WO2017088208A1 - 一种由数据差异驱动的间歇过程自学习动态优化方法 - Google Patents

一种由数据差异驱动的间歇过程自学习动态优化方法 Download PDF

Info

Publication number
WO2017088208A1
WO2017088208A1 PCT/CN2015/096377 CN2015096377W WO2017088208A1 WO 2017088208 A1 WO2017088208 A1 WO 2017088208A1 CN 2015096377 W CN2015096377 W CN 2015096377W WO 2017088208 A1 WO2017088208 A1 WO 2017088208A1
Authority
WO
WIPO (PCT)
Prior art keywords
batch
data
variable
optimization
period
Prior art date
Application number
PCT/CN2015/096377
Other languages
English (en)
French (fr)
Inventor
栾小丽
王志国
刘飞
Original Assignee
江南大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 江南大学 filed Critical 江南大学
Priority to US15/566,902 priority Critical patent/US10254729B2/en
Publication of WO2017088208A1 publication Critical patent/WO2017088208A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/04—Programme control other than numerical control, i.e. in sequence controllers or logic controllers
    • G05B19/042—Programme control other than numerical control, i.e. in sequence controllers or logic controllers using digital processors
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0205—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric not using a model or a simulator of the controlled system
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32077—Batch control system

Definitions

  • the invention belongs to the field of chemical process manufacturing, and is a self-learning dynamic optimization method involving a model-free intermittent process driven by data difference.
  • the method is applicable to the dynamic optimization of operating trajectories including batch reactors, batch distillation columns, batch drying, batch fermentation, batch crystallization, and other processes and systems operating in a batch mode.
  • the batch process is widely used in the production and preparation of food, polymer, pharmaceutical and other products, and plays an important role in the chemical production and process industries.
  • process control and optimization technology With the continuous development of computer technology, process control and optimization technology, the quality control and optimization of batch process has become one of the hotspots of current industrial and academic research, and has important significance for the development of batch process industry.
  • the quality control and optimization problems of most batch processes are based on the optimization model based on mechanism model. It has high optimization efficiency and can be used for online control and optimization, but it must rely on complex process models.
  • the flexibility of the batch process determines that the processed product is subject to change at any time and does not have the large number of experimental and time conditions required to identify the model. At the same time, the nonlinearity, uncertainty and interference of the batch process are more serious than the continuous process, which brings many problems to the establishment of the accurate and reliable mechanism model of the intermittent process.
  • the data-driven optimization method has no obvious prior knowledge of the process and has obvious advantages.
  • the optimization performance has limitations. Therefore, how to use the repeating characteristics of the batch process to update the trajectory of the current batch optimization variables through the iterative algorithm according to the historical batch information, so as to continuously improve the quality index and improve the production efficiency, which becomes the difficult point and focus of the batch process optimization field. .
  • the invention relates to a model-free self-learning dynamic optimization method driven by data difference for a batch process, and uses an perturbation method to establish an initial optimization strategy of an optimized variable setting curve.
  • a model-free self-learning dynamic optimization method based on data difference driving for intermittent processes, based entirely on the operational data of the production process, without prior knowledge and mechanism model of the process mechanism.
  • the method comprises two parts.
  • the first part is the establishment of initial data collection and initial optimization strategies.
  • the second part is a self-learning algorithm and calculation steps based on new batch data.
  • Step 1 Collect the variables to be optimized and the final quality or yield indicators by batch for the complete batch process data.
  • the data collection time interval may be an equal time interval or an unequal time interval.
  • the variable to be optimized of the process does not change significantly, or its change does not have a significant impact on the final quality or yield indicator. 20-30 sets of valid data are generally required.
  • Step 2 Perform the principal component analysis on the collected data according to the batch and eliminate the singular points in the pivot mode, so that all data points are within a certain degree of confidence.
  • Step 3 The operation data after the singular point is removed is divided into N segments at equal intervals on the time axis or unequal intervals.
  • the value of the time period variable consists of the individual batch data of the variable to be optimized in a specific time interval.
  • a data matrix composed of multi-batch time variables is called a time-variation matrix, denoted as L p ⁇ N , and p is the number of batches.
  • Step 5 The batch quality or yield indicator corresponding to step 4 is referred to as the index variable Y p ⁇ 1 .
  • the value of the indicator variable is a continuous variable formed by the final mass or yield of the p batches.
  • Step 6 Calculate the covariance matrix S LL and the joint covariance matrix S LY according to the period variable matrix L p ⁇ N and the index variable Y p ⁇ 1 formed in steps 4 and 5.
  • Step 8 Classify the PLS coefficient variable elements in step 7 by symbol size, and define the action symbols as follows:
  • e is the threshold definition for noise and sign(i) is the PLS coefficient symbol corresponding to the ith period.
  • Step 9 Calculate the mean and standard deviation of the variables for each time period.
  • the optimization strategy for each period variable of the collected batch data is established according to the following perturbation calculation formula:
  • J i , M i and ⁇ i are the optimization target values, mean and standard deviation of the i-th period variable, respectively.
  • the self-learning algorithm and calculation steps based on the new batch data are as follows:
  • Step 12 Perform a principal component analysis on the covariance matrix S LL (k+1) and the joint covariance matrix S LY (k+1) and obtain a PLS coefficient vector F i (k+1).
  • Step 13 Calculate the mean and standard deviation of the variables for each period under the new batch data using the following recursive formula:
  • M i (k+1) M i (k)+[C i (k+1)-M i (k)]/(k+1)
  • ⁇ i (k+1) ⁇ i (k)+[C i (k+1) ⁇ M i (k)] ⁇ [C i (k+1) ⁇ M i (k+1)]
  • Step 14 Calculate the optimization target value of each time period variable under the new batch data according to the perturbation formula in step 9.
  • the optimization variable curve established in the above step fifteenth generally needs to be digitally filtered, so that the new optimization curve is relatively smooth and easy to track control.
  • the invention utilizes the perturbation method to establish an initial optimization strategy for optimizing the variable setting curve.
  • the self-learning iterative update of the mean and standard deviation based on the statistical difference of data, to achieve continuous improvement of the optimization index provides a new method for solving the intermittent process optimization strategy of practical industrial problems.
  • the invention is based entirely on operational data of the production process and does not require prior knowledge and mechanism models of the process mechanism.
  • Figure 1 is a graph of the temperature operation of a batch process.
  • Figure 2 is a principal mode diagram of the batch process temperature as an optimized variable.
  • FIG. 3 is an exemplary diagram of a time period variable.
  • FIG. 4 is a diagram showing an example of a PLS coefficient of a period variable.
  • Figure 5 is a calculation diagram of the optimization target variable curve.
  • Figure 6 is a block diagram of the recursive calculation process of the present invention.
  • Figure 7 is a comparison of the optimized temperature profile and the original temperature profile.
  • Figure 8 is an optimization graph before and after filtering.
  • Figure 9 is a comparison of the original operation results of the self-learning optimization strategy and the initial optimization strategy.
  • the method is divided into three parts.
  • the first part is data collection and preprocessing.
  • the second part is to calculate the initial optimization strategy.
  • the third part is to calculate the recursive optimization strategy based on the updated batch data.
  • Step 1 For the complete batch crystallization process, select the temperature operation curve closely related to the product yield as the variable to be optimized, and collect 35 sets of temperature variables and final yield indicator data by batch. The data collection interval is 1 minute.
  • Figure 1 is an example (partial) of a temperature profile of a batch crystallization process.
  • Step 2 For the temperature data of all the collected batches, perform the principal component analysis according to the batch temperature variable and eliminate the singular points in the pivot mode, so that all data points are within a certain confidence level.
  • Figure 2 is a batch process with temperature as the principal mode diagram of the optimized variable. It can be seen from the figure that there is a batch mode in the lower right corner and all other points have a large distance, and the batch data should be eliminated.
  • Step 3 The temperature data of the remaining 34 batches are equally divided into 221 time intervals on the time axis to form period variables C 1 , C 2 ... C 221 .
  • Figure 3 shows the period variables for C 40 to C 70 .
  • Step 4 The multi-batch period variable is used to construct the period variable matrix L 34 ⁇ 221 , and the index variable Y 34 ⁇ 1 .
  • Step 5 Calculate the covariance matrix S LL and the joint covariance matrix S LY using the period variable matrix L and the index variable Y formed in step 4, respectively.
  • Step 6 Perform a principal component analysis on the covariance matrix S LL and the joint covariance matrix S LY and obtain a PLS coefficient vector F 221 .
  • Figure 4 is a plot of PLS coefficients with 221 time-varying variables.
  • Step 7 For the PLS coefficient variable element in step 6, define the following action symbols:
  • Step 8 Calculate the mean and standard deviation of each period variable, and calculate the optimization target value of each period variable of the collected batch data according to the following perturbation formula:
  • J i , M i and ⁇ i are the optimization target values, mean and standard deviation of the i-th period variable, respectively.
  • Figure 5 is an example of establishing an optimized target variable curve.
  • the optimization strategy established by the above batch data is used as the initial value of the recursive learning of the new batch data.
  • the covariance matrix S LL (k+1) and the joint covariance matrix S LY (k+1) are updated using the following recursive self-learning regression formula:
  • Step 11 Perform a principal component analysis on the covariance matrix S LL (k+1) and the joint covariance matrix S LY (k+1) and obtain a PLS coefficient vector F i (k+1).
  • Step 12 Calculate the mean and standard deviation of the variables for each time period under the new batch data using the following recursive self-learning formula:
  • M i (k+1) M i (k)+[C i (k+1)-M i (k)]/(k+1)
  • ⁇ i (k+1) ⁇ i (k)+[C i (k+1) ⁇ M i (k)] ⁇ [C i (k+1) ⁇ M i (k+1)]
  • the optimization variable curve established in the above step 14 generally needs to be digitally filtered, so that the new optimization curve is smooth and easy to track control.
  • Figure 7 is a comparison of the optimized curve and the original curve.
  • Fig. 8 is an example of an optimization curve before and after the moving average filtering.
  • Figure 9 is an illustration of an optimization result for a batch crystallization process. 35 sets of batch production data were used as initial optimized drive data, and 15 sets of batch data were used as recursive update data. The yield increased from 90.80% before the optimization strategy was applied to 92.81% of the initial optimization strategy. Finally, the self-learning optimization strategy increased the yield to 95.51%.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Feedback Control In General (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种由数据差异驱动的间歇过程自学习动态优化方法，包括离线采集生产过程数据，PCA运算剔除奇异批次，构造时段和指标方差矩阵进行PLS运算生成初始优化策略，采集新批次数据，运行递归算法，更新优化策略等步骤。利用摄动法建立优化变量设定曲线的初始优化策略，在此基础之上，基于数据统计差异量对均值和标准差进行自学习迭代更新，实现优化指标的持续改进，为解决实际工业问题的间歇过程优化策略提供了新的方法。该优化方法完全基于生产过程的操作数据，不需要过程机理的先验知识和机理模型。适用于包括间歇反应器、间歇精馏塔、间歇干燥、间歇发酵，间歇结晶及其它采用间歇方式操作的过程和系统的操作轨线动态优化。

Description

一种由数据差异驱动的间歇过程自学习动态优化方法 技术领域
本发明属于化工流程制造业领域，是一种涉及由数据差异驱动的无模型间歇过程自学习动态优化方法。本方法适用于包括间歇反应器、间歇精馏塔、间歇干燥、间歇发酵，间歇结晶以及其它采用间歇方式操作的过程和系统的操作轨线动态优化。
背景技术
间歇过程被广泛应用于食品、聚合物、药品等多种产品的生产和制备上，在化工生产以及流程工业中占有重要的地位。伴随着计算机技术、过程控制和优化技术的持续发展，间歇过程的质量控制与优化成为了当前工业界和学术界研究的热点之一，对间歇过程工业的发展有着重要意义。
目前大多数间歇过程的质量控制与优化问题，都采用基于机理模型的优化方法。它具有较高的优化效率，可以用于在线控制与优化，然而它必须依赖于复杂的过程模型进行。间歇过程的柔性决定了加工产品随时可能发生改变，且不具备辨识模型所需的大量实验和时间条件。同时间歇过程的非线性、不确定性及干扰都较连续过程更为严重，这给建立间歇过程精准可靠的机理模型带来了很多问题。
基于数据驱动的优化方法无需过程的先验知识，具有明显优势。但对间歇生产过程进行操作优化时，如果仅进行一次优化，优化性能具有局限性。因此如何利用间歇过程的重复特性，根据历史批次的信息，通过迭代算法更新当前批次优化变量的轨线，从而持续改进质量指标，提高生产效率，成为间歇过程优化领域亟待解决的难点和焦点。
发明内容
本发明涉及一种针对间歇过程的由数据差异驱动的无模型自学习动态优化方法，利用摄动法建立优化变量设定曲线的初始优化策略。
本发明为实现上述目的，采用如下技术方案：
一种针对间歇过程的基于数据差异驱动的无模型自学习动态优化方法，完全基于生产过程的操作数据，不需要过程机理的先验知识和机理模型。
本方法包括两个部分。第一部分是初始数据收集和初始优化策略的建立。第二部分是建立在新批次数据基础上的自学习算法和计算步骤。
初始数据收集和初始优化策略的建立步骤如下：
步骤一：针对操作完整的间歇过程，按批次收集待优化变量和最终质量或产率指标 数据。数据的收集时间间隔可以是等时间间隔或非等时间间隔，在一个时间间隔内，过程的待优化变量没有显著变化，或它的变化不会对最终质量或产率指标有显著影响。一般要求20-30组有效数据。
步骤二：对所采集的数据，按批次为变量进行主元分析并在主元模式图中剔除奇异点，使得所有数据点在一个可信度之内。
步骤三：将剔除奇异点后的操作数据在时间轴上等间隔划分或不等间隔划分成N段。
步骤四：将每一个间隔所包含的各批次数据表达为一个连续变量Ci,i＝1,2,…N，这些变量被称为分解后的时段变量。时段变量的值由待优化变量在一个特定时间区间的各个批次数据所组成。由多批次时段变量构成的数据矩阵称为时段变量矩阵，记为Lp×N，p是批次数。
步骤五：将步骤四中所对应的每个批次质量或收率指标，称为指标变量Yp×1。指标变量的值是由p个批次最终质量或收率形成的连续变量。
步骤六：根据步骤四和步骤五中形成的时段变量矩阵Lp×N和指标变量Yp×1，分别计算协方差矩阵SLL和联合协方差矩阵SLY。
步骤七：对协方差矩阵SLL和联合协方差矩阵SLY做主元分析并得到PLS系数向量Fi,i＝1,2,…N。
步骤八：对步骤七中的PLS系数变量元素，按符号大小进行分类，定义作用符号如下：
Figure PCTCN2015096377-appb-000001
其中e是对噪声的阈值限定，sign(i)是第i个时段对应的PLS系数符号。
步骤九：计算出各个时段变量的均值和标准差。按以下摄动量计算公式建立已采集批次数据各时段变量的优化策略：
Ji＝Mi+signi×3σi
此处的Ji，Mi和σi分别是第i个时段变量的优化目标值，均值和标准差。
步骤十：将步骤九中所得各时段的优化目标值，按时段顺序i＝1,2,…N组合成一条针对整个批次过程的新优化变量曲线。
上述初始优化策略建立后，将作为自学习算法的初始值。建立在新批次数据基础上的自学习算法和计算步骤如下：
步骤十一：收集新批次的时段变量Ci(k+1),i＝1,2,…N和指标变量数据Y(k+1)，利用以下递归公式，计算更新协方差矩阵SLL(k+1)和联合协方差矩阵SLY(k+1)：
SLL(k+1)＝λSLL(k)+C(k+1)TC(k+1)
SLY(k+1)＝λSLY(k)+C(k+1)TY(k+1)
其中C(k+1)＝[C1(k+1),C2(k+1) … CN(k+1)]，0<λ<1是对现有协方差矩阵的遗忘因子。当λ＝1时，代表没有数据从旧的协方差矩阵中去除。
步骤十二：对协方差矩阵SLL(k+1)和联合协方差矩阵SLY(k+1)做主元分析并得到PLS系数向量Fi(k+1)。
步骤十三：利用如下递归公式，计算新批次数据下各时段变量的均值和标准差：
Mi(k+1)＝Mi(k)+[Ci(k+1)-Mi(k)]/(k+1)
σi(k+1)＝σi(k)+[Ci(k+1)-Mi(k)]×[Ci(k+1)-Mi(k+1)]
步骤十四：按步骤九中的摄动量公式计算新批次数据下各时段变量的优化目标值。
步骤十五：将步骤十四中所得各时段的优化目标值，按时段顺序i＝1,2,…N结合构成一个新的优化变量曲线。
判断是否有新数据更新。有新数据转步骤十一，继续做自学习更新运算。否则结束学习过程。
在以上步骤十五中所建立的优化变量曲线，一般地需要其进行数字滤波，使得新的优化曲线比较光滑，易于跟踪控制。
本发明利用摄动法建立优化变量设定曲线的初始优化策略。在此基础之上，基于数据统计差异量对均值和标准差进行自学习迭代更新，实现优化指标的持续改进，为解决实际工业问题的间歇过程优化策略提供了新的方法。本发明完全基于生产过程的操作数据，不需要过程机理的先验知识和机理模型。
附图说明
图1为一个间歇过程的温度操作曲线图。
图2为间歇过程温度为优化变量的主元模式图。
图3为时段变量的示例图。
图4为时段变量的PLS系数示例图。
图5为优化目标变量曲线的计算图。
图6为本发明递归计算过程框图。
图7为优化的温度曲线和原始温度曲线的比较图。
图8为滤波前后的优化曲线图。
图9为自学习优化策略和初始优化策略对原始操作结果比较图。
具体实施方式
本例以一个间歇结晶过程为例，所述方法不构成本发明的范围限制。
本方法分为三个部分。第一部分为数据收集和预处理。第二部分是计算初始优化策略。第三部分是在取得更新批次数据的基础上，进行递归优化策略的计算。
本方法实施步骤框图如图6所示，具体实施步骤和算法如下：
步骤1：针对操作完整的间歇结晶过程，选择与产品收率密切相关的温度操作曲线作为待优化变量，按批次收集35组温度变量以及最终产率指标数据。数据的收集时间间隔为1分钟。图1是一个间歇结晶过程的温度曲线示例(局部)。
步骤2：对采集的所有批次的温度数据，按批次温度变量进行主元分析并在主元模式图中剔除奇异点，使得所有数据点在一个可信度之内。图2是一个间歇过程，温度为优化变量的主元模式图，从图中可以看出，右下角有一个批次的模式和其它所有点都有很大距离，应剔除这个批次数据。
步骤3：将剩余34个批次的温度数据在时间轴上等间隔划分为221个时段，形成时段变量C1,C2…C221。为清晰起见，图3给出了C40到C70的时段变量。
步骤4：用多批次时段变量构成时段变量矩阵L34×221，和指标变量Y34×1。
步骤5：用步骤4中形成的时段变量矩阵L和指标变量Y，分别计算协方差矩阵SLL和联合协方差矩阵SLY。
步骤6：对协方差矩阵SLL和联合协方差矩阵SLY做主元分析并得到PLS系数向量 F221。图4是一个有221个时段变量的PLS系数图。
步骤7：对步骤6中的PLS系数变量元素，定义如下作用符号：
Figure PCTCN2015096377-appb-000002
其中e是对噪声的阈值限定，此处取为e＝0.01。
步骤8：计算出各个时段变量的均值和标准差，并按以下摄动量公式计算已采集批次数据各时段变量的优化目标值：
Ji＝Mi+signi×3σi
此处的Ji，Mi和σi分别是第i个时段变量的优化目标值，均值和标准差。
步骤9：将步骤8中所得各时段的优化目标值，按时段顺序i＝1,2,…221构成一个优化变量曲线。图5是一个建立优化目标变量曲线的示例。如图5所示，当i＝40时，均值M40＝159.687，标准差σ40＝0.577，F40＝0.000149，sign40＝0，所得优化策略为J40＝159.687；当i＝45时，均值M45＝170.26，标准差σ45＝0.416，F45＝-0.046091，sign45＝-1，所得优化策略为J45＝169.020。
将以上批次数据建立的优化策略，作为新批次数据递归学习的初始值。
建立在新批次数据基础上的递归自学习算法和计算步骤如下：
步骤10：收集新一批次数据的时段变量Ci(k+1),i＝1,2,…N和指标变量数据Y(k+1)，此处k＝35。
利用以下递归自学习归公式，对协方差矩阵SLL(k+1)和联合协方差矩阵SLY(k+1))进行更新：
SLL(k+1)＝λSLL(k)+C(k+1)TC(k+1)
SLY(k+1)＝λSLY(k)+C(k+1)TY(k+1)
其中k＝35,36,…，C(k+1)＝[C1(k+1),C2(k+1) … CN(k+1)]，0<λ<1是对现有协方差矩阵的遗忘因子。当λ＝1时，代表没有数据从旧的协方差矩阵中去除。本例中为保留所有批次信息，取λ＝1。
步骤11：对协方差矩阵SLL(k+1)和联合协方差矩阵SLY(k+1)做主元分析并得到PLS系数向量Fi(k+1)。
步骤12：利用以下递归自学习公式，计算新批次数据下各时段变量的均值和标准差：
Mi(k+1)＝Mi(k)+[Ci(k+1)-Mi(k)]/(k+1)
σi(k+1)＝σi(k)+[Ci(k+1)-Mi(k)]×[Ci(k+1)-Mi(k+1)]
在本例中，当i＝40时，原始35批数据的均值M40(35)＝159.687，标准差σ40(35)＝0.577，新一批数据更新后的均值M40(36)＝159.772，标准差σ40(36)＝0.585；当i＝45时，原始35批数据的均值M45(35)＝170.26，标准差σ45(35)＝0.416，新一批数据更新后的均值M45(36)＝170.336，标准差σ45(36)＝0.510。
步骤13：根据上述步骤计算所得更新后的时段变量的均值和标准差，按照步骤8中的摄动量计算公式获取各时段变量的优化目标值。如i＝40时，均值M40＝159.772，标准差σ40＝0.585，F40＝-0.008856，所得优化策略为J40＝159.772；当i＝45时，均值M45＝170.336，标准差σ45＝0.510，F40＝-0.04199，所得优化策略为J40＝168.832。
步骤14：将步骤13中所得各时段的优化目标值，按时段顺序i＝1,2,…221构成一个更新的优化变量曲线。
判断是否有新数据更新。有新数据转步骤10，做自学习递归运算，否则结束计算过程。本例中，有15批新数据，此递归算法连续计算15次后，即k＝49，终止递归计算。
在以上步骤14中所建立的优化变量曲线，一般地需要将其进行数字滤波，使得新的优化曲线比较光滑，易于跟踪控制。图7是优化曲线和原始曲线的比较。图8是滑动平均滤波前后的优化曲线的示例。
为说明本发明方法的有效性，图9是一个间歇结晶过程的优化结果示例。35组批次生产数据作为初始优化的驱动数据，15组批次数据作为递归更新数据。收率从未施加优化策略前的90.80％增加到初始优化策略92.81％，最终经过自学习优化策略，收率增长到95.51％。

Claims (4)

  1. 一种由数据差异驱动的间歇过程自学习动态优化方法，其特征在于包括下述步骤：
    (1)针对操作完整的间歇过程，按批次收集待优化变量和最终质量或产率指标；
    (2)对所述步骤(1)中采集的数据，以批次为变量进行主元分析并在主元模式图中剔除奇异点，使得所有数据点在一个可信度之内；
    (3)将剔除奇异点后的操作数据在时间轴上间隔划分成N段；将每一个时间间隔所包含的各批次数据表达为一个连续变量Ci,i＝1,2,…N，这些变量被称为时段变量；所述时段变量的值由待优化变量在一个特定时间区间的各个批次数据所组成；由多批次时段变量构成的数据矩阵称为时段变量矩阵，记为Lp×N，p是批次数；
    (4)将步骤(3)中所对应的每个批次质量或收率指标，称为指标变量Yp×1；所述指标变量的值是由p个批次最终质量或收率形成的连续变量；
    (5)根据步骤(3)和步骤(4)中形成的时段变量矩阵Lp×N和指标变量Yp×1，分别计算协方差矩阵SLL和联合协方差矩阵SLY；
    (6)对协方差矩阵SLL和联合协方差矩阵SLY做主元分析并得到PLS系数向量Fi,i＝1,2,…N；
    (7)对步骤(6)中的PLS系数变量元素，按符号大小进行分类，定义作用符号如下：
    Figure PCTCN2015096377-appb-100001
    其中e是对噪声的阈值限定，sign(i)是第i个时段对应的PLS系数符号；
    (8)计算出各个时段变量的均值和标准差；按以下摄动量计算公式建立已采集批次数据各时段变量的初始优化策略：
    Ji＝Mi+signi×3σi
    此处的Ji，Mi和σi分别是第i个时段变量的优化目标值，均值和标准差；
    (9)将步骤(8)中所得各时段的优化目标值，按时段顺序i＝1,2,…N组合成一条针对整个批次过程的初始优化变量曲线；
    (10)收集新批次的时段变量Ci(k+1),i＝1,2,…N和指标变量数据Y(k+1)，利用以下递归公式，计算更新协方差矩阵SLL(k+1)和联合协方差矩阵SLY(k+1)：
    SLL(k+1)＝λSLL(k)+C(k+1)TC(k+1)
    SLY(k+1)＝λSLY(k)+C(k+1)TY(k+1)
    其中C(k+1)＝[C1(k+1),C2(k+1)…CN(k+1)]，0<λ<1是对现有协方差矩阵的遗忘因子；当λ＝1时，代表没有数据从旧的协方差矩阵中去除；
    (11)对协方差矩阵SLL(k+1)和联合协方差矩阵SLY(k+1)做主元分析并得到PLS系数向量Fi(k+1)；
    (12)利用如下递归公式，计算新批次数据下各时段变量的均值和标准差：
    Mi(k+1)＝Mi(k)+[Ci(k+1)-Mi(k)]/(k+1)
    σi(k+1)＝σi(k)+[Ci(k+1)-Mi(k)]×[Ci(k+1)-Mi(k+1)]；
    (13)按步骤(8)中的摄动量公式计算新批次数据下各时段变量的优化目标值；
    (14)将步骤(13)中所得各时段的优化目标值，按时段顺序i＝1,2,…N结合构成一个新的优化变量曲线；
    (15)判断是否有新数据更新，有新数据转步骤(10)，继续做自学习更新运算，否则结束学习过程。
  2. 根据权利要求1所述的由数据差异驱动的间歇过程自学习动态优化方法，其特征在于：所述步骤(1)中间歇过程数据收集的时间间隔相等或不相等。
  3. 根据权利要求1所述的由数据差异驱动的间歇过程自学习动态优化方法，其特征在于：所述步骤(3)中间隔划分为等间隔划分或不等间隔划分。
  4. 根据权利要求1所述的由数据差异驱动的间歇过程自学习动态优化方法，其特征在于：所述步骤(14)中所建立的优化变量曲线，对其进行数字滤波，使得新的优化曲线比较光滑，易于跟踪控制。
PCT/CN2015/096377 2015-11-26 2015-12-04 一种由数据差异驱动的间歇过程自学习动态优化方法 WO2017088208A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/566,902 US10254729B2 (en) 2015-11-26 2015-12-04 Data-difference-driven self-learning dynamic optimization method for batch process

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510834482.5 2015-11-26
CN201510834482.5A CN105353607B (zh) 2015-11-26 2015-11-26 一种由数据差异驱动的间歇过程自学习动态优化方法

Publications (1)

Publication Number Publication Date
WO2017088208A1 true WO2017088208A1 (zh) 2017-06-01

Family

ID=55329596

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/096377 WO2017088208A1 (zh) 2015-11-26 2015-12-04 一种由数据差异驱动的间歇过程自学习动态优化方法

Country Status (3)

Country Link
US (1) US10254729B2 (zh)
CN (1) CN105353607B (zh)
WO (1) WO2017088208A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114020861A (zh) * 2021-10-14 2022-02-08 同济大学 基于调度知识自学习更新的智能车间生产控制方法及设备
CN114609981A (zh) * 2021-12-16 2022-06-10 南京工业大学 一种基于参数区间变化趋势的动态操作模式优化方法、系统及储存介质
CN115327903A (zh) * 2022-08-11 2022-11-11 辽宁石油化工大学 二维状态时滞批处理过程的离轨策略最优跟踪控制方法

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3514741A1 (en) * 2018-01-19 2019-07-24 Siemens Aktiengesellschaft A method and apparatus for dynamically optimizing industrial production processes
CN109101758A (zh) * 2018-09-03 2018-12-28 江南大学 基于t-pls模型的间歇过程工艺条件设计方法
CN110632908B (zh) * 2019-10-18 2022-06-14 太原理工大学 基于最小熵控制的化学批次反应器系统控制性能评价方法
US20220138614A1 (en) * 2020-10-30 2022-05-05 International Business Machines Corporation Explaining machine learning based time series models
CN112925202B (zh) * 2021-01-19 2022-10-11 北京工业大学 基于动态特征提取的发酵过程阶段划分方法
CN113674057B (zh) * 2021-08-19 2023-08-25 广东工业大学 一种基于分层聚合策略的组批优化方法
CN113803647B (zh) * 2021-08-25 2023-07-04 浙江工业大学 一种基于知识特征与混合模型融合的管道泄漏检测方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060151688A1 (en) * 2003-05-29 2006-07-13 Waters Investments Limited System and method for metabonomics directed processing of LC-MS or LC-MS/MS data
CN101872444A (zh) * 2010-05-21 2010-10-27 杭州电子科技大学 一种结合中期修正策略的间歇过程批到批优化方法
CN103092078A (zh) * 2013-01-07 2013-05-08 北京中医药大学 多阶段间歇生产过程的全程优化方法

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1417629B1 (en) * 2001-07-06 2008-02-20 Lipomics Technologies, Inc. Generating, viewing, interpreting, and utilizing a quantitative database of metabolites
US6873915B2 (en) * 2001-08-24 2005-03-29 Surromed, Inc. Peak selection in multidimensional data
US7457708B2 (en) * 2003-03-13 2008-11-25 Agilent Technologies Inc Methods and devices for identifying related ions from chromatographic mass spectral datasets containing overlapping components
US7949475B2 (en) * 2005-08-08 2011-05-24 Metabolon Inc. System and method for analyzing metabolomic data
CN103116306B (zh) * 2013-02-05 2015-06-17 浙江大学 一种自动的步进式有序时段划分方法
CN103279123B (zh) * 2013-05-21 2015-12-23 沈阳化工大学 对间歇控制系统进行分段故障监视的方法
CN103336507B (zh) * 2013-06-24 2015-08-19 浙江大学 基于多模态协同时段自动划分的统计建模与在线监测方法
CN103853152B (zh) * 2014-03-21 2016-08-17 北京工业大学 一种基于ar-pca的间歇过程故障监测方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060151688A1 (en) * 2003-05-29 2006-07-13 Waters Investments Limited System and method for metabonomics directed processing of LC-MS or LC-MS/MS data
CN101872444A (zh) * 2010-05-21 2010-10-27 杭州电子科技大学 一种结合中期修正策略的间歇过程批到批优化方法
CN103092078A (zh) * 2013-01-07 2013-05-08 北京中医药大学 多阶段间歇生产过程的全程优化方法

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
CAMACHO, JOSE: "Multi-phase principal component analysis for batch processes modelling", CHEMOMETRICS AND INTELLIGENT LABORATORY SYSTEMS, vol. 81, no. 2, 15 April 2006 (2006-04-15), pages 127 - 136, XP005351930, ISSN: 0169-7439, DOI: 10.1016/j.chemolab.2005.11.003 *
CHEN, JUNGHUI ET AL.: "On-line batch process monitoring using dynamic PCA and dynamic PLS models", CHEMICAL ENGINEERING SCIENCE, vol. 57, no. 1, 31 January 2002 (2002-01-31), pages 63 - 75, XP055598203, ISSN: 0009-2509, DOI: 10.1016/S0009-2509(01)00366-9 *
YE, LINGJIAN ET AL.: "A Real-Time Optimization Approach for Uncertain Batch Process", JOURNAL OF CHEMICAL INDUSTRY AND ENGINEERING (CHINA), vol. 65, no. 9, 30 September 2014 (2014-09-30), pages 3535 - 3543 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114020861A (zh) * 2021-10-14 2022-02-08 同济大学 基于调度知识自学习更新的智能车间生产控制方法及设备
CN114609981A (zh) * 2021-12-16 2022-06-10 南京工业大学 一种基于参数区间变化趋势的动态操作模式优化方法、系统及储存介质
CN114609981B (zh) * 2021-12-16 2024-04-16 南京工业大学 一种基于参数区间变化趋势的动态操作模式优化方法、系统及储存介质
CN115327903A (zh) * 2022-08-11 2022-11-11 辽宁石油化工大学 二维状态时滞批处理过程的离轨策略最优跟踪控制方法

Also Published As

Publication number Publication date
US10254729B2 (en) 2019-04-09
CN105353607B (zh) 2017-10-27
US20180259921A1 (en) 2018-09-13
CN105353607A (zh) 2016-02-24

Similar Documents

Publication Publication Date Title
WO2017088208A1 (zh) 一种由数据差异驱动的间歇过程自学习动态优化方法
CN111401573B (zh) 一种工况状态建模与修正模型方法
CN105574587B (zh) 一种塑料注塑过程的在线工况过程监控方法
CN104914723B (zh) 基于协同训练偏最小二乘模型的工业过程软测量建模方法
CN109308063B (zh) 基于得分矩阵的发酵过程阶段划分方法
Lv et al. Batch process monitoring based on just-in-time learning and multiple-subspace principal component analysis
CN109858572A (zh) 一种用于污水异常检测的改进型层次聚类方法
CN111340110B (zh) 一种基于工业过程运行状态趋势分析的故障预警方法
CN107403196B (zh) 基于谱聚类分析的即时学习建模对脱丁烷塔底丁烷浓度进行预测的方法
US20190094838A1 (en) Model-Free Online Recursive Optimization Method for Batch Process Based on Variable Period Decomposition
CN109634233A (zh) 工业大数据智能分析决策方法、可读存储介质和终端
CN103488561A (zh) 一种在线升级主样本模型的kNN故障检测方法
CN108459579B (zh) 基于时间序列模型系数的半导体批次过程故障诊断方法
CN107808209B (zh) 基于加权kNN距离的风电场异常数据辨识方法
WO2024139313A1 (zh) 多工况工业系统自适应稀疏辨识和控制方法、系统和设备
CN109614451A (zh) 工业大数据智能分析决策装置
CN118211082A (zh) 基于数据分析的油液滤芯剩余寿命预测方法及系统
CN114118401A (zh) 基于神经网络的配电网络流量预测方法、系统、装置及存储介质
CN105159247B (zh) 基于信息传递的间歇过程实时采样点阶段归属判断方法
CN113609672A (zh) 一种基于增量模型的数字孪生系统协同进化方法
CN117669373A (zh) 一种锻造成形装备液压系统能耗预测方法及系统
CN117571312A (zh) 面向噪声标签工业场景的旋转机械故障诊断方法
CN114360634A (zh) 基于时序大数据及转移熵的基因表达调控网络构建方法
CN114967624A (zh) 一种基于邻域差异特征分析提取的批次生产过程监测方法
CN103439964A (zh) 一种ocsvm监控模型的在线更新系统及方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15909119

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15566902

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15909119

Country of ref document: EP

Kind code of ref document: A1