基于CNN-LSTM的二阶段辊式淬火过程板形预测方法  

Plate shape prediction method in two stage based on CNN-LSTM for quenching process with roller-hearth quenching machine

作　　者：刘艾 张廷虎 王忠亮 Liu Ai;Zhang Tinghu;Wang Zhongliang(Heqing Beiya Mining Co.,Ltd.,Dali Yunnan 671000,China)

机构地区：[1]鹤庆北衙矿业有限公司,云南大理671000

出　　处：《金属热处理》2024年第10期295-300,共6页Heat Treatment of Metals

摘　　要：钢板的板形是淬火过程的关键质量指标,针对钢板在淬火过程中板形预测的问题,提出一种基于卷积神经网络-长短时记忆网络(CNN-LSTM)的二阶段辊式淬火过程板形预测方法。该方法分为两个阶段,首先,利用CNN提取板形特征,捕捉板形的空间信息。其次,以淬火过程参数、历史板形特征为输入,采用LSTM建立板形预测模型。最后将两个阶段串联,使预测模型能够同时考虑板形的空间信息和时间信息。基于实际生产数据进行试验,其结果表明,预测误差的均方值从0.0471降低到了0.0264,即预测误差降低了43.9%,达到了提高板形预测精度的目标。Shape of the steel plate is a key quality indicator during the quenching process.In order to solve the problem of plate shape prediction of steel plates during the quenching process,a two-stage shape prediction method for steel plates in roller-hearth machine quenching process based on convolutional neural network and long short-term memory network(CNN-LSTM)was proposed.This method was divided into two stages.Firstly,the CNN was used to extract the plate shape features and capture the spatial information of the plate shape.Secondly,using quenching process parameters and historical plate shape characteristics as inputs,a plate shape prediction model was established through LSTM.Finally,by concatenating these two stages,both spatial and temporal information of the plate shape could be considered simultaneously.Based on the experiments with actual production data,the results show that the proposed method reduces the root mean squared error of the prediction is reduced from 0.0471 to 0.0264,which represents a 43.9%reduction in prediction error,achieving the goal of improving the plate shape prediction accuracy.

关 键 词：辊式淬火 板形预测 卷积神经网络 长短时记忆网络 

分 类 号：TG162[金属学及工艺—热处理]