基于改进核主元分析的过程监测方法研究

发布时间：2018-07-18 20:44

【摘要】：随着现代工业过程规模日益扩大,流程日益复杂,如何保证过程运行安全并且提高产品质量是目前工业生产企业急需解决的两个问题,过程监测技术是解决这两个问题的有效方法。但由于实际工业过程的复杂性以及波动性,准确的过程模型建立和应用都十分困难,传统的基于定性和定量模型的理论和方法受到一定局限。由于智能仪表和计算机技术在工业过程应用中的发展,大量高维的且具有强相关性的过程状态数据被采集并存储,很难从中去掉冗余和干扰提取有用信息。多元统计过程监测技术作为应对多变量的相关性的方法,在过去的十几年间得到持续的关注与发展。本文在前人工作成果的基础上,针对非线性工业过程中由于设备老化、过程漂移、传感器测量误差等因素引起的参数漂移问题,做了以下的研究工作：(1)针对过程样本逐渐增大或参数漂移的问题,本文结合基于滑动窗口机制的核主元分析方法和指数加权的核主元分析方法,提出一种自适应核主元分析方法。当过程采集到新样本,先利用滑动窗口判断样本是否满足模型更新的条件。如果满足模型更新条件,则采用指数加权核主元分析方法更新模型；反之,则不进行模型更新,直至采集到下一个正常样本。将该方法用于电熔镁炉工作过程的监测,仿真结果验证方法的可行性。(2)传统的核主元分析方法基于样本不包含劣点的假设,但实际的工业过程所采集的数据往往含有劣点。即使将其映射到特征空间,劣点问题依然存在,对模型产生很大影响,导致过程监测结果不准确。针对这个问题,本文提出一种改进的核主元分析方法,在重建误差最小的意义下定义特征空间的损失函数,采用惩罚因子的迭代核主元分析求解,消除劣点影响,并且用基于遗忘因子的核矩阵更新方法,保证模型更好的符合过程的变化；对新样本先计算重建误差,判断是否是劣点,如果是劣点则重建后更新,如果不是则正常直接更新模型。将此方法应用于电熔镁炉工作过程中进行过程监测,仿真结果表明,改进的核主元分析方法能够减小劣点对模型的影响,提高模型准确性。
[Abstract]:With the increasing scale and complexity of modern industrial process, how to ensure the safety of process operation and improve the quality of products are two urgent problems to be solved by industrial production enterprises. Process monitoring technology is an effective method to solve these two problems. However, due to the complexity and volatility of actual industrial processes, it is very difficult to establish and apply accurate process models, and the traditional theories and methods based on qualitative and quantitative models are limited to a certain extent. Due to the development of intelligent instruments and computer technology in industrial process applications, a large number of high-dimensional and strongly correlated process state data are collected and stored, it is difficult to remove redundancy and interference to extract useful information. As a method to deal with multivariate correlation, multivariate statistical process monitoring technology has been continuously concerned and developed in the past ten years. On the basis of previous work, this paper aims at the problem of parameter drift caused by equipment aging, process drift and sensor measurement error in nonlinear industrial process. The following research works have been done: (1) aiming at the problem of process samples increasing gradually or parameter drift, this paper combines the kernel principal component analysis method based on sliding window mechanism and the exponential weighted kernel principal component analysis method. An adaptive kernel principal component analysis method is proposed. When a new sample is collected, the sliding window is used to determine whether the sample satisfies the condition of model updating. If the model updating condition is satisfied, the exponential weighted kernel principal component analysis method is used to update the model, whereas the model update is not carried out until the next normal sample is collected. The method is used to monitor the working process of the fused magnesium furnace and the simulation results verify the feasibility of the method. (2) the traditional kernel principal component analysis method is based on the assumption that the sample does not contain the inferior points, but the actual data collected in the industrial process often contain the inferior points. Even if it is mapped to the feature space, the problem of inferior points still exists, which has a great influence on the model and results in inaccurate process monitoring. In order to solve this problem, an improved kernel principal component analysis method is proposed in this paper, which defines the loss function of the feature space in the sense of minimum reconstruction error, and solves the problem by using the penalty factor iterative kernel principal component analysis to eliminate the influence of inferior points. Moreover, the kernel matrix updating method based on forgetting factor is used to ensure that the model is more consistent with the change of process, and the reconstruction error is first calculated for the new sample to determine whether it is a bad point, and if it is a bad point, the reconstruction is later updated. If not, update the model directly. The simulation results show that the improved kernel principal component analysis method can reduce the influence of inferior points on the model and improve the accuracy of the model.
【学位授予单位】：东北大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TB49

【相似文献】