基于统计方法的模型在线校正方法研究

发布时间：2018-01-26 22:17

本文关键词： 过程建模模型校正统计方法模型监控改进RPLS　出处：《东北大学》2014年硕士论文　论文类型：学位论文

【摘要】：在连续工业生产过程中,一般难以甚至无法建立严格精确的数学模型来描述实际工业过程,因此模型误差必然存在。此外,即使模型本身可以较为准确地描述过程变化,在模型投入运行后,由于系统的时变性和不确定性等,随着时间的推移,模型误差也可能加大。解决以上问题一般采用的方法是通过在线的模型校正,使得模型始终反映最新的过程变化。对模型进行校正,对于模型的正常运行有着重要意义。本文在分析过程建模、模型校正存在的主要问题及现有方法、思路的基础上,利用统计方法对模型进行建模、监控及在线校正。首先介绍了基于数据驱动的建模方法和模型监控方法,研究了偏最小二乘算法和主成分分析法,在监控指标的选择上,选取了一种由常规统计量T2统计量及SPE统计量结合的组合监控指标,在实际应用中更为简洁。基于该统计监控指标,提出了一种改进的递推偏最小二乘算法——基于PCA组合监控指标的遗忘因子-滑窗块式RPLS算法。该算法利用一定规模的数据块对模型进行校正,通过控制数据块个数来控制数据规模,防止数据饱和；通过遗忘因子保证新数据块对模型的影响程度,同时降低旧数据块的置信度；将模型的实时监控指标加入到遗忘因子的计算当中,能根据当前模型的运行状态动态调整遗忘因子。最后,利用该方法参数的动态调整性能,进一步提出了模型的在线校正策略。选择田纳西-伊斯曼过程作为仿真对象,在仿真研究中,证明了本文提出的改进递推偏最小二乘算法及校正策略的可行性及有效性。
[Abstract]:In the process of continuous industrial production, it is difficult or even impossible to establish a strict and accurate mathematical model to describe the actual industrial process, so the model error must exist. Even if the model itself can accurately describe the process changes, after the model is put into operation, due to the system time variability and uncertainty, with the passage of time. The model error may also be increased. The general method to solve the above problem is to correct the model through on-line model correction, so that the model always reflects the latest process changes and corrects the model. It is very important for the normal operation of the model. Based on the analysis of the process modeling, the main problems of model correction and the existing methods, this paper uses the statistical method to model the model. Monitoring and online correction. Firstly, the data-driven modeling method and model monitoring method are introduced, and the partial least squares algorithm and principal component analysis method are studied in the selection of monitoring indicators. A combined monitoring index, which combines T2 statistics with SPE statistics, is selected, which is more succinct in practical application. An improved recursive partial least squares algorithm (RPLS) based on PCA combined monitoring index is proposed. The algorithm uses a certain size of data blocks to correct the model. The data size is controlled by controlling the number of data blocks to prevent data saturation. The influence degree of the new data block on the model is guaranteed by forgetting factor, and the confidence degree of the old data block is reduced at the same time. The real-time monitoring index of the model can be added to the calculation of the forgetting factor, and the forgetting factor can be dynamically adjusted according to the running state of the current model. Finally, the dynamic adjustment performance of the parameters of the method is used. Furthermore, the online calibration strategy of the model is proposed. Tennessee Eastman process is selected as the simulation object in the simulation research. The feasibility and effectiveness of the improved recursive partial least squares algorithm and its correction strategy are proved.
【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】