【作者】 李赣平；

【导师】 邵惠鹤；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2007， 博士

【摘要】 智能计算的发展已有较悠久的历史,许多经典的智能算法已取得成功的应用。随着智能计算技术的发展,经典智能算法与来自生命科学中其它生物理论的结合,使得这类算法有了较大进展,最终形成了现代智能计算理论。目前现代智能计算领域已呈现较多的新智能工具,如支持向量机、核方法、粒子群优化算法、迭代学习控制理论等。本文将其中的一部分内容应用到过程控制与优化领域,取得了卓有成效的结果。本文创新的工作主要体现在以下几点:1提出了间歇过程的批次优化控制方法。通常情况下间歇过程的精确机理模型很难获得,由于支持向量机在解决小样本、非线性和高维数的问题时具有强大的功能,支持向量回归模型被用于间歇过程的终端优化控制。为了达到间歇过程所要求的终端性能指标,批次控制方法通过利用间歇过程重复运行的特性来获取间歇过程的优化操作方案,其中二次规划法被用来解决优化控制问题。本文的批次优化控制策略被证明在模型失配和扰动存在的情况下也是收敛的,因而该控制方法具有一定程度的鲁棒性。基于支持向量回归模型的批次优化控制方法是一种综合性的控制方法,充分利用了支持向量机建模智能化的特点和批次优化控制消除建模误差及克服干扰的特点,是一种可靠的优化控制方案。2在线监控和故障诊断在工业过程中对操作安全和产品质量起着重要作用。本文提出了基于核主元和多支持向量机分类的过程监控和故障诊断方法。其中,核主元用来进行故障特征的提取,多支持向量机用来对故障的来源进行分类。该方法首先构造系统正常时的核主元模型,然后将新的数据映射到该核主元模型,对数据进行重构,重构的数据用多元统计指标T2或SPE判断监测过程是否超出了正常的控制限,若有故障发生,则监测程序将给于警告,提示过程出现了异常操作状况。由于原始数据经过核主元的非线性映射后难以求得核主元空间到原始空间的逆映射,因此给故障诊断带来困难。本文采用多支持向量机学习的方法对故障进行分类,避开了求解逆映射的数学方法,直接用智能的方法获得故障的信息,为过程的监控和故障诊断提供了一个新的方法。3在迭代学习控制方面,本文对已有的成果作了总结和分类。在此基础上,本文针对两类基于逆模型的前馈—反馈迭代学习控制方案的鲁棒性作了分析,分别提出了各自的鲁棒收敛性条件。理论上获得的鲁棒收敛域是这两类迭代学习控制方法在学习空间全局收敛的充分条件。该理论结果可以为此类型的迭代学习控制器设计提供参考。4针对无独立状态和终端约束的间歇过程鲁棒优化问题,本文将迭代方法与粒子群优化算法相结合,提出了迭代粒子群算法。对于该算法,首先将控制变量离散化,用标准粒子群优化算法搜索离散控制变量的最优解,然后在随后的迭代过程中将基准移到刚解得的最优值处,同时收缩控制变量的搜索域,使优化性能指标和控制轨线在迭代过程中不断趋于最优解。算法简洁、可行、高效,避免了求解大规模的微分方程组问题。该方法尤其适合系统梯度信息不可得的情况。当系统的梯度信息不可得时,一般的数学方法很难获得优化问题的最优解,而迭代粒子群优化算法利用智能寻优的特点却可获得满意的解。通过控制变量离散化,迭代粒子群优化算法将一个连续问题转化为一个离散问题。优化算法具有并行计算的特点,所有时间段的控制变量被同时优化。算法兼具了迭代法和智能算法的优点,但不需离散化状态变量,仅需在解空间随机寻优,因此又克服了二者的不足之处。该方法在处理间歇过程的鲁棒优化问题时更具优势,编程简单,且可大大减小计算量。更多还原

【Abstract】 Intelligent computation has a long history. Many of the theories have been put into successful applications. With the development of intelligent computation, classical intelligent computation combines other biological theories of the life science, which makes the intelligent computation have big progress and hence leads to form the modern intelligent computation theories. Now there are many new intelligent tools in the modern intelligent computation region such as support vector machines, kernel method, particle swarm optimization algorithm and iterative learning control theory. In the article, some of them are applied for the process control and optimization and hence acquired successful achievements.The main original points of this paper lie below:1. A run-to-run method is presented for the optimal control of batch processes. Generally it is very difficult to acquire an accurate mechanistic model for a batch process. Because Support Vector Machine is powerful for the problems characterized by small samples, nonlinearity, high dimension and local minima, support vector regression models are developed for end-point optimal control of batch processes. To reach the desired end-point properties, a run-to-run control method is used to exploit the repetitive nature of batch processes to determine the optimal operating policy. Quadratic programming (QP) is employed to solve the optimal control problem. The run-to-run control method is proved convergent and robust even when model mismatches and disturbances exist. Therefore the run-to-run optimal control method based on support vector regression model is a comprehensive method which sufficiently utilizes the intelligence of support vector machine’s modeling and the characters of the run-to-run control that can move away the model mismatches and overcome the disturbances. Hence we can see the run-to-run optimal control based on support vector regression model is a reliable optimal control method.2. Online monitoring and fault diagnosis of industry process is extremely important for operation safety and product quality. In the paper, an integrated method is applied for process monitoring and fault diagnosis, which combines kernel principle component analysis (KPCA) for fault feature extraction and multiple support vector machines (MSVMs) for identification of different fault sources. For the algorithm, the kernel primary component model according to the normal system is constructed firstly. Second, the new acquired data are mapped to the kernel primary component model to reconstruct the data. Then the reconstructed data are analyzed by multiple statistical indexes such as T2 or SPE to determine if the supervised process exceeds the normal control limits. If fault occurs, then the supervising program will alarm to prompt that the process has abnormal states. Because it is hard to get the inverse mapping from the kernel space to the original space after the original data are nonlinear mapped by the KPCA method, hence it is difficult to diagnose how the fault happened. The paper adopts the learning method of MSVMs to classify the faults, which avoids the mathematical solution for the inverse mapping and acquires the fault information through intelligent method directly. Such an integrated method provides a new way for process monitoring and fault diagnosis.3. As far as iterative learning control concerned, the paper summarized and classified the iterative learning control theories. On the basis, the paper analyzed the robust performances of two kinds of feedforward-feedback iterative learning control based on inverse model. Then robust convergent conditions are presented respectively. The theoretically acquired robust convergent regions are the sufficient conditions for the two kinds of iterative learning control schemes’global convergences in the learning space. The acquired theoretical results can provide the references for the design of iterative learning controllers.4. To solve optimization problems of batch processes without state independent and end-point constraints, the article combined the iteration method and the particle swarm optimization algorithm together and proposed an iterative particle swarm algorithm. For the algorithm, the control variables are discretized firstly and the standard particle swarm optimization algorithm is used to search for the best solution of the discretized control variables. Next, the benchmark is moved to the acquired optimal values in the subsequent iterations and the searching space contracted at the same time, hence the optimization performance index and control profile could achieve the best value gradually through iterations. The algorithm is simple, feasible and efficient. It avoided the problem solving large-scale differential equation group. The algorithm is especially practical when the system’s gradient information is unavailable. In the case, it is hard for general mathematical method to acquire the optimal solutions. However, the iterative particle swarm algorithm can acquire the satisfying results based on its intelligent researching. By discretizing the control variables, a continuous dynamic optimization problem is transformed to a discrete problem. The algorithm has the characters of parallel computations which make the control profile at all time stage optimized simultaneously. The algorithm has the advantages of iteration methods and intelligent algorithms. Nevertheless, it don’t need discretization of state variables and only search the optimal control profile in the solution space randomly, hence overcome the disadvantages of iteration methods and intelligent algorithms. The algorithm is of more advantage when solve the robust optimization of batch process which make the program simple and can reduce the computation largely.更多还原