硅单晶等径阶段直径模型辨识与控制研究
发布时间：2019-06-07 10:08
【摘要】：硅单晶作为半导体行业基础材料之一,对集成电路产业发展起着重要作用。随着极大规模集成电路发展,对硅单晶提出了大尺寸、高品质等要求。目前制备硅单晶最为重要的方法是基于传统控制结构的直拉法。在传统控制结构中,目标温度跟踪曲线依赖人工经验且控制器参数常通过反复实验获得。除此之外,热场温度控制晶体直径环节存在非线性、大时滞、缓时变特性。为避免目标温度跟踪曲线和控制器参数设置不合理,最有效策略是辨识热场温度-晶体直径非线性动态模型,并采用基于模型的控制策略实现晶体直径控制,从而达到提高晶体品质的目的。本文在硅单晶生长原理及传统硅单晶控制结构研究基础上,提出一种恒拉速热场温度-晶体直径辨识和控制方案。利用等径阶段的热场温度和晶体直径数据及输出相关性时滞确定算法获得热场温度-晶体直径非线性动态模型时滞;采用利普希茨商确定模型输入输出阶次;在模型参数辨识后采用增强型相关检验算法优化模型阶次并对辨识模型进行检验;参数辨识分别基于动态BP神经网络和栈式稀疏自动编码器实现。获得有效热场温度-晶体直径模型后,分别采用基于动态BP神经网络和基于栈式稀疏自动编码器预测模型的广义预测控制算法对晶体直径进行控制仿真实验。实验结果表明:提出的辨识方案可以有效获得热场温度-晶体直径大时滞非线性动态模型;将栈式稀疏自动编码器引入广义预测控制能够有效跟踪设定晶体直径,但该算法在线修正实时性较差。基于动态BP神经网络的广义预测控制算法不仅能够准确跟踪设定晶体直径,同时能够通过在线学习对预测模型予以修正,以适应系统缓时变特征。
[Abstract]:Silicon single crystal, as one of the basic materials in semiconductor industry, plays an important role in the development of integrated circuit industry. With the development of very large scale integrated circuits, large size and high quality requirements have been put forward for silicon single crystals. At present, the most important method to prepare silicon single crystal is Czochralski method based on traditional control structure. In the traditional control structure, the target temperature tracking curve depends on artificial experience and the controller parameters are often obtained by repeated experiments. In addition, the diameter link of thermal field temperature control crystal has nonlinear, large time delay and slow time-varying characteristics. In order to avoid unreasonable setting of target temperature tracking curve and controller parameters, the most effective strategy is to identify the nonlinear dynamic model of temperature-crystal diameter in thermal field, and to realize crystal diameter control by using model-based control strategy. So as to achieve the purpose of improving the crystal quality. Based on the growth principle of silicon single crystal and the control structure of traditional silicon single crystal, a temperature-crystal diameter identification and control scheme for constant pull speed thermal field is proposed in this paper. The nonlinear dynamic model delay of thermal field temperature and crystal diameter is obtained by using the thermal field temperature and crystal diameter data of equal diameter stage and the time delay determination algorithm of output correlation, and the input and output order of the model is determined by Lipschitz quotient. After the model parameter identification, the enhanced correlation test algorithm is used to optimize the model order and test the identification model, and the parameter identification is realized based on dynamic BP neural network and stack sparse automatic encoder, respectively. After the effective thermal field temperature-crystal diameter model is obtained, the generalized predictive control algorithm based on dynamic BP neural network and stack sparse automatic encoder prediction model is used to simulate the crystal diameter control. The experimental results show that the proposed identification scheme can effectively obtain the nonlinear dynamic model of thermal field temperature-crystal diameter with large time delay. The introduction of stack sparse automatic encoder into generalized predictive control can effectively track the set crystal diameter, but the real-time performance of the algorithm is poor. The generalized predictive control algorithm based on dynamic BP neural network can not only accurately track the set crystal diameter, but also modify the prediction model through online learning to adapt to the slow time-varying characteristics of the system.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN304.12
[Abstract]:Silicon single crystal, as one of the basic materials in semiconductor industry, plays an important role in the development of integrated circuit industry. With the development of very large scale integrated circuits, large size and high quality requirements have been put forward for silicon single crystals. At present, the most important method to prepare silicon single crystal is Czochralski method based on traditional control structure. In the traditional control structure, the target temperature tracking curve depends on artificial experience and the controller parameters are often obtained by repeated experiments. In addition, the diameter link of thermal field temperature control crystal has nonlinear, large time delay and slow time-varying characteristics. In order to avoid unreasonable setting of target temperature tracking curve and controller parameters, the most effective strategy is to identify the nonlinear dynamic model of temperature-crystal diameter in thermal field, and to realize crystal diameter control by using model-based control strategy. So as to achieve the purpose of improving the crystal quality. Based on the growth principle of silicon single crystal and the control structure of traditional silicon single crystal, a temperature-crystal diameter identification and control scheme for constant pull speed thermal field is proposed in this paper. The nonlinear dynamic model delay of thermal field temperature and crystal diameter is obtained by using the thermal field temperature and crystal diameter data of equal diameter stage and the time delay determination algorithm of output correlation, and the input and output order of the model is determined by Lipschitz quotient. After the model parameter identification, the enhanced correlation test algorithm is used to optimize the model order and test the identification model, and the parameter identification is realized based on dynamic BP neural network and stack sparse automatic encoder, respectively. After the effective thermal field temperature-crystal diameter model is obtained, the generalized predictive control algorithm based on dynamic BP neural network and stack sparse automatic encoder prediction model is used to simulate the crystal diameter control. The experimental results show that the proposed identification scheme can effectively obtain the nonlinear dynamic model of thermal field temperature-crystal diameter with large time delay. The introduction of stack sparse automatic encoder into generalized predictive control can effectively track the set crystal diameter, but the real-time performance of the algorithm is poor. The generalized predictive control algorithm based on dynamic BP neural network can not only accurately track the set crystal diameter, but also modify the prediction model through online learning to adapt to the slow time-varying characteristics of the system.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN304.12
【参考文献】
相关期刊论文 前10条
1 赵飞翔;刘永祥;霍凯;;基于栈式降噪稀疏自动编码器的雷达目标识别方法[J];雷达学报;2017年02期
2 刘丁;赵小国;赵跃;;直拉硅单晶生长过程建模与控制研究综述[J];控制理论与应用;2017年01期
3 戴晓爱;郭守恒;任m，
本文编号：2494716
本文链接：https://www.wllwen.com/kejilunwen/dianzigongchenglunwen/2494716.html
