摘要
在带钢冷轧过程中,轧制力预报精度直接决定板带材的轧制精度以及产品质量。传统的基于单隐层的神经网络建模方法结构简单,对复杂函数的表达能力与泛化能力都受到一定制约;轧制现场环境复杂,数据测量存在噪声干扰,都会直接影响预报精度。针对这些问题,提出一种基于非监督学习的改进深度信念网络预测模型。深层网络的构建以及去噪机制的引入可提高系统对输入数据特征学习的能力,同时采用改进的对比散度算法对网络进行训练,提高网络训练速度。最后,利用某钢厂1200 mm轧机组的实测数据对模型进行检验,对比分析3种不同模型,结果表明该模型对轧制力预测的平均相对误差控制在4.5%以内,建模所需时间相比于栈式自编码网络减少26%。
In the cold rolling process of strip steel,the accuracy of rolling force prediction directly determines the rolling precision and product quality of the strip.The traditional single-hidden layer-based neural network modeling method is simple in structure,and the expression ability and generalization ability of complex functions are restricted.The rolling site environment is complex,and data measurement has noise interference,which will directly affect the forecasting accuracy.Regarding the issue above,an improved deep belief network prediction model based on unsupervised learning is proposed.The construction of denoising-restricted Boltzmann machines and deep networks can improve the system s ability to learn the characteristics of input data,while training the deep network with improved contrast divergence algorithm.Finally,the model is tested by using the measured data of a steel mill s 1200 mm rolling mill,and three different models are compared and analyzed.The results show that the average relative error of the rolling force prediction of the model is controlled within 4.5%,and the time required for modeling is reduced by 26%compared to the self-encoding network.
作者
魏立新
王恒
孙浩
呼子宇
WEI Li-xin;WANG Heng;SUN Hao;HU Zi-yu(Intelligent Control System and Intelligent Equipment Engineering Research Center of Ministry of Education,Yanshan University,Qinhuangdao,Hebei 066004,China;Key Lab of Industrial Computer Control Engineering of Hebei Province,Yanshan University,Qinhuangdao,Hebei 066004,China)
出处
《计量学报》
CSCD
北大核心
2021年第7期906-912,共7页
Acta Metrologica Sinica
基金
国家自然科学基金(61803327)
河北省自然科学基金青年项目(E2018203162)。
关键词
计量学
轧制力预报
深度信念网络
去噪机制
metrology
rolling force prediction
deep belief network
denoising mechanism
