基于ELM的时序数据建模及应用研究

发布时间：2018-05-12 14:58

本文选题：数据建模 + 自适应滤波　；参考：《广西大学》2017年硕士论文

【摘要】：使用模型来进行预测、诊断与分析评价,一直是工业过程优化研究的重点手段。随着工业过程的复杂化,依据物理化学机理建立数学模型的方法已变得十分困难,人们转而投向基于数据驱动的建模研究,使其在数十年间得到了飞速的发展。作为数据驱动建模的一种,人工神经网络凭借其强大的非线性映射能力成为了建模领域的热点之一,但随着数据量的增加,传统神经网络固有的缺陷和不足被放大,例如训练速度过慢、参数难以选择等,另外由于工业环境的特殊性,过程数据必定会携带着不同程度的噪声干扰,这些因素都将使得模型无法满足实际需求。针对上述问题,本论文研究了基于超限学习机(Extreme learning machine,ELM)的数据建模方法,主要的内容和研究成果包括:1.针对工业过程数据的特点分析了传统非线性滤波算法的不足,然后详细介绍了一种基于局部最优拟合的自适应滤波算法,并通过洛伦兹系统及电功率两组数据的去噪实验将该算法与小波域阈值滤波及扩展卡尔曼滤波进行了对比,证明了该算法能够更有效地减小工业过程数据中的噪声干扰。2.针对传统神经网络训练过程中耗时过长、参数难以选择的问题,对固定型ELM及增量型ELM进行了研究,并针对增量型ELM收敛速度不稳定的问题介绍了一种十分有效的改进算法,然后通过回归问题的仿真实验将该改进算法与多种神经网络进行了对比,从收敛性能、预测精度、训练速度及网络稳定性等多个角度评价了改进效果,从而证明了该改进算法的可行性与优越性。3.从MATLAB环境与工业现场环境存在本质区别的角度出发,对能真实反映工业对象特性的仿真系统提出了需求。以流程工业综合自动化半实物仿真实验系统为平台,对我国选矿工业中常用的两段全闭路磨矿回路进行了动态仿真,并详细介绍了仿真的实现流程,为后续的数据建模及模型验证提供了支撑。4.以螺旋分级机为建模对象,利用磨矿过程仿真系统产生的数据,基于自适应滤波算法和超限学习机改进算法建立了螺旋分级机的预测模型,并通过离线及在线验证的实验验证了所建模型的有效性,进而证明了本文所研究的自适应滤波及超限学习机改进算法应用于工业数据建模时是切实可行的。
[Abstract]:The use of models for prediction, diagnosis and analysis and evaluation has always been the focus of industrial process optimization research. With the complication of industrial process, the method of establishing mathematical model based on physicochemical mechanism has become very difficult. People turn to data-driven modeling research, which makes it develop rapidly in decades. As a kind of data-driven modeling, artificial neural network (Ann) has become one of the hotspots in modeling field by virtue of its strong nonlinear mapping ability. However, with the increase of data volume, the inherent defects and shortcomings of traditional neural networks are amplified. For example, the training speed is too slow, the parameters are difficult to choose and so on. In addition, because of the particularity of the industrial environment, the process data must carry different degrees of noise interference, these factors will make the model can not meet the actual needs. To solve the above problems, this paper studies the data modeling method based on extreme learning machine. The main content and research results include: 1. According to the characteristics of industrial process data, the shortcomings of traditional nonlinear filtering algorithm are analyzed, and then an adaptive filtering algorithm based on local optimal fitting is introduced in detail. The algorithm is compared with wavelet domain threshold filter and extended Kalman filter through the experiments of Lorentz system and electric power data. It is proved that the algorithm can reduce the noise disturbance of industrial process data more effectively. Aiming at the problem that the training process of traditional neural network is too long and the parameters are difficult to select, the fixed ELM and incremental ELM are studied, and a very effective improved algorithm is introduced to solve the problem of unstable convergence rate of incremental ELM. Then the improved algorithm is compared with various neural networks through the simulation experiment of regression problem. The improved algorithm is evaluated from the convergence performance, prediction accuracy, training speed and network stability. The feasibility and superiority of the improved algorithm are proved. From the point of view of the essential difference between the MATLAB environment and the industrial field environment, the requirements of the simulation system which can truly reflect the characteristics of the industrial objects are put forward. Based on the process industry integrated automation hardware-in-the-loop simulation system, the dynamic simulation of the two-stage closed circuit grinding circuit commonly used in China's mineral processing industry is carried out, and the realization flow of the simulation is introduced in detail. Provide support for subsequent data modeling and model validation. 4. 4. Taking the spiral classifier as the modeling object, the prediction model of the spiral classifier is established based on the adaptive filtering algorithm and the improved algorithm of the over-limit learning machine, using the data generated by the grinding process simulation system. The validity of the proposed model is verified by off-line and on-line verification experiments, and it is proved that the adaptive filter and the improved algorithm of over-limit learning machine are feasible when applied to industrial data modeling.
【学位授予单位】：广西大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【参考文献】