摘要
高炉炼铁过程中难以实时测量的硅含量指标,快速预测方法对指导流程后续操作和保障最终质量具有重要意义。基于深度学习算法设计了一种新型高炉炼铁硅含量软测量框架;考虑到高温高压的生产环境对传感器的影响,采集数据多包含异常值和噪声,设计了基于单类支持向量机(One-Class SVM)的异常数据过滤模块和基于傅里叶变换(Fourier Transform,FT)和高斯平滑(Gaussian Smoothing)的混合特征降噪模块;考虑到工业数据的强非线性和时序依赖性,设计了基于改进Transformer-LSTM模型的回归模块;最后,使用高炉炼铁的实际工业过程数据集,通过多模型对比实验和模型功能的消融实验,异常过滤模块和降噪模块使模型在均方误差上分别降低了1.1%和16.07%;Transformer-LSTM模型在预测误差上对比常用的LSTM、GRU和TCN在均方误差上分别降低了0.63%、4.7%和2.4%;并有效减少了循环神经网络堆叠层数。
Silicon content indicators that are difficult to measure in real-time during blast furnace ironmaking process.It is of great significance to study the rapid prediction method to guide the follow-up operation of the process and guarantee the final quality.Therefore,a new soft measurement framework for silicon content in blast furnace ironmaking is designed based on deep learning algorithm.Considering the impact of high temperature and pressure production environment on the sensor,the collected data mostly contains outliers and noise.The abnormal data filtering module based on One-class SVM and the mixed feature noise reduction module based on Fourier Transform and Gaussian Smoothing are designed.Considering the strong nonlinearity and temporal dependence of industrial data,a regression module based on the improved Transformer-LSTM model is designed.Finally,using the actual industrial process data set of blast furnace ironmaking,the mean square error of the model is reduced by 1.1%and 16.07%by the anomaly filtering module and the noise reduction module through the multi-model comparison experiment and the ablation experiment of model functions.The prediction error of Transformer LSTM model is 0.63%,4.7%and 2.4%lower than that of LSTM,GRU and TCN,respectively.And the number of stacking layers of the recurrent neural network is effectively reduced.
作者
钱子豪
王磊
靳晟
Qian Zihao;Wang Lei;Jin Sheng(College of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi 830052,China;Xinjiang Electronics Research Institute Co.,Ltd.,Urumqi 830013,China)
出处
《机电工程技术》
2025年第19期35-40,114,共7页
Mechanical & Electrical Engineering Technology
基金
中央引导地方科技发展资金项目(ZYYD2024JD28)。
关键词
软测量
高炉炼铁
质量预测
特征降噪
变换
LSTM
soft measurement
blast-furnace ironmaking
quality prediction
feature noise reduction
transformer
LSTM
