对转炉炼钢终点的实时精准控制能够有效提高钢铁产出的质量, 炉口火焰光谱在炼钢不同时期的变化明显, 对其进行分析处理并与机器学习方法相结合可有效用于炼钢终点的实时控制。 针对炉口火焰光谱数据量大、 现有方法对光谱特征提取在可信度和实时性上不足的缺陷, 提出一种基于窗口竞争性自适应重加权采样(WCARS)结合迭代式连续投影算法(ISPA)的光谱特征波长选择方法, 该方法在有效解决模型过拟合问题的同时, 能够降低高维数据计算的复杂度。 将火焰光谱数据沿波长方向进行窗口划分后, 使用CARS进行计算选出特征窗口波段, 再将迭代式选择与传统连续投影算法相结合, 通过重复迭代精选出特征波长, 在此基础上使用支持向量机回归(SVR)建立炼钢终点碳含量预测模型。 实验采集363组炼钢后期的炉口火焰光谱数据作为样本, 并对其进行Savitzky-Golay平滑预处理。 使用WCARS-ISPA算法从全光谱数据中选出10个特征波长作为SVR模型的输入, 碳含量为模型输出, Kennard-stone算法对训练集和测试集进行划分, 选择碳含量的平均预测误差、 预测误差在±2%以内的命中率以及运行30次的平均时间作为模型评价指标。 实验结果显示, 模型的平均碳含量预测误差为1.413 2%, 命中率高达90.63%, 运行时间为0.019 679 s。 与使用全光谱和WCARS-ISPA, CARS-SPA, WCARS和SPA四种不同特征选择方法选出的特征波长建模得到的结果进行对比, 基于WCARS-ISPA方法选出的特征波长建立的终点碳含量预测模型误差最小、 命中率最高。 提出一种新的炉口火焰光谱特征波长提取方法, 使用窗口竞争性自适应重加权采样结合迭代式连续投影算法选取特征波长, 并在此基础上建立转炉炼钢终点碳含量预测模型, 实验结果表明, 该方法能够有效提取火焰光谱特征, 所建模型能够对转炉炼钢终点进行准确预测, 满足工业生产的实时控制要求, 为实际生产提供可靠帮助。

Real-time precise control of the BOF steelmaking end-point can effectively improve the quality of steel output. The flame spectra change obviously in different stages of steelmaking. It can be used to control the end-point of steelmaking effectively with the machine learning method. Due to a large amount of spectral data and the lack of reliability and real-time performance of the existing methods for spectral feature extraction, a characteristic spectral wavelength selection method based on window competitive adaptive reweighted sampling (WCARS) combined with iterative successive projection algorithm (ISPA) was proposed in this paper. This method can effectively solve the problem of over-fitting and reduce the complexity of high-dimensional data calculation. After dividing the spectral data along the wavelength direction in the window, CARS was used to select the feature window band. The iterative selection was combined with a traditional successive projection algorithm, and the characteristic wavelengths were selected through repeated iteration. On this basis, support vector machine regression (SVR) was used to establish the carbon content prediction model of steelmaking end-point. 363 sets of spectral data of the later stage of steelmaking were collected as an experimental sample and preprocessed by Savitzky-Golay smoothing. The input of the SVR model was 10 characteristic wavelength data selected by WCARS-ISPA, and the output was carbon content. The training set and test set were divided by the Kennard-Stone algorithm. The average prediction error of carbon content, the hit ratio of prediction error within ±2% and the average running time of 30 times were selected as the evaluation indexes. The results indicated that the average prediction error is 1.413 2%, the hit ratio is 90.63%, and the running time is 0.019 679 s. Compared with the model of full spectra and characteristic wavelengths selected by four different feature selection methods of WCARs-ISPA, CARS-SPA, WCARS and SPA, the WCARS-ISPA model has the lowest error and the highest hit ratio. In this paper, a new flame spectral characteristic wavelength extraction method was proposed. Window competitive adaptive reweighted sampling was combined with an iterative successive projection algorithm to select the wavelength, and a prediction model of end-point carbon content is established on this basis. The experimental results showed that this method could effectively extract the spectral characteristics of flame. This model can accurately predict the endpoint of BOF steelmaking and meet the requirements of real-time control of industrial production.