TL;DR
Researchers have developed a high-precision "soft sensing" model called ILGSSA-LSSVM to predict the surface temperature of steel billets during the continuous casting process. By combining a Least Square Support Vector Machine (LSSVM) with a bio-inspired Sparrow Search Algorithm—further refined by chaotic mapping and trigonometric optimization—the system achieves a mean prediction error of just 0.058°C, providing a reliable "virtual sensor" for harsh industrial environments.
What Problem Does This Paper Address?
In the steel industry, "continuous casting" turns molten steel into solid billets. A critical stage is secondary cooling, where water is sprayed on the hot metal. To prevent internal cracks and ensure high quality, engineers need to know the billet's surface temperature in real-time to adjust the water flow.
However, the environment inside a casting machine is a nightmare for sensors. High temperatures, thick water mist, and "scale" (peeled iron oxide) on the billet surface make contact thermometers impossible to use and infrared cameras highly inaccurate. This lack of reliable data creates a "blind spot" in the control loop. This paper addresses the need for a soft sensing approach: using easily measurable data (like water pressure and flow rates) to mathematically estimate the temperature that cannot be measured directly.
Methods
The authors built their solution on a Least Square Support Vector Machine (LSSVM), a powerful machine learning tool for regression. The performance of an LSSVM depends heavily on two "hyperparameters" (the penalty factor $\gamma$ and kernel parameter $\sigma$). Finding the perfect values for these is usually a trial-and-error process.
To automate and optimize this, the authors introduced a hybrid optimization chain:
- Sparrow Search Algorithm (SSA): A bio-inspired algorithm that mimics how sparrows forage and avoid predators to find the "best" mathematical solution.
- Improved Logistic Chaotic Mapping: Standard SSA often suffers from poor "initialization"—the "sparrows" start in a clump, missing parts of the search space. The authors used chaotic mapping to ensure the algorithm starts with a more uniform, diverse spread of potential solutions.
- Golden Sine Algorithm: To prevent the algorithm from getting stuck in "local optima" (good-but-not-best solutions), they integrated the Golden Sine Algorithm, which uses trigonometric functions to scan the search space more effectively.
The resulting model, ILGSSA-LSSVM, uses 12 input variables—including molten steel temperature, cooling water flow, and valve positions—to predict the surface temperature.
Datasets
The model was trained and validated using real-world industrial data:
- Source: Fujian Sanming Steel Co., Ltd.
- Material: Q235 steel billets (150 mm × 150 mm).
- Scale: A massive dataset of 255,000 data points collected over 30 days in June 2022.
- Structure: The data was divided into 850 groups, with 200 samples per group used for training and 100 for testing.
- Ground Truth: For validation, the authors used an infrared colorimetric thermometer at the exit site, applying heavy data filtering and emissivity correction to ensure the "true" temperature was as accurate as possible.
Key Insights and Ideas
- Initialization Matters: The most significant "aha moment" is the demonstration of how much initial population distribution affects AI performance. By using chaotic mapping (Fig. 1 & 2), the authors showed that a more "randomized" start leads to a much more stable and accurate final model.
- Efficiency vs. Accuracy: The ILGSSA model reached stability in just 12 iterations, compared to 18 for the standard SSA. In an industrial setting, faster convergence means the model can be updated or retrained more efficiently.
- LSSVM > Neural Networks: The study found that LSSVM-based models significantly outperformed traditional Back-Propagation (BP) neural networks for this specific task, likely because LSSVM is better at handling the smaller, non-linear datasets typical of specific industrial batches.
Key Results
The ILGSSA-LSSVM model was compared against five other methods (including standard LSSVM, BP networks, and Gray Wolf optimized models). The results were decisive:
- Mean Error: Only 0.05805 °C, significantly lower than the BP network's 4.79 °C.
- Maximum Error: Capped at 3.857 °C, whereas other models strayed by as much as 9.47 °C.
- Fit Quality ($R^2$): Achieved a score of 0.900, indicating a very high correlation between predicted and actual temperatures.
- Robustness: The error fluctuations (Fig. 7) were much tighter for the proposed model, suggesting it is less likely to produce "wild" incorrect guesses during production.
Limitations and Future Work
While the results are impressive, a few limitations remain:
- Steel Grade Specificity: The model was tested specifically on Q235 steel. Different alloys have different thermal properties, meaning the model might need recalibration for different products.
- Sensor Dependency: As a "soft sensor," the model is only as good as its inputs. If the water flow meters or pressure sensors in the factory drift or fail, the temperature prediction will also fail.
- Future Directions: The authors suggest this model provides a foundation for a fully automated, closed-loop secondary cooling control system, where the AI doesn't just predict the temperature but actively adjusts the valves to maintain it.
Summary updated