鲁棒模型预测控制与基于数据重构的故障检测

发布时间：2018-05-17 16:06

本文选题：模型预测控制 + 过程监测　；参考：《天津大学》2016年博士论文

【摘要】：近年来,随着工业规模的不断扩大、复杂程度的不断提高以及大量数据的涌现,工业过程的质量控制和性能控制面临着巨大的挑战。预测和监测逐渐成为了两种必不可少的手段,所以发展有效的工具实现对工业过程的及时、稳定预测和监测,降低各种因素对工业运行性能的负面影响是具有非常重要意义的。本文的主要工作如下:1.提出了一个延迟依赖记忆型鲁棒模型预测算法。该系统的时间延迟虽然大小未知,但具有明确上下界。将最小最大优化问题转换为求“在最糟糕状况下”代价函数上界最小问题,利用线性矩阵不等式得到了一个新的代价函数单调性充分条件;记忆型状态反馈控制率被首次引入到鲁棒模型预测控制中,使用所得到的充分条件证明了引入的控制率能使代价函数的上界最小,且能保证闭环系统渐近稳定;通过一个非线性系统的例子说明了所给算法良好的性能。2.提出了一个新的多尺度非线性过程质量监测与故障检测方法,被称为尺度筛选多尺度算法(Scale-Sifting Multi-Scale Algorithm,SMA)。这个算法包括尺度筛选基准、数据分解与重构以及改进的动态核偏最小二乘等三个部分。与现在流行的多尺度算法相比,尺度筛选多尺度算法的关键特点在于能够在没有任何先验假设的条件下实现关键尺度数据的筛选和重构,数据在没有任何先验假设的情况下被分解;尺度筛选基准被用于筛选出包含过程异常状况关键特征的关键尺度;根据所选出的尺度进行全局数据重构;改进的动态核偏最小二乘被用于分析中心化后的重构数据。仿真和实验结果表明尺度筛选多尺度算法在多尺度故障检测方面优越的性能。3.提出了一种新的压缩稀疏主元分析算法(Compressive Sparse Principal Component Analysis,CSPCA)用于过程监测与故障检测。该方法由压缩部分重构算法以及改进的稀疏主元分析算法构成。CSPCA算法在没有任何先验假设的情况下实现对于异常信号的压缩和部分重构。根据主元分析与数据矩阵奇异值分解之间的关系,通过将2,1L范数作为目标函数和惩罚项得到一个获取稀疏主元负载的凸优化问题,并通过一个迭代算法求解。2,1l惩罚项的引入使得获得稀疏主元的同时,还能兼顾不大得分主元与小得分主元在监测算法中的作用。改进的稀疏主元分析算法的单调性和全局收敛性得到证明。两个标准算例的结果显示了CSPCA优越的性能。
[Abstract]:In recent years, with the continuous expansion of industrial scale, the increasing complexity and the emergence of a large number of data, the quality control and performance control of industrial processes are facing great challenges. Prediction and monitoring have gradually become two indispensable means, so it is of great significance to develop effective tools to realize timely, stable prediction and monitoring of industrial processes, and to reduce the negative effects of various factors on industrial performance. The main work of this paper is as follows: 1. A delay dependent memory robust model prediction algorithm is proposed. The time delay of the system is unknown, but it has definite upper and lower bounds. The minimum maximum optimization problem is transformed into the upper bound minimum problem of the cost function in the worst case, and a new sufficient condition for monotonicity of the cost function is obtained by using the linear matrix inequality (LMI). The memory-type state feedback control rate is introduced into robust model predictive control for the first time. The sufficient conditions obtained prove that the proposed control rate can minimize the upper bound of the cost function and ensure the asymptotic stability of the closed-loop system. An example of a nonlinear system is given to illustrate the good performance of the proposed algorithm. 2. 2. A new multi-scale nonlinear process quality monitoring and fault detection method is proposed, which is called scaling sifting Multi-Scale algorithm. The algorithm consists of three parts: scale screening benchmark, data decomposition and reconstruction, and improved dynamic kernel partial least squares. Compared with the popular multiscale algorithm, the key feature of the scale filtering algorithm is that it can filter and reconstruct the critical scale data without any prior assumptions. The data is decomposed without any prior hypothesis, the scale screening datum is used to screen the key scale which contains the key characteristics of the abnormal state of the process, and the global data is reconstructed according to the selected scale. The improved dynamic kernel partial least squares is used to analyze the reconstructed data after centralization. Simulation and experimental results show that the multi-scale filtering algorithm has excellent performance in multi-scale fault detection. A new compressed sparse principal component analysis (PCA) algorithm, Compressive Sparse Principal Component Analysis (CSPCA), is proposed for process monitoring and fault detection. The proposed method consists of a compression partial reconstruction algorithm and an improved sparse principal component analysis algorithm. The CSPCA algorithm realizes the compression and partial reconstruction of abnormal signals without any prior assumptions. According to the relationship between principal component analysis and singular value decomposition of data matrix, a convex optimization problem for obtaining sparse principal component load is obtained by using 2L norm as objective function and penalty term. By using an iterative algorithm to solve the penalty term of .2ll, the sparse principal component can be obtained, and the function of small score principal component and small score principal component in monitoring algorithm can be taken into account at the same time. The monotonicity and global convergence of the improved sparse principal component analysis algorithm are proved. The results of two standard examples show the superior performance of CSPCA.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP13

【参考文献】