融合深度学习与模型优化的连铸坯壳漏钢预报  

Steel leakage prediction of continuous casting billet shell by combining deep learning and model optimization

作　　者：成彬 黄诺金 雷华[2] 左水利[2] CHENG Bin;HUANG Nuojin;LEI Hua;ZUO Shuili(School of Mechanical and Electrical Engineering,Xi'an University of Architecture and Technology,Xi'an 710055,Shaanxi,China;China Nation Heavy Machinery Research Institute Co.,Ltd.,Xi'an 710018,Shaanxi,China)

机构地区：[1]西安建筑科技大学机电工程学院,陕西西安710055 [2]中国重型机械研究院股份公司,陕西西安710018

出　　处：《钢铁研究学报》2024年第7期885-895,共11页Journal of Iron and Steel Research

基　　金：陕西省自然科学基础研究计划资助项目(2021JM-360)。

摘　　要：针对传统漏钢预报模型时序建模能力不足,且大多没有对热电偶温度数据进行处理,直接作为模型输入导致计算量大的问题,通过分析黏结漏钢的形成机制和时序特征,基于长短期记忆网络(LSTM),添加能聚焦关键信息的注意力机制(Attention Mechanism),提出了一种基于深度学习的Attention-LSTM漏钢预报模型。首先,对热电偶温度数据进行数据增广、帧间差分、阈值分割等处理,提取黏结漏钢的共性特征作为模型输入;其次,通过遍历选参搭建模型进行训练,以均方误差作为模型损失函数,寻找最优预报模型;最后,为提高模型预测精度,在Dense输出层前添加注意力机制模块完成模型优化。应用连铸现场数据对模型进行测试,测试结果表明,提出的Attention-LSTM漏钢预报模型以97.3%的预报率和100%的报出率,验证了模型的可行性和有效性。Aiming to address the issue that traditional steel leakage forecasting models lack sufficient time-sequence modeling capability and most of them do not process the thermocouple temperature data,which is directly used as model input,resulting in high computational complexity.By analyzing the formation mechanism and timesequence characteristics of adhesive steel leakage and leveraging the Long Short-Term Memory network(LSTM),adding the Attention Mechanism that can focus on the key information,a deep learning-based Attention-LSTM forecasting model was proposed for predicting steel leakage.Firstly,the thermocouple temperature data is processed through data augmentation,inter-frame differencing,threshold segmentation,etc.,and the common features of bonded steel leakage are extracted as model inputs.Secondly,the model is constructed by traversing the selected parameters for training,using the mean square error as the model's loss function to find the optimal forecasting model.Lastly,to enhance the model's prediction accuracy,the Attention Mechanism module is added in front of the Dense output layer to complete the model optimization.The model is tested by applying the continuous casting field data,and the test results show that the proposed Attention-LSTM steel leakage forecasting model verifies the feasibility and effectiveness of the model with a 97.3% forecasting rate and 100% reporting rate.

关 键 词：连铸 漏钢预报 LSTM网络 注意力机制 特征提取 模型优化 

分 类 号：TF777.1[冶金工程—钢铁冶金]