基于LSTM-KAN与光谱技术的转炉出钢钢水温度反演

Molten Steel Temperature Inversion during Converter Tapping Based on LSTM-KAN and Spectral Technology

摘要: 针对钢铁冶炼转炉出钢阶段，高温辐射、高浓度烟尘及复杂动态环境导致钢水温度难以精确、连续测量的问题，本文提出一种基于光谱技术与监督学习双阶段架构的远程非接触式温度反演方法。首先，为解决工业现场无效背景干扰，构建了基于梯度提升机(GBM)的信号识别模型，实现对有效钢水辐射光谱的精准筛选；其次，针对提取的有效光谱数据，通过K-means聚类结合肘部法则确定波段划分，并利用核主成分分析(KPCA)与Lasso回归提取非线性显著特征；最后，提出一种融合长短时记忆网络(LSTM)与科尔莫戈罗夫–阿诺德网络(KAN)的混合模型(LSTM-KAN)，利用LSTM捕获光谱的时序演化特性，并结合KAN增强对光谱强度与温度间复杂映射关系的非线性建模能力。实验结果表明，该双阶段框架具有极强的鲁棒性，信号识别准确率达97.17%；温度反演模型预测精度高，平均绝对误差(MAE)仅为4.9443℃，误差超过15℃的样本比例降至1.08%。该研究为复杂工业环境下金属熔池温度的在线监测提供了一种高精度、强鲁棒的智能化方案。

Abstract: To address the challenges of accurate and continuous temperature measurement of molten steel during the converter tapping process—characterized by high-temperature radiation, dense dust, and dynamic environmental interference—this paper proposes a remote non-contact temperature inversion method based on a two-stage supervised learning framework. First, to mitigate invalid background interference in industrial sites, a signal recognition model based on Gradient Boosting Machine (GBM) is constructed to precisely filter valid radiation spectra of molten steel. Second, for the extracted spectral data, the optimal band division is determined using K-means clustering guided by the elbow method, and non-linear salient features are extracted via Kernel Principal Component Analysis (KPCA) and Lasso regression. Finally, a hybrid model combining Long Short-Term Memory (LSTM) and Kolmogorov-Arnold Network (KAN), termed LSTM-KAN, is developed. This model leverages LSTM to capture the temporal evolution characteristics of spectral data and employs KAN to enhance the non-linear modeling of the complex mapping between spectral intensity and temperature. Experimental results demonstrate that the proposed two-stage framework exhibits superior robustness, with a signal recognition accuracy of 97.17%. The temperature inversion model achieves high precision, with a Mean Absolute Error (MAE) of 4.9443˚C and only 1.08% of samples exceeding a 15˚C error margin. This research provides a high-precision and robust intelligent solution for the online monitoring of metal molten pool temperatures in complex industrial environments.