基于径向基神经网络和遗传算法建立转炉提钒终点预报模型的研究
|
摘要
钒在钢铁冶金、电子及国防工业等方面有着重要的应用价值，是一种具有战略价值的金属材料。我国目前的转炉提钒为人工操作模式，由操作人员根据经验和感觉进行操作，自动化水平低，存在着钒渣质量和半钢质量不稳定的问题；在国外，俄罗斯等一些国家已经在使用静态模型对提钒过程进行控制，取得了较好的效果，但是这些模型大都是根据复杂的物理化学规律研制的机理模型，要求有非常稳定的生产流程和工艺条件，对于铁水成分、生产设备等变化的适应性差，不但移植困难，而且模型价格异常昂贵。研制具有高性价比的转炉提钒模型是我国钒生产企业的迫切需要，对如提钒这样的复杂冶金工业过程建模也是目前国内外的研究热点之一。随着计算智能研究的兴起，一些新的建模方法如RBF神经网络和遗传算法等为复杂的冶金工业过程建模提供了新的思路和方法。本文使用RBF神经网络和遗传算法建立了转炉提钒终点预报模型，并取得了较好的效果，为计算智能方法在复杂工业过程建模中的应用作出了有益的尝试。
本文首先根据冶金学原理，找出影响转炉提钒终点状态的主要因素，并从某大型钢厂的原始数据中提取合符要求的数据进行建模。从降低模型的复杂度的角度，本文将整个终点预报模型分成三个独立的模型，即终点温度模型、终点碳模型、终点钒模型。在深入研究神经网络理论的基础上，本文使用RBF神经网络对终点的三个指标分别建模。
由于RBF神经网络本身对其神经元个数、Spread等参数比较敏感，所以这些参数的选取对RBF神经网络的逼近效果影响很大，靠经验方法和手工测试方法选取的神经元个数、Spread等网络参数得到的网络模型对测试集误差较大，也就是说模型的泛化能力不好。本文通过对泛化理论的研究，分析了影响泛化能力的主要因素，提出了使用遗传算法对RBF神经网络的神经元个数、Spread等参数进行优化，求取具有较好泛化能力的神经网络的方法，试验证明通过这种方法求得的一组网络模型具有较好泛化能力，并成功的对测试集进行了准确的预报。最后得到的转炉提钒终点状态预报模型对终点温度的预报命中率达到86.4%，对终点碳的预报命中率达到83.7%，终点钒的预报命中率达到59.4%，碳温同时命中率达到76.3%，三者同时命中率达到43.2%。
Vanadium is important and valuable in steel-making, electronic production and national defence industry etc. Now, the operation pattern of refining vanadium in our nation is based on human experiences, and the automation degree is still at a low level. Operation mode based on human experiences is one of the main reasons make the quality of semi-steel and vanadium product unstable. In developed contury such as Russian, static model was used to control the process of vanadium refining. But most of these models are based on complex physical and chemical reactions. It's difficult to transplant these models from equipment to another, and these models are very expensive. Cheep and good models were strong desired. Building model for a complex metallurgic process such as vanadium refining is one of the main focal point in Control Theory Researchs. This paper discusses the construction procedure of Converting Furnace Endpoint Prediction Model in Refining vanadium based on RBF neural networks and Genetic Algorithms.
At first this paper find out the main factor that affects endpoint status according to metallurgy, then build the Endpoint Prediction Model. From the aspect of diminishing the complex degree of the model, this paper separates the Endpoint Prediction Model to three detached models, Endpoint Temperature Model, Endpoint Carbon Model and Endpoint Vanadium model. Based on the penetrating research of neural networks, the author use RBF neural networks to construct models of the three endpoint criterions. Because RBF is very sensitive to some parameters such as the number of the neuron and the spread, choosing those parameters contributes a lot to the approach effect of the RBF neural networks. The valid data collecting from the scene is not enough and inevitably has some noise because of the restriction of measurement methods, so when use the network model obtained by choosing parameters such as the number of neuron and spread based on experiences and handmade test, there will be a large deviation, that is, the generalization ability of the model is not good. Based on the research of the Generalization Theory, this paper analyses main factors effecting the generalization ability, and present a method that uses the genetic algorithm optimize parameters such as the number of neuron and spread and get a method with a better generalization ability. It is proved by experiments that the group of network models obtained by this method have better generalization abilities, and those experiments forecast the independent check set accurately and successfully. The hit
ratio of the Converting Furnace Endpoint Prediction Model in Refining Vanadium is 86.4% to the endpoint temperature, 83.7% to the endpoint carbon, 59.4% to the endpoint vanadium, 76.3% to both the temperature and the carbon, and 43.2% to all of the three.
本文首先根据冶金学原理，找出影响转炉提钒终点状态的主要因素，并从某大型钢厂的原始数据中提取合符要求的数据进行建模。从降低模型的复杂度的角度，本文将整个终点预报模型分成三个独立的模型，即终点温度模型、终点碳模型、终点钒模型。在深入研究神经网络理论的基础上，本文使用RBF神经网络对终点的三个指标分别建模。
由于RBF神经网络本身对其神经元个数、Spread等参数比较敏感，所以这些参数的选取对RBF神经网络的逼近效果影响很大，靠经验方法和手工测试方法选取的神经元个数、Spread等网络参数得到的网络模型对测试集误差较大，也就是说模型的泛化能力不好。本文通过对泛化理论的研究，分析了影响泛化能力的主要因素，提出了使用遗传算法对RBF神经网络的神经元个数、Spread等参数进行优化，求取具有较好泛化能力的神经网络的方法，试验证明通过这种方法求得的一组网络模型具有较好泛化能力，并成功的对测试集进行了准确的预报。最后得到的转炉提钒终点状态预报模型对终点温度的预报命中率达到86.4%，对终点碳的预报命中率达到83.7%，终点钒的预报命中率达到59.4%，碳温同时命中率达到76.3%，三者同时命中率达到43.2%。
Vanadium is important and valuable in steel-making, electronic production and national defence industry etc. Now, the operation pattern of refining vanadium in our nation is based on human experiences, and the automation degree is still at a low level. Operation mode based on human experiences is one of the main reasons make the quality of semi-steel and vanadium product unstable. In developed contury such as Russian, static model was used to control the process of vanadium refining. But most of these models are based on complex physical and chemical reactions. It's difficult to transplant these models from equipment to another, and these models are very expensive. Cheep and good models were strong desired. Building model for a complex metallurgic process such as vanadium refining is one of the main focal point in Control Theory Researchs. This paper discusses the construction procedure of Converting Furnace Endpoint Prediction Model in Refining vanadium based on RBF neural networks and Genetic Algorithms.
At first this paper find out the main factor that affects endpoint status according to metallurgy, then build the Endpoint Prediction Model. From the aspect of diminishing the complex degree of the model, this paper separates the Endpoint Prediction Model to three detached models, Endpoint Temperature Model, Endpoint Carbon Model and Endpoint Vanadium model. Based on the penetrating research of neural networks, the author use RBF neural networks to construct models of the three endpoint criterions. Because RBF is very sensitive to some parameters such as the number of the neuron and the spread, choosing those parameters contributes a lot to the approach effect of the RBF neural networks. The valid data collecting from the scene is not enough and inevitably has some noise because of the restriction of measurement methods, so when use the network model obtained by choosing parameters such as the number of neuron and spread based on experiences and handmade test, there will be a large deviation, that is, the generalization ability of the model is not good. Based on the research of the Generalization Theory, this paper analyses main factors effecting the generalization ability, and present a method that uses the genetic algorithm optimize parameters such as the number of neuron and spread and get a method with a better generalization ability. It is proved by experiments that the group of network models obtained by this method have better generalization abilities, and those experiments forecast the independent check set accurately and successfully. The hit
ratio of the Converting Furnace Endpoint Prediction Model in Refining Vanadium is 86.4% to the endpoint temperature, 83.7% to the endpoint carbon, 59.4% to the endpoint vanadium, 76.3% to both the temperature and the carbon, and 43.2% to all of the three.
引文
[1] 刘浏 转炉炼钢的技术进步 炼钢2000：1-6
[2] 陶钧，谢书明，柴天佑 转炉炼钢控制模型的研究与展望 钢铁1999 34（8）
[3] Iwamura K.New Endpoint Control System With Auto-parameter-tuning in BOF. Steelmaking
Conference Proceedings.1995.715～719.
[4] Yamane H. Development of Expert System and Mn Online Sensor for BOF Process. Steelmaking
Conference Proceedings.1991.453～458.
[5] 樊俊飞，李永如，姚海石.宝钢转炉吹炼控制模拟在线专家系统.宝钢技术，1996，(4)：50～
58.
[6] 丁容，刘浏 转炉炼钢过程人工智能静态控制模型 钢铁1997 32（1） 22-26
[7] 黄云，齐振亚，董履仁 利用人工神经网络系统预报钢水温度 炼钢2001 Vol.17 No.5 43-46
[8] 陶钧，谢书明，柴天佑 基于遗传算法和径向基函数神经网络的转炉炼钢模型 系统仿真学
报2000 Vol.12 No.3 241-244
[9] 黄道鑫 提钒炼钢 冶金工业出版社
[10] 潘树范 国内外氧气顶吹转炉提钒现状及对攀钢转炉提钒有关问题的探讨 钢铁钒钛1995
Vol.16 No.1 6-16
[11] 王永骥 涂健 神经元网络控制 机械工业出版社，1998.2
[12] 胡守仁 神经网络应用技术 电子工业出版社 1993.12
[13] 李聪 张勇 高智 一种新的聚类算法 模式识别与人工智能 1999（6）：205-208
[14] Chen. S, C.F.N . Cowan, P. M. Grant , "Orthogonal Least Squares Learning Algorithm for Radial
Basis Function Networks" IEEE Transactions on Neural networks. 1991. 3.Vol.2, No.2
[15] 高大启，有教师的线性基本函数前向三层神经网络结构研究，计算机学报，1998. Vol.21
No.1：80-83
[16]李杰 韩正之 一种估计人工神经网络泛化误差的新方法 控制理论与应用2001. Vol.18 No.2
257-259
[17] Moody J E. The Effective Number of Parameters: An Analysis of Generalization and
Regularization in Nonlinear Learning System. NIPS 4, 1992, San Mateo, CA:847-854
[18] Cohn D. Improving generalization with active learning. Machine Learning. 1994(15) 201-221
[19] Vladimir N. Cherkassky, Vapnik-Chervonenkis(VC) Learning Theory and Its Applications, IEEE
Trasactions on Neural Networks, Vol.10, No.5, 1999
[20] Wang C, Venkatesh S S. Optimal Stopping and Effective Machine Complexity in Learning,
NIPS 6, 1994 , San Mateo, CA:303-310
[21] Wang C, Venkatesh S S. Temporal Dynamics of Generalization in Neural Networks. NIPS 7,
1995, Cambridge, MA:263-270
[22] 韩正之 李杰 神经网络的学习误差函数及泛化能力 控制与决策2000.1 Vol.15 No.1 95-97
[23] Sjoberg J, Ljung L. Overtraining, regularization and searching for a minimum, with application to neural networks. International Journal of Control, 1995, 62(6):1391~1407
[24] 潘正君，康立山，陈毓屏 著，演化计算，北京：清华大学出版社，1998.5
[25] 吴斌 吴坚 涂序彦，快速遗传算法研究，电子科技大学学报，1999，Vol.28 No.1：49-53
[26] 陈铁男，陈广义，模拟生物种族形成的进化算法与多峰函数优化，控制与决策，1999（2）：
5-9
[27] 董聪，关仪，遗传算法，大自然探索，1998，16（1）：37~41
[28] 董聪 郭晓华 袁曾任，基于广义遗传算法的全局优化方法，计算机科学，Vol.26 No.6 ,1999：
7~10
[29] Baker J E. Adaptive Selection Methods for Genetic Algorithms. in Proc. of the 1st Int'l. Conf. on
Genetic Algorithms , 110-111, 1985
[30] 张静 基于神经网络和遗传算法的板胚缺陷预报和参数优化系统研究 重庆大学硕士学位论
文2001 ：23-24
[31] Simon Haykin. Neural Networks: A Comprehensive Foundation (Second Edition). Prentice Hall,
清华大学出版社 2001
[32] 陈宝林 最优化理论与算法 清华大学出版社 1989
[2] 陶钧，谢书明，柴天佑 转炉炼钢控制模型的研究与展望 钢铁1999 34（8）
[3] Iwamura K.New Endpoint Control System With Auto-parameter-tuning in BOF. Steelmaking
Conference Proceedings.1995.715～719.
[4] Yamane H. Development of Expert System and Mn Online Sensor for BOF Process. Steelmaking
Conference Proceedings.1991.453～458.
[5] 樊俊飞，李永如，姚海石.宝钢转炉吹炼控制模拟在线专家系统.宝钢技术，1996，(4)：50～
58.
[6] 丁容，刘浏 转炉炼钢过程人工智能静态控制模型 钢铁1997 32（1） 22-26
[7] 黄云，齐振亚，董履仁 利用人工神经网络系统预报钢水温度 炼钢2001 Vol.17 No.5 43-46
[8] 陶钧，谢书明，柴天佑 基于遗传算法和径向基函数神经网络的转炉炼钢模型 系统仿真学
报2000 Vol.12 No.3 241-244
[9] 黄道鑫 提钒炼钢 冶金工业出版社
[10] 潘树范 国内外氧气顶吹转炉提钒现状及对攀钢转炉提钒有关问题的探讨 钢铁钒钛1995
Vol.16 No.1 6-16
[11] 王永骥 涂健 神经元网络控制 机械工业出版社，1998.2
[12] 胡守仁 神经网络应用技术 电子工业出版社 1993.12
[13] 李聪 张勇 高智 一种新的聚类算法 模式识别与人工智能 1999（6）：205-208
[14] Chen. S, C.F.N . Cowan, P. M. Grant , "Orthogonal Least Squares Learning Algorithm for Radial
Basis Function Networks" IEEE Transactions on Neural networks. 1991. 3.Vol.2, No.2
[15] 高大启，有教师的线性基本函数前向三层神经网络结构研究，计算机学报，1998. Vol.21
No.1：80-83
[16]李杰 韩正之 一种估计人工神经网络泛化误差的新方法 控制理论与应用2001. Vol.18 No.2
257-259
[17] Moody J E. The Effective Number of Parameters: An Analysis of Generalization and
Regularization in Nonlinear Learning System. NIPS 4, 1992, San Mateo, CA:847-854
[18] Cohn D. Improving generalization with active learning. Machine Learning. 1994(15) 201-221
[19] Vladimir N. Cherkassky, Vapnik-Chervonenkis(VC) Learning Theory and Its Applications, IEEE
Trasactions on Neural Networks, Vol.10, No.5, 1999
[20] Wang C, Venkatesh S S. Optimal Stopping and Effective Machine Complexity in Learning,
NIPS 6, 1994 , San Mateo, CA:303-310
[21] Wang C, Venkatesh S S. Temporal Dynamics of Generalization in Neural Networks. NIPS 7,
1995, Cambridge, MA:263-270
[22] 韩正之 李杰 神经网络的学习误差函数及泛化能力 控制与决策2000.1 Vol.15 No.1 95-97
[23] Sjoberg J, Ljung L. Overtraining, regularization and searching for a minimum, with application to neural networks. International Journal of Control, 1995, 62(6):1391~1407
[24] 潘正君，康立山，陈毓屏 著，演化计算，北京：清华大学出版社，1998.5
[25] 吴斌 吴坚 涂序彦，快速遗传算法研究，电子科技大学学报，1999，Vol.28 No.1：49-53
[26] 陈铁男，陈广义，模拟生物种族形成的进化算法与多峰函数优化，控制与决策，1999（2）：
5-9
[27] 董聪，关仪，遗传算法，大自然探索，1998，16（1）：37~41
[28] 董聪 郭晓华 袁曾任，基于广义遗传算法的全局优化方法，计算机科学，Vol.26 No.6 ,1999：
7~10
[29] Baker J E. Adaptive Selection Methods for Genetic Algorithms. in Proc. of the 1st Int'l. Conf. on
Genetic Algorithms , 110-111, 1985
[30] 张静 基于神经网络和遗传算法的板胚缺陷预报和参数优化系统研究 重庆大学硕士学位论
文2001 ：23-24
[31] Simon Haykin. Neural Networks: A Comprehensive Foundation (Second Edition). Prentice Hall,
清华大学出版社 2001
[32] 陈宝林 最优化理论与算法 清华大学出版社 1989