TL;DR
This paper introduces a novel method for precisely and continuously monitoring the stockline (material level) in blast furnaces, crucial for optimizing energy efficiency and production. By combining radar and mechanical probe data, and employing a new high-dimensional spatial feature extraction technique (MLRBFM) with a dynamic RBF neural network (SDSO-RBFNN), the system achieves significantly improved accuracy and reliability compared to existing methods, even under challenging industrial conditions.
What Problem Does This Paper Address?
Blast furnaces are massive, high-temperature, and dusty environments where maintaining the optimal material level (stockline) is critical for efficient iron production and energy conservation. Traditional methods for measuring this stockline have significant drawbacks: mechanical probes are discontinuous and can't operate during material charging, while radar probes, though non-contact and real-time, suffer from low accuracy, poor stability, and are easily affected by the harsh furnace environment (dust, airflow). This leads to inaccurate stockline information, hindering precise process control and potentially causing energy waste. The paper aims to bridge this gap by developing a system that provides continuous, accurate, and reliable stockline data by effectively integrating the strengths of both measurement techniques.
Methods
The proposed solution involves a multi-stage approach:
- Discrete Time Series Joint Partition: This initial step synchronizes and segments the data from mechanical and radar probes, accounting for their asynchronous sampling rates and the periodic nature of material charging.
- Maximum Likelihood Radial Basis Function Model (MLRBFM): This is a core novel contribution. MLRBFM is used to extract "high-dimensional spatial features" from the radar data. It models the radar data's distribution using a Gaussian normal distribution and then uses maximum likelihood estimation to capture complex, non-linear relationships within the data, effectively representing the stockline's spatial characteristics.
- Eigen QR Decomposition Direct Clustering Algorithm (EQRDD): This algorithm is used to cluster the input samples (which include spatial-temporal features from radar and periodic features) to determine optimal parameters for the subsequent neural network.
- Structural Dynamic Self-Optimization RBF Neural Network (SDSO-RBFNN): This is another key novel component. SDSO-RBFNN is designed to dynamically adapt its structure (number of hidden nodes, centers, and widths) based on the input data. It integrates the extracted high-dimensional spatial features and periodic features from the radar data, along with mechanical probe data, to achieve accurate and continuous stockline estimation. The "self-optimization" aspect allows the network to adjust its complexity and parameters in real-time, improving its robustness and accuracy.
Datasets
The study utilized actual measurement data from a 2650 m³ blast furnace in a steel plant. This included:
- Mechanical Probe Data: Sampled every 3-6 minutes, with 3393 data points used.
- Radar Probe Data: Sampled every 10 seconds, with 78,816 data points used.
The data was collected over 12 hours of real operation and also during periods of normal and abnormal working conditions. No new benchmark datasets were introduced; instead, existing industrial data was leveraged for validation.
Key Insights and Ideas
- High-Dimensional Spatial Features are Key: The paper strongly emphasizes that simply using time-series data isn't enough. The "high-dimensional spatial characteristics" of the stockline, captured by MLRBFM, are crucial for accurately reflecting the material's surface dynamics, especially during fluctuations.
- Dynamic Network Adaptation is Essential: The harsh and variable blast furnace environment necessitates a flexible model. SDSO-RBFNN's ability to dynamically adjust its structure (add/remove/merge nodes) is vital for handling changing conditions and maintaining accuracy without manual intervention.
- Fusion of Data Sources is Powerful: The synergy between mechanical probes (providing a stable, albeit discontinuous, reference) and radar probes (providing continuous, real-time data) is effectively harnessed. The proposed method uses the mechanical data to correct and validate the radar data, leading to superior results.
- MLRBFM's Robustness: The MLRBFM model, when properly tuned (specifically with m=10⁶ cycles), demonstrates a strong ability to extract meaningful features from radar data, even when the blast furnace is in abnormal operating conditions where radar data alone becomes highly unreliable.
Key Results
- Accuracy Improvement: The proposed SDSO-RBFNN method achieved a relative error of approximately ±5% and an absolute error within ±0.3 m, which is comparable to the mechanical probe's inherent error (±2%). This is a significant improvement over traditional RBFNN (relative error within ±10-18%) and raw radar data (relative error within ±20%, absolute error up to ±0.5 m).
- High Real-Time Performance: The method provides continuous and real-time stockline information.
- Robustness: The system demonstrated good performance and stability under both normal and abnormal working conditions, a critical requirement for industrial applications.
- Best Cycle Number: Ablation studies identified m = 10⁶ as the optimal number of cycles for MLRBFM to extract high-dimensional spatial features, balancing accuracy and model complexity.
Limitations and Future Work
- Extreme Abnormal Conditions: The paper acknowledges that under some extreme abnormal working conditions, the radar data can be severely affected by the blast furnace environment (high-frequency, high-amplitude jumps). This can lead to a loss of clustering characteristics and deviation from Gaussian distribution, hindering the effective extraction of spatial features and thus impacting the method's accuracy. This area requires further study and improvement.
- Data Availability: The authors mention that the industrial data used is not publicly available due to privacy concerns related to real-time industrial operation data. This might limit independent verification by other researchers.
The authors suggest that further research is needed to address the limitations encountered under the most extreme abnormal working conditions, aiming to improve the method's ability to track real material levels in such scenarios.
Summary updated