the autocorrelation function approached white noise, suggesting good generalization and minimal systematic bias.

Limitations and Future Work

  • Limitations: While the paper demonstrates strong performance, the specific industrial context (mineral processing) might limit direct applicability to vastly different domains without re-calibration. The computational complexity of some operations, though optimized, could still be a factor in extremely resource-constrained edge devices.
  • Future Work: The authors suggest that the proposed model framework is versatile and can be applied to other industrial domains, such as sensor data monitoring for equipment fault prediction and dynamic load regulation in energy management. Further exploration into optimizing the computational cost for even more constrained environments could also be a future direction.

Summary updated

Abstract

In industrial environments, slurry density detection models often suffer from performance degradation due to concept drift. To address this, this article proposes an intelligent detection method tailored for slurry density in concept drift data streams. The method begins by building a model using Gaussian process regression (GPR) combined with regularized stochastic configuration. A sliding window-based online GPR is then applied to update the linear model's parameters, while a forgetting mechanism enables online recursive updates for the nonlinear model. Network pruning and stochastic configuration techniques dynamically adjust the nonlinear model's structure. These approaches enhance the mechanistic model's ability to capture dynamic relationships and reduce the data-driven model's reliance on outdated data. By focusing on recent data to reflect current operating conditions, the method effectively mitigates concept drift in complex process data. Additionally, the method is applied in industrial settings through collaborative computing, ensuring real-time slurry density detection and model adaptability. Experimental results on industrial data show that the proposed method outperforms other algorithms in all density estimation metrics, significantly improving slurry density detection accuracy.
BibTeX
@inproceedings{wang2025online,
  author = {Lanhao Wang and Hao Wang and Tao Wei and Wei Dai and Hongyan Wang},
  title = {{An online intelligent detection method for slurry density in concept drift data streams based on collaborative computing}},
  booktitle = {PeerJ},
  year = {2025},
  doi = {10.7717/peerj-cs.2683},
}
Figure