非线性变量误差系统的辨识方法研究

发布时间：2018-06-07 03:33

本文选题：EIV系统 + 合成模型　；参考：《东华大学》2016年博士论文

【摘要】：随着工业大数据时代的来临,发展基于数据的建模方法显得尤为重要。基于工业过程数据的建模问题,一个重要的挑战是收集到的数据含有噪声。传统的基于工业过程输入输出数据的建模方法只考虑输出测量数据存在噪声,而忽略输入测量数据存在噪声,会导致已建立的模型存在较大的偏差甚至是无效的。变量误差(EIV)系统是一种考虑可观测或测量输入输出数据均含有误差的系统。传统的EIV系统并没有考虑输入生成动态过程,本论文针对具有输入生成动态过程的EIV系统进行研究,主要内容如下:(1)针对具有线性输入生成过程的线性EIV系统进行研究,即线性EIV系统包括线性输入生成过程和线性EIV过程。通过分析线性EIV系统中所有可观测或测量变量之间的因果关系,提出合成线性EIV模型,合成线性EIV模型考虑了线性EIV系统中所有观测或测量变量对被估计的输入变量(实际输入)产生的影响。进一步,基于期望最大化算法估计合成线性EIV模型的参数。最后,采用数值仿真例子及混合水箱实验验证了提出方法的有效性。(2)考虑到工业过程本身具有非线性性,对非线性EIV系统进行研究。其中非线性EIV系统是由非线性输入生成过程和非线性EIV过程组成。考虑到直接辨识非线性EIV系统存在较大的困难,采用多个线性EIV模型逼近非线性EIV模型的思想,对非线性模型进行辨识。通过分析非线性EIV系统中所有观测或测量变量之间的因果关系,提出了合成非线性EIV模型。在提出的合成非线性EIV模型中,设计了一种并行粒子滤波的策略用于估计每个线性EIV模型的实际输入(被估计的状态),即对多个线性EIV模型中的每一个线性模型均执行一个粒子滤波,且与多个线性EIV模型相对应的多个粒子滤波并行执行。进而,在最大似然的估计框架下,基于采集到的非线性EIV系统的输入输出数据,通过期望最大化算法对提出的合成非线性EIV模型的参数进行估计。数值模拟例子和实验例子验证了提出方法的有效性。(3)考虑到工业过程可观测或测量数据通常存在异常观测数据,针对非线性EIV系统,提出了一种鲁棒辨识方法。考虑到t分布可以通过调整自由度,有效地解决异常观测数据对辨识过程造成的干扰。采用t分布对测量噪声进行建模,而不是传统的高斯分布。进一步,为了避免直接辨识非线性模型的复杂性,提出了采用多个鲁棒线性模型近似非线性EIV系统。进而,考虑指数函数作为权重函数加权合成多个鲁棒线性局部模型的输出得到非线性EIV系统的全局输出。在极大似然估计的框架下,通过期望最大化算法同时估计得到每个鲁棒线性局部模型的参数及权重函数的参数。最后,提出方法的有效性通过使用数字仿真例子及实验例子进行了验证。(4)基于变分贝叶斯方法,针对存在异常观测数据的非线性EIV系统,提出了一种鲁棒辨识方法。基于观测或测量得到的非线性EIV系统的输入输出数据,在贝叶斯框架下,通过使用变分贝叶斯期望最大化算法,解决了非线性EIV系统的参数估计问题,得到了非线性EIV系统模型参数的分布,而不是参数的点估计。并通过一个连续发酵过程作为仿真例子及一个三串联水箱系统作为实验例子验证了提出方法的优越性。最后,对本论文的研究内容进行了总结,并对下一步的研究工作和相关问题进行了展望。
[Abstract]:With the coming of the era of industrial data, it is very important to develop the modeling method based on data. Based on the modeling of industrial process data, an important challenge is that the collected data contains noise. The traditional modeling method based on the input and output data of industrial process only considers the noise of the output data, and neglects the input. There is a noise in the measured data, which leads to the existence of large deviations or even ineffectiveness of the established model. The variable error (EIV) system is a system that takes into account the error of the input and output data which can be observed or measured. The traditional EIV system does not consider the dynamic process of the input generation, and this paper is aimed at the EI with the dynamic input generation process. The main contents of the V system are as follows: (1) a linear EIV system with linear input generating process is studied, that is, linear EIV system includes linear input generating process and linear EIV process. By analyzing the causality between all observable or measured variables in linear EIV system, a linear EIV model is proposed and a linear E is synthesized. The IV model considers the effects of all observation or measurement variables on the estimated input variables (actual input) in linear EIV systems. Further, based on the expectation maximization algorithm, the parameters of the synthetic linear EIV model are estimated. Finally, the effectiveness of the proposed method is verified by numerical simulation examples and mixed water tank experiments. (2) consideration of industry. The nonlinear process itself is nonlinear, and the nonlinear EIV system is studied. The nonlinear EIV system is composed of nonlinear input generating process and nonlinear EIV process. Considering the difficulties in identifying the nonlinear EIV system directly, multiple linear EIV models are used to approximate the non linear EIV model, and the nonlinear model is identified. By analyzing the causality between all observed or measured variables in the nonlinear EIV system, a synthetic nonlinear EIV model is proposed. In the proposed nonlinear EIV model, a parallel particle filtering strategy is designed to estimate the actual input (estimated state) of each linear EIV model, that is, multiple linear EIV models. Each linear model in the linear model performs a particle filter and performs parallel execution of multiple particle filters corresponding to multiple linear EIV models. Then, under the maximum likelihood estimation framework, the input and output data of the collected nonlinear EIV system are based on the expected maximization algorithm for the parameters of the proposed synthetic nonlinear EIV model. A line estimation. Numerical simulation examples and experimental examples show the effectiveness of the proposed method. (3) a robust identification method is proposed for nonlinear EIV systems considering the observable or measured data in industrial processes. A robust identification method is proposed for nonlinear EIV systems. Considering that the t distribution can be used to solve the abnormal observation data effectively by adjusting the degree of freedom. The disturbance caused by the recognition process is modeled by t distribution rather than the traditional Gauss distribution. Further, in order to avoid the complexity of nonlinear model identification directly, a multi robust linear model is proposed to approximate the nonlinear EIV system. Then, the exponential function is considered as weighting function to synthesize multiple robust linear bureaus. The output of the model is obtained from the global output of the nonlinear EIV system. Under the framework of the maximum likelihood estimation, the parameters and weights of each robust linear local model are estimated by the expected maximization algorithm. Finally, the effectiveness of the proposed method is verified by using digital simulation examples and experimental examples. (4) In the variational Bayesian method, a robust identification method is proposed for nonlinear EIV systems with abnormal observation data. Based on the observation or measurement of the input and output data of the nonlinear EIV system, the parameter estimation of the nonlinear EIV system is solved by using the variational Bayesian expectation maximization algorithm. The distribution of the parameters of the nonlinear EIV system model, not the point estimation of the parameters, is obtained. The superiority of the proposed method is verified by a continuous fermentation process as a simulation example and a three series water tank system as an experimental example. Finally, the research content of this paper is summarized, and the next step of the research work is also discussed. The related issues were prospected.
【学位授予单位】：东华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：N945.14

【相似文献】