摘要
为了更好地应用BP神经网络对连铸板坯质量进行在线诊断,基于连铸生产特点,利用采集的过程数据建立了符合生产实际的均一化函数。通过分析BP神经网络中各参数对网络性能及诊断准确率的影响,对BP神经网络的结构及学习算法进行修正,使该网络有选择和有区分地学习铸坯质量知识。结合某钢厂连铸现场数据,以黏结为例,建立了6种网络模型,对各模型算法进行了比较测试。结果表明:采用自定义函数均一化样本或采用提出的差异性算法训练神经网络,均可明显提高诊断准确率;采用选择性算法可确保诊断准确率不变的同时,提高学习速度;修正的算法更能很好地符合连铸生产实际。
In order to apply BP neural network in the continuous casting slab on-line diagnostics even better, the homogeneous functions which met the requirements of manufacture were established upon characteristics of the continuous casting. The effect of the parameters of BP neural network on performance and the diagnostic accuracy rate was analyzed. By amending the structure and algorithm of BP neural network, knowledge of the slab quality was learned selectively and discriminatively. Finally, BP neural network was applied to the sticking prediction in the continuous casting processes, and six models were established and compared with the historical data collected from a steel mill. The result show: Sample homogenized by custom functions and difference training algorithm can significantly improve the diagnostic accuracy rate; Selective training algorithm can speed up learning process, but also ensure the same diagnostic accuracy rate. The improved algorithm is in keeping with the real conditions of the continuous casting processes verified by the research result.
出处
《钢铁研究学报》
CAS
CSCD
北大核心
2013年第7期58-62,共5页
Journal of Iron and Steel Research
基金
"十二五"科技支撑计划资助项目(2012BAE03B02)