基于案例推理增强学习的磨矿过程设定值优化
|
摘要
磨矿粒度和循环负荷是磨矿过程产品质量与生产效率的关键运行指标,相对于底层控制偏差,回路设定值对其影响要严重的多.然而,磨矿过程受矿石成分与性质、设备状态等变化因素影响,运行工况动态时变,难以建立模型,因此难以通过传统的模型方法优化回路设定值.本文将增强学习与案例推理相结合,提出一种数据驱动的磨矿过程设定值优化方法.首先根据当前运行工况,采用基于Prey-Predator优化的案例推理方法,决策出可行的基于Elman神经网络的Q函数网络模型;然后利用实际运行数据,在增强学习的框架下,根据Q函数网络模型优化回路设定值.在基于METSIM的磨矿流程模拟系统上进行实验研究,结果表明所提方法可根据工况变化在线优化回路设定值,实现磨矿运行指标的优化控制.
In grinding processes, particle size and circulating load are two key operation indexes for product quality and production efficiency. With respect to the economic performance, the basic loop controller performance is most probably not as important as the right selection of the loop set-points. The industrial grinding processes, however, are affected by the factors such as composition and properties of ore, the equipment status and so on. When the large fluctuation of the factors occurs, the operation will be time-varying, thereby making the process modeling very difficult. Therefore, it is hard to employ the traditional model-based methods to optimize the loop set-points. In this paper, a data-driven optimalsetting control method is proposed by using case-based reasoning(CBR) and reinforcement learning(RL) technologies.The method first employs a Prey-Predator optimization-based CBR method to determine a feasible Elman neural networkbased Q function model in accordance with current operation condition. Then, under the RL framework, the Q function model is adopted to optimize the loop set-points according to the operation data. Experiments studies are carried out in a METSIM-based grinding simulation system. Results show that the proposed method can realize the optimization control of the grinding operation indexes by optimizing the loop set-points online according to the varied operation conditions.
In grinding processes, particle size and circulating load are two key operation indexes for product quality and production efficiency. With respect to the economic performance, the basic loop controller performance is most probably not as important as the right selection of the loop set-points. The industrial grinding processes, however, are affected by the factors such as composition and properties of ore, the equipment status and so on. When the large fluctuation of the factors occurs, the operation will be time-varying, thereby making the process modeling very difficult. Therefore, it is hard to employ the traditional model-based methods to optimize the loop set-points. In this paper, a data-driven optimalsetting control method is proposed by using case-based reasoning(CBR) and reinforcement learning(RL) technologies.The method first employs a Prey-Predator optimization-based CBR method to determine a feasible Elman neural networkbased Q function model in accordance with current operation condition. Then, under the RL framework, the Q function model is adopted to optimize the loop set-points according to the operation data. Experiments studies are carried out in a METSIM-based grinding simulation system. Results show that the proposed method can realize the optimization control of the grinding operation indexes by optimizing the loop set-points online according to the varied operation conditions.
引文
[1]ZHOU Ping,CHAI Tianyou.Intelligent operation feedback control of typical hematite grinding process.Control Theory&Applications,2014,31(10):1352-1359.(周平,柴天佑.典型赤铁矿磨矿过程智能运行反馈控制.控制理论与应用,2014,31(10):1352-1359.)
[2]WEI D,CRAIG I K.Grinding mill circuits-a survey of control and economic concerns.International Journal of Mineral Processing,2009,90(1):56-66.
[3]STEYN C W,SANDROCK C.Benefits of optimisation and model predictive control on a fully autogenous mill with variable speed.Minerals Engineering,2013,53(6):113-123.
[4]NIU D,CHEN X,YANG J,et al.Composite control for raymond mill based on model predictive control and disturbance observer.Advances in Mechanical Engineering,2016,8(3):1-10.
[5]MINCHALA L I,ZHANG Y,GARZA-CASTANON L E.Predictive control of a closed grinding circuit system in cement industry.IEEETransactions on Industrial Electronics,2018,65(5):4070-4079.
[6]CRAIG I K.Grinding mill modeling and control:past,present and future.Control Conference.Najing,china:IEEE,2012:16-21.
[7]ZHOU P,CHAI T Y.Grinding circuit control:a hierarchical approach using extended 2-DOF decoupling and model approximation.Powder Technology,2011,213(1):14-26.
[8]CHEN X S,YANG J,LI S H.Disturbance observer based multivariable control of ball mill grinding circuits.Journal of Process Control,2009,19(7):1205-213.
[9]ZHANO Dayong,CHAI Tianyou.Fuzzy switching control for sump level interval and hydrocyclone pressure in regrindingprocess.Acta Automatica Sinica,2013,39(5):556-564.(赵大勇,柴天佑.再磨过程泵池液位区间与给矿压力模糊切换控制.自动化学报,2013,39(5):556-564.)
[10]REMES A,AALTONEN J,KOIVOV H.Grinding circuit modeling and simulation of particle size control at Silinjarci concentrator.International Journal Miner Process,2010,96(1/2/3/4):70-78.
[11]ZHANG Xiaolin.Research and application of grinding process expert system.Changchun:Jilin University,2010.(张孝临.磨矿过程专家系统研究与应用.长春:吉林大学,2010.)
[12]CHEN X S,ZHAI J Y,LL Q,et al.Fuzzy logic based on-line efficiency optimization control of a ball mill grinding circuit.Proceedings of the Fourth International Conference on Fuzzy Systems and Knowledge Discovery.Haikou,china:IEEE,2007,2:575-580.
[13]ZHOU P,CHAI T,WANG H.Intelligent optimal-setting control for grinding circuits of mineral processing process.IEEE Transactions on Automation Science and Engineering,2009,6(4):730-743.
[14]ZHOU P,CHAI T,SUN J.Intelligence-based supervisory control for optimal operation of a DCS-controlled grinding system.IEEE Transactions on Control Systems Technology,2013,21(1):162-175.
[15]SUTTON R S,BARTO A G.Reinforcement Learning:An Introduction.Cambridge:MIT Press,1998.
[16]KIUMARSI B,LEWIS F L,JIANG Z P.H infinite control of linear discrete-time systems:off-policy reinforcement learning.Automatica,2017,78(4):144-152.
[17]WANG Xuesong,TIAN Xilan,CHENG Yuhu,et al.Q learning based on Cooperative least squares support vector machines.Journal of Automation,2009,35(2):214-219.(王雪松,田西兰,程玉虎,等.基于协同最小二乘支持向量机的Q学习.自动化学报,2009,35(2):214-219.)
[18]VALENZUELA J,NAJIM K,VILLAR R D,et al.Learning control of an autogenous grinding circuit.International Journal of Mineral Processing,1993,40(1/2):45-56.
[19]CONRADIE A V E,ALDRICH C.Neurocontrol of a ball mill grinding circuit using evolutionary reinforcement learning.Minerals Engineering,2001,14(10):1277-1294.
[20]DAI Wei,CHAI Tianyou.Data driven complex grinding process operation optimization control method.Journal of Automation,2014,40(9):2005-2014.(代伟,柴天佑.数据驱动的复杂磨矿过程运行优化控制方法.自动化学报,2014,40(9):2005-2014.)
[21]JAEGER H.The“echo state”approach to analysing and training recurrent neural networks.Bonn,Germany:German National Research Center for Information Technology GMD Technical Report,2001,148(34):13.
[22]TILAHUM S L,ONG H C.Prey-predator algorithm:a new metaheuristic algorithm for optimization problems.International Journal of Information Technology and Decision Making,2015,14(6):1331-1352.
[23]TILAHUM S,SATHASIVAM S,ONG H C,et al.Prey-predator algrorithm as a new optimization technique using in radia basis function neural networks.Research Journal of Applied Sciences,2013,7(8):383-387.
[2]WEI D,CRAIG I K.Grinding mill circuits-a survey of control and economic concerns.International Journal of Mineral Processing,2009,90(1):56-66.
[3]STEYN C W,SANDROCK C.Benefits of optimisation and model predictive control on a fully autogenous mill with variable speed.Minerals Engineering,2013,53(6):113-123.
[4]NIU D,CHEN X,YANG J,et al.Composite control for raymond mill based on model predictive control and disturbance observer.Advances in Mechanical Engineering,2016,8(3):1-10.
[5]MINCHALA L I,ZHANG Y,GARZA-CASTANON L E.Predictive control of a closed grinding circuit system in cement industry.IEEETransactions on Industrial Electronics,2018,65(5):4070-4079.
[6]CRAIG I K.Grinding mill modeling and control:past,present and future.Control Conference.Najing,china:IEEE,2012:16-21.
[7]ZHOU P,CHAI T Y.Grinding circuit control:a hierarchical approach using extended 2-DOF decoupling and model approximation.Powder Technology,2011,213(1):14-26.
[8]CHEN X S,YANG J,LI S H.Disturbance observer based multivariable control of ball mill grinding circuits.Journal of Process Control,2009,19(7):1205-213.
[9]ZHANO Dayong,CHAI Tianyou.Fuzzy switching control for sump level interval and hydrocyclone pressure in regrindingprocess.Acta Automatica Sinica,2013,39(5):556-564.(赵大勇,柴天佑.再磨过程泵池液位区间与给矿压力模糊切换控制.自动化学报,2013,39(5):556-564.)
[10]REMES A,AALTONEN J,KOIVOV H.Grinding circuit modeling and simulation of particle size control at Silinjarci concentrator.International Journal Miner Process,2010,96(1/2/3/4):70-78.
[11]ZHANG Xiaolin.Research and application of grinding process expert system.Changchun:Jilin University,2010.(张孝临.磨矿过程专家系统研究与应用.长春:吉林大学,2010.)
[12]CHEN X S,ZHAI J Y,LL Q,et al.Fuzzy logic based on-line efficiency optimization control of a ball mill grinding circuit.Proceedings of the Fourth International Conference on Fuzzy Systems and Knowledge Discovery.Haikou,china:IEEE,2007,2:575-580.
[13]ZHOU P,CHAI T,WANG H.Intelligent optimal-setting control for grinding circuits of mineral processing process.IEEE Transactions on Automation Science and Engineering,2009,6(4):730-743.
[14]ZHOU P,CHAI T,SUN J.Intelligence-based supervisory control for optimal operation of a DCS-controlled grinding system.IEEE Transactions on Control Systems Technology,2013,21(1):162-175.
[15]SUTTON R S,BARTO A G.Reinforcement Learning:An Introduction.Cambridge:MIT Press,1998.
[16]KIUMARSI B,LEWIS F L,JIANG Z P.H infinite control of linear discrete-time systems:off-policy reinforcement learning.Automatica,2017,78(4):144-152.
[17]WANG Xuesong,TIAN Xilan,CHENG Yuhu,et al.Q learning based on Cooperative least squares support vector machines.Journal of Automation,2009,35(2):214-219.(王雪松,田西兰,程玉虎,等.基于协同最小二乘支持向量机的Q学习.自动化学报,2009,35(2):214-219.)
[18]VALENZUELA J,NAJIM K,VILLAR R D,et al.Learning control of an autogenous grinding circuit.International Journal of Mineral Processing,1993,40(1/2):45-56.
[19]CONRADIE A V E,ALDRICH C.Neurocontrol of a ball mill grinding circuit using evolutionary reinforcement learning.Minerals Engineering,2001,14(10):1277-1294.
[20]DAI Wei,CHAI Tianyou.Data driven complex grinding process operation optimization control method.Journal of Automation,2014,40(9):2005-2014.(代伟,柴天佑.数据驱动的复杂磨矿过程运行优化控制方法.自动化学报,2014,40(9):2005-2014.)
[21]JAEGER H.The“echo state”approach to analysing and training recurrent neural networks.Bonn,Germany:German National Research Center for Information Technology GMD Technical Report,2001,148(34):13.
[22]TILAHUM S L,ONG H C.Prey-predator algorithm:a new metaheuristic algorithm for optimization problems.International Journal of Information Technology and Decision Making,2015,14(6):1331-1352.
[23]TILAHUM S,SATHASIVAM S,ONG H C,et al.Prey-predator algrorithm as a new optimization technique using in radia basis function neural networks.Research Journal of Applied Sciences,2013,7(8):383-387.