Model-Free Online Recursive Optimization Method for Batch Process Based on Variable Period Decomposition

Abstract

The present invention discloses a model-free online recursive optimization method for a batch process based on variable period decomposition. Variable operation data closely related to product quality is acquired, optimization action on each subset is integrated on the basis of time domain variable division on the process by utilizing a data driving method and a global optimization strategy is formed, based on which an online recursive error correction optimization strategy is implemented. According to the method, the online optimization strategy is formed completely based on the operation data of the batch process without needing prior knowledge or a model of a process mechanism. Meanwhile, the optimized operation locus line has better adaptability by using the online recursive correction strategy, and thus the anti-interference requirement of the actual industrial production is better met.

Claims

1. A model-free online recursive optimization method for a batch process based on variable period decomposition, characterized by comprising the following steps: (1) for operating a complete batch process, acquiring variables to be optimized and final quality or yield indicators in batches; (2) for the data acquired in step (1), performing principal component analysis on the variables in batches, and removing singular points from a principal component mode diagram, so as to enable all data points to be within one degree of credibility; (3) performing interval division on the remaining data after the singular points are removed on a time axis; expressing each batch of data included in each interval as a continuous variable, wherein these variables are referred to as decomposed period variables, and a value of the period variable is composed of each batch of data of the variable to be optimized in a specific time interval; (4) referring to each corresponding batch quality or yield indicator in step (3) as an indicator variable, wherein a value of the indicator variable is a continuous variable formed by the quality or yield of each batch; (5) combining the period variables and the indicator variables formed in step (3) and step (4) to form a combined data matrix of the period variables and the indicator variables, and performing principal component analysis on the combined data matrix to form a principal component load diagram; (6) classifying the action directions and magnitudes of the period variables on the indicator variables for the principal component load diagram in step (5); (7) calculating an optimization strategy for each period variable according to the following perturbation formula:
J(i)=(i)+sign(i)3(i) wherein J(i), M(i) and (i) herein are respectively optimization target value, mean value and standard deviation of the ith period variable; and sign(i) is a cosine symbol of an included angle formed by the ith period variable and the indicator variable; (8) constituting a basic optimization variable curve for the whole batch process by using the optimization target values of all periods obtained in step (7) according to a period sequence; (9) in the (i1)th time period, calculating an error of an offline basic optimization target value J(i1) and an actual measured value RV(i1):
E(i1)=J(i1)RV(i1); (10) on the offline basic optimization strategy, constituting a new optimization target value of next period:
J.sub.o(i)=J(i)+E(i1); and (11) sequentially calculating step (9) and step (10) according to the period sequence i=1, 2, . . . , N and applying them to the process, till the operation of the whole batch process is over.

2. The model-free online recursive optimization method for the batch process based on variable period decomposition according to claim 1, characterized in that the time intervals of batch process data acquisition in step (1) are equal or unequal.

3. The model-free online recursive optimization method for the batch process based on variable period decomposition according to claim 1, characterized in that the interval division in step (3) is equal interval division or unequal interval division.

4. The model-free online recursive optimization method for the batch process based on variable period decomposition according to claim 1, characterized in that the classification in step (6) comprises positive action, reverse action and no/micro action.

5. The model-free online recursive optimization method for the batch process based on variable period decomposition according to claim 1, characterized in that the value of the included angle cosine symbol sign(i) is +1 when the included angle is smaller than 90 degrees, 1 when the included angle is greater than 90 degrees, or 0 when the included angle is equal to 90 degrees.

6. The model-free online recursive optimization method for the batch process based on variable period decomposition according to claim 1, characterized in that the optimization variable curve is digitally filtered in step (8), so that the new optimization variable curve is smooth.

Description

BRIEF DESCRIPTION OF FIGURES

[0029] FIG. 1 is a temperature curve example of a batch process.

[0030] FIG. 2 is a principal component mode diagram indicating that the temperature of a batch process is an optimization variable.

[0031] FIG. 3 is a composition diagram of period variables.

[0032] FIG. 4 is a principal component load diagram of period variables and indicator variables.

[0033] FIG. 5 is an action classification diagram of the period variables on the indicator variables.

[0034] FIG. 6 is a comparison diagram of an optimized temperature curve and an original temperature curve of a batch process.

[0035] FIG. 7 is a generation diagram of an online recursive error correction strategy.

[0036] FIG. 8 is a block diagram of implementation steps of the present invention.

[0037] FIG. 9 shows an optimized curve after moving average filtering and an original optimized curve.

[0038] FIG. 10 is an optimization result (partial) diagram of a batch crystallization process.

DETAILED DESCRIPTION

[0039] A batch crystallization process is taken as the example, and the method does not limit the scope of the present invention.

[0040] This implementation method is divided into four parts. The first part is data acquisition and preprocessing. The second part is construction of a combined data matrix. The third part is calculation of a basic optimization strategy. The fourth part is establishment of a recursive error correction online optimization strategy.

[0041] The block diagram of the implementation steps of the present method is shown as FIG. 8, and the specific implementation steps and algorithms are as follows.

[0042] Step 1: For operating a complete batch crystallization process, operation temperature closely related to product yield is selected as a variable to be optimized, and 50 groups of temperature variables and final yield indicator data are acquired in batches. The acquisition time interval of the data is 1 minute. FIG. 1 is a temperature curve data acquisition example of a batch crystallization process, and for the sake of clarity, only the temperature curves of 2 batches are drawn in the figure.

[0043] Step 2: For all the acquired 50 batches of temperature data, principal component analysis is performed on the temperature variables in batches, and singular points are removed from a principal component mode diagram, so that all data points are within one degree of credibility. FIG. 2 is a principal component mode diagram indicating that the temperature of a batch process is an optimization variable, and it can be seen from the figure that a batch of temperature data on the right is greatly different from the overall data mode and thus should be removed.

[0044] Step 3: The remaining 49 batches of temperature data are divided into 300 periods at equal intervals on a time axis to constitute 300 period variables C1, C2, . . . , C300. For the sake of clarity, FIG. 3 shows C40 to C70 period variables.

[0045] Step 4: Each corresponding batch of yield indicator data in step 3 forms an indicator variable Q.

[0046] Step 5: The 300 period variables C1, C2, . . . , C300 and one indicator variable Q formed in step 3 and step 4 are combined to generate a 49301-dimensional combined data matrix L.

[0047] Step 6: Principal component analysis is performed on the combined matrix L to form a principal component load diagram. For the sake of clarity, FIG. 4 shows a principal component load diagram example generated by combining 25 period variables C36 to C60 and the indicator variable Q.

[0048] Step 7: The action directions and magnitudes of the period variables on the indicator variable are classified for the principal component load diagram in step 6. FIG. 5 is a classification example, it can be seen from FIG. 5 that the actions of C154, C155, C156 and C273 on the indicator variable Q are maximum, wherein C154, C155 and C156 are reverse actions, and C273 is a positive action. C66, C111 and the like having an included angle of about 90 degrees with the indicator variable Q in the directions nearly do not act on the indicator variable Q.

[0049] Step 8: Mean value and standard deviation of each period variable are calculated respectively. For example, the mean value of C154 having a reverse action on the indicator variable Q is 134.58 DEG C, and the standard deviation is 6.08 DEG C.

[0050] Step 9: The optimization target value of the ith period variable is acquired according to the following perturbation calculation formula:

J(i)=M(i)+sign(i)3(i)

[0051] wherein J(i), M(i) and (i) herein are respectively optimization target value, mean value and standard deviation of the ith period variable; and sign(i) is a cosine symbol of an included angle formed by the ith period variable and the indicator variable. On the classification diagram of FIG. 5, the sign(i) is +1 when the included angle is smaller than 90 degrees, 1 when the included angle is greater than 90 degrees, and 0 when the included angle is equal to 90 degrees.

[0052] Step 10: The optimization target values of all periods obtained in step 9 constitute a basic optimization variable curve according to a period sequence i=1, 2, . . . , 300.

[0053] Step 11: Moving average filtering is performed on the basic optimization curve, so that the filtered optimization curve is relatively smooth and facilitates later tracking control design. FIG. 6 is a comparison of the optimized temperature curve and the original temperature curve, and FIG. 9 shows an optimization curve after moving average filtering and an original optimization curve. It can be seen from FIG. 9 that the filtered optimization curve is smoother and facilitates implementation of a tracking controller.

[0054] Step 12: When the basic optimization control locus obtained by the above series of steps is used on line, recursive error correction is performed in each time period:

[0055] (1) for the (i1)th time period, the error of the offline basic optimization target value J(i1) and the actual measured value RV(i1) is calculated:

E(i1)=J(i1)RV(i=1)

[0056] (2) on the offline basic optimization strategy, a new optimization target value of next period is constituted:

J.sub.o(i)=J(i)+E(i1).

[0057] Step 12 is sequentially calculated according to the period sequence i=1, 2, . . . , 300, till the operation of the whole batch process is over. FIG. 7 is a generation calculation schematic diagram of an online recursive error correction strategy.

[0058] FIG. 10 is an optimization result example of a batch crystallization process. It can be seen from the result in the figure that the optimization-free yield is 90.25%, whereas under the recursive correction optimization strategy, the actual operation yield is 94.88% and is substantially close to the theoretical optimal yield 95.51%. This result shows the effectiveness and practicability of the method of the present invention.

[0059] While the present invention has been described in some detail for purposes of clarity and understanding, one skilled in the art will appreciate that various changes in form and detail can be made without departing from the true scope of the invention. All figures, tables, appendices, patents, patent applications and publications, referred to above, are hereby incorporated by reference.