基于粒子滤波的间歇过程时滞状态估计方法

发布时间：2018-12-12 08:02

【摘要】：间歇过程广泛应用于各类生产行业,在国民经济发展中发挥越来越大的作用。但受限于传感检测技术,关键变量难以实现在线检测,且间歇过程具有高度的非线性、非高斯性和时变性,导致生产过程的监控优化异常困难。基于间歇过程的状态空间模型,状态估计方法能够根据检测值的统计规律,通过滤波的形式实现对关键变量的准确估计。实际生产过程中,存在两种检测值:在线检测值和离线检测值,在线检测值无时滞但精度低,离线检测值有时滞但精度高。另一方面,典型的间歇过程具有周期性批量生产的特点,批次较多。考虑时滞检测值的存在和多批次特性,本文关于间歇过程状态估计的研究内容如下:(1)针对间歇过程中关键变量,在线检测精度低、离线检测时滞大的问题,基于贝叶斯方法提出一种融合时滞检测值信息的状态估计方法。鉴于在线和离线检测值的采样周期不同,分仅有在线检测值和两种检测值并存等两种情况进行分析。该算法以粒子滤波算法为基础,并基于贝叶斯方法对其进行扩展,实现两种检测值的信息融合。数字仿真和生物制氢过程的实验结果表明,该方法能够较好地处理含时滞检测值的状态估计问题,且效果优于不考虑时滞检测值的情况。(2)针对间歇过程的多批次特性,建立时间维度及批次维度的双维状态空间模型,同时考虑时滞检测值的信息,提出一种基于双维状态空间模型的时滞状态估计算法。该算法通过贝叶斯方法及前/后向平滑融合先前批次和时滞检测值的信息,估计效果随批次维度增加而提高,且考虑时滞检测值的状态估计效果优于不考虑时滞检测值时的情况。在数字仿真和生物制氢过程中的仿真应用验证了该算法的有效性。(3)针对双维状态空间模型难以准确建立的问题,结合间歇过程的多批次特性及重复特性,提出一种基于迭代学习的时滞状态估计算法。该算法根据先前批次相同采样时刻的估计状态及测量模型,估计该时刻的期望检测值,并与真实检测值作差得到跟踪误差,通过迭代此误差来提高当前批次的估计精度。对于时滞检测值,同样采用贝叶斯方法和前/后向平滑融合其信息。最后,通过数字仿真和生物制氢过程中的仿真应用验证了所提方法的有效性和实用性。
[Abstract]:Intermittent process is widely used in all kinds of production industries and plays a more and more important role in the development of national economy. However, due to the sensor detection technology, the key variables are difficult to realize on-line detection, and the batch process is highly nonlinear, non-Gao Si and time-varying, which makes the monitoring and optimization of the production process extremely difficult. Based on the state space model of batch process, the state estimation method can estimate the key variables accurately by filtering according to the statistical rule of detection value. In the actual production process, there are two kinds of detection values: the on-line detection value and the off-line detection value, the on-line detection value has no delay but the precision is low, the off-line detection value has the delay but the precision is high. On the other hand, the typical batch process has the characteristic of periodic batch production. Considering the existence of time-delay detection values and the characteristics of multiple batches, the research contents of state estimation for batch processes are as follows: (1) for the key variables in batch processes, the on-line detection accuracy is low, and the off-line detection delay is large. Based on Bayesian method, a state estimation method based on time-delay detection value information is proposed. In view of the different sampling periods of on-line and off-line detection values, only on-line detection values and two detection values coexist to analyze. This algorithm is based on particle filter algorithm and extends it based on Bayesian method to realize the information fusion of two detection values. The experimental results of digital simulation and biological hydrogen production process show that the proposed method can deal with the state estimation problem with time-delay detection value, and the effect is better than that without considering the time-delay detection value. (2) in view of the multi-batch characteristics of the batch process, Two dimensional state space models of time dimension and batch dimension are established, and a delay state estimation algorithm based on two dimensional state space model is proposed, considering the information of time delay detection value at the same time. Based on Bayesian method and forward / backward smoothing, the estimation effect is improved with the increase of batch dimension. The effect of state estimation with time delay detection value is better than that without time delay detection value. The simulation results in digital simulation and biological hydrogen production show that the algorithm is effective. (3) aiming at the problem that the two-dimension state space model is difficult to establish accurately, the multi-batch and repeated characteristics of batch process are combined. A delay state estimation algorithm based on iterative learning is proposed. According to the estimation state and the measurement model of the same sampling time in the previous batch, the algorithm estimates the expected detection value at that time, and the tracking error is obtained from the real detection value, and the estimation accuracy of the current batch is improved by iterating the error. Bayesian method and forward / backward smoothing are also used to fuse the time delay detection values. Finally, the effectiveness and practicability of the proposed method are verified by the digital simulation and the simulation application in the biological hydrogen production process.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN713

【参考文献】