Just-in-Time Correntropy Soft Sensor with Noisy Data for Industrial Silicon Content Prediction

Kun Chen, Yu Liang, Zengliang Gao, Yi Liu2017 · DOI

TL;DR

This paper introduces a novel "Just-in-Time Correntropy Soft Sensor" (JCSVR) to predict silicon content in industrial blast furnaces, tackling the challenges of noisy, non-Gaussian, and unevenly distributed data. By integrating outlier detection with a local learning approach, JCSVR achieves more accurate and reliable predictions compared to traditional methods, making it suitable for long-term industrial applications.

What Problem Does This Paper Address?

Industrial blast furnaces are complex systems where precise control of silicon content in hot metal is crucial for stable production. However, accurately predicting this content online is challenging due to several factors:

  • Noisy and Non-Gaussian Data: Sensor readings are often corrupted by noise, and the data distribution deviates significantly from the ideal Gaussian assumption.
  • Uneven Data Distribution: Data can be clustered in certain operating regions, making it difficult for global models to generalize well.
  • Outliers: The presence of outliers, caused by issues like instrument degradation or transmission problems, can severely degrade the performance of predictive models, leading to over-fitting and unreliable predictions.
  • Dynamic Process: Blast furnace processes are dynamic, meaning that fixed, global models can become outdated and inaccurate over time.

Existing soft-sensing approaches often treat outlier detection and model building as separate tasks, which can be inefficient and less effective. The paper aims to bridge this gap by proposing a unified framework that simultaneously handles noisy data, outlier detection, and local, adaptive modeling.

Methods

The core of the proposed solution is the Just-in-Time Correntropy Soft Sensor (JCSVR). This method builds upon two key concepts:

  1. Correntropy Support Vector Regression (CSVR): This is a novel modeling framework that integrates the maximization of the correntropy criterion with Support Vector Regression (SVR). Correntropy is a robust measure of similarity that is less sensitive to outliers than traditional measures like mean squared error. By using correntropy, CSVR can automatically assign lower weights to outliers during the training process, effectively reducing their negative impact without explicit outlier removal. This allows for the simultaneous development of a soft sensor and detection of outliers.
  2. Just-in-Time (JIT) Learning with Clustering: To address the dynamic nature of the process and improve computational efficiency, JCSVR employs a local learning approach.
  3. Offline Stage: * The initial training data is first processed using the CSVR method to identify and down-weight outliers.
  4. The "clean" data is then clustered using a Fuzzy C-Means (FCM) algorithm into several sub-classes. Each sub-class represents data with similar characteristics, and its center is identified.
  5. Online Stage: * When a new query sample arrives, the system first determines which sub-class center it is closest to.
  6. Then, it searches for similar data points only within that specific sub-class using similarity criteria (like Euclidean distance).
  7. A JCSVR model is then constructed using this local set of similar data points to predict the silicon content for the query sample.
  8. Database Maintenance: The system is designed to incorporate new data over time, allowing the database to be updated and the models to remain relevant for long-term utilization.

Datasets

The paper utilizes a real-world industrial dataset from a blast furnace in China for training and testing the proposed JCSVR model.

  • Training Set: 240 data samples.
  • Testing Set: 200 data samples.

The dataset consists of input variables such as blast temperature, blast volume, gas permeability, top pressure, top temperature, and ore/coke ratio, along with the offline measured silicon content. The paper notes that the data is noisy and exhibits non-Gaussian distributions and nonlinear relationships.

Key Insights and Ideas

  • Unified Outlier Detection and Modeling: The most significant insight is the successful integration of outlier detection directly into the soft sensor modeling framework using CSVR. This avoids the need for separate, potentially complex, outlier preprocessing steps.
  • Robustness through Correntropy: The use of correntropy as a similarity measure provides inherent robustness against outliers, a critical advantage for industrial data.
  • Efficiency and Adaptability with JIT and Clustering: The combination of Just-in-Time learning and FCM clustering allows the model to focus on relevant local data for prediction, significantly improving computational efficiency and enabling adaptation to changing process dynamics.
  • Improved Long-Term Reliability: By continuously updating the database and employing local models, JCSVR offers better long-term prediction performance compared to static global models.

Key Results

The JCSVR method was compared against three other soft sensor models: JLSSVR (Just-in-Time Least Squares SVR), CSVR, and LSSVR. The evaluation was performed on the test set using Root Mean Square Error (RMSE), Relative RMSE (RE), and Hit Rate (HR).

  • Prediction Accuracy: JCSVR achieved the best prediction performance on the test set, with the lowest RMSE (0.079) and RE (17.70%), and the highest HR (80.5%).
  • Outlier Handling: The CSVR component effectively identified candidate outliers in the training data, with approximately 18.3% of the data being flagged.
  • Comparison: JCSVR outperformed JLSSVR, CSVR, and LSSVR across all metrics, demonstrating its superiority in handling noisy and dynamic industrial data. The box plot of prediction relative errors further illustrated JCSVR's narrower distribution and median closest to zero, indicating superior performance.

Limitations and Future Work

The authors acknowledge several areas for future research:

  • Exploring Other Correntropy Forms: Investigating different forms of correntropy could further adapt the method to various types of sensor uncertainties.
  • Efficient Feature Extraction: Developing efficient feature extraction methods specifically for noisy data could enhance model performance.
  • Incorporating Process Knowledge: Integrating domain-specific process knowledge into the modeling framework is suggested as a way to further improve prediction accuracy, though this is noted as a challenging task.
  • Model Selection Complexity: The paper mentions that model selection (e.g., choosing the number of clusters in FCM) can be relatively difficult for JCSVR, suggesting this as an area for potential improvement.

Summary updated

BibTeX
@misc{chen2017just,
  author = {Kun Chen and Yu Liang and Zengliang Gao and Yi Liu},
  title = {{Just-in-Time Correntropy Soft Sensor with Noisy Data for Industrial Silicon Content Prediction}},
  year = {2017},
  doi = {10.3390/s17081830},
}
Figure