WO2013171862A1

WO2013171862A1 - Setting calculation system learning device and learning method

Info

Publication number: WO2013171862A1
Application number: PCT/JP2012/062535
Authority: WO
Inventors: 治樹井波
Original assignee: 東芝三菱電機産業システム株式会社
Priority date: 2012-05-16
Filing date: 2012-05-16
Publication date: 2013-11-21
Also published as: JP5846303B2; CN104303114A; KR20140143225A; KR101622068B1; JPWO2013171862A1; CN104303114B

Abstract

Provided is a setting calculation system learning device with which, even if a portion of actual measured values which are necessary for mathematical model learning can not be obtained, it is possible to appropriately update learning terms of each function which configures the mathematical model. To this end, a learning device, which, using an actual measured value, updates a learning term of a mathematical model with a setting calculation system which, using the mathematical model, calculates a setting value of mechanical equipment, comprises: a learning term calculation unit which respectively calculates, using actual measured values, learning terms of upstream-side and downstream-side functions which configure a mathematical model; an actual measured value determination unit which, in the calculation of the learning term of the downstream-side function, determines whether a first actual measured value which is inputted into the downstream-side function is anomalous; a performance calculation value calculation unit which, when it is determined that the first actual measured value is anomalous, inputs an output from the upstream-side function into the downstream-side function and calculates a performance calculation value which is outputted from the downstream-side function; and a complementary learning calculation unit which allocates to each learning term of the upstream-side and the downstream-side function an error of the performance calculation value with respect to the second actual measured value which corresponds to the output from the downstream-side function.

Description

Learning apparatus and learning method for setting calculation system

The present invention relates to a learning apparatus and learning method for a setting calculation system.

In general, for example, as a method of determining set values for controlling machine equipment in a process line of a rolling plant, a mathematical model that represents and reproduces a physical phenomenon that occurs in an environment including the controlled object by a mathematical formula is constructed. There is known a method for determining a setting value by obtaining a setting value that can achieve the target result.

In determining the set value using the mathematical model, increasing the reproduction accuracy of the target physical phenomenon in the mathematical model to be used leads to a better set value. Thus, in order to improve the accuracy of the mathematical model, it has been conventionally performed to incorporate a learning term in the mathematical model and to correct the mathematical model based on the actual value.

As a learning method of such a mathematical model, the actual value calculated by the mathematical model (actual calculated value) is compared with the actual value obtained from the actual value actually measured with a measuring instrument, etc. A method of updating the learning term of the mathematical model is often adopted so that the difference between the two becomes small.

And in the conventional learning method, the factor that causes the difference between the actual calculation value by the mathematical model and the actual actual value is processed by the process line and the time-series fluctuation of the process line due to the change of equipment and environment. It is known that the calculation value of the mathematical formula model is corrected based on the learning coefficient calculated for each of these two types of variations after roughly dividing into two types of variations due to the types of materials and processing patterns. (For example, refer to Patent Document 1).

Japanese Patent No. 2839746

By the way, when a physical phenomenon to be reproduced by a mathematical model (ie, a physical phenomenon that occurs in a controlled object) is complicated, a mathematical model that reproduces the physical phenomenon by a composite function obtained by synthesizing a plurality of simpler functions. May be configured. In such a case, it is also preferable from the viewpoint of improving the reproducibility of the mathematical model to provide a learning term for each function constituting the composite function and perform learning for each function.

However, when actual measurement values cannot be obtained due to problems such as the environment around the measuring instrument, it may be impossible to obtain actual values necessary for learning for some of the functions constituting the mathematical model. As the synthesis function constituting the mathematical model is synthesized from a larger number of functions, the more types of actual values necessary for learning, the higher the possibility that (partial) actual values will be lost.

Then, the problem that the mathematical model learning cannot be completed due to a deficiency in a part of the actual values necessary for learning, and the prediction accuracy of the mathematical model temporarily decreases is the conventional setting. It exists in a learning apparatus and a learning method of a calculation system. In addition, in the situation immediately after starting learning, if the learning of the mathematical model cannot be completed due to a deficiency in a part of the actual values necessary for learning, the mathematical model is expected to have the expected accuracy. There is also a problem that the time required until the prediction can be performed becomes long.

The present invention has been made to solve such a problem, and even when a part of the actual value necessary for learning the mathematical model is not obtained, the actual value for the obtained amount is used. A learning apparatus and learning method for a setting calculation system capable of appropriately updating learning terms of each function constituting a mathematical model by performing learning and suppressing a decrease in accuracy of the mathematical model Is.

In the learning apparatus of the setting calculation system according to the present invention, in the setting calculation system that calculates the setting value of the mechanical facility using the mathematical model, the learning device updates the learning term of the mathematical model using the actual measurement value. In the calculation of the learning term of the downstream function in the learning term calculation unit, the learning term calculation unit that calculates the learning term of each of the upstream function and the downstream function constituting the mathematical model using the actual measurement value When it is determined that the first actual measurement value input to the downstream function is abnormal, the actual measurement value determination unit that determines whether the first actual measurement value is abnormal, or the first actual measurement value is abnormal , An actual calculation value calculation unit that inputs an output from the upstream function to the downstream function and calculates an actual calculation value output from the downstream function, and the actual measurement corresponding to the output from the downstream function Value The error of the actual calculated values for the second measured value, and a complementary learning calculation section for distributing the learning section learning term and the downstream function of the upstream function.

Further, in the learning method of the setting calculation system according to the present invention, in the setting calculation system for calculating the setting value of the mechanical equipment using the mathematical model, the learning method of updating the learning term of the mathematical model using the actual measurement value. A first step of calculating learning terms of each of the upstream function and the downstream function constituting the mathematical model using the actual measurement value, and a learning term of the downstream function in the first step. A second step of determining whether or not the first actual measurement value that is the actual measurement value input to the downstream function in the calculation is abnormal; and the first actual measurement value is determined to be abnormal A third step of inputting an output from the upstream function to the downstream function and calculating an actual calculation value output from the downstream function, and the output corresponding to the output from the downstream function Fruit Comprising an error of the actual calculated value for the second measured value is a value, a fourth step of distributing the learning section learning term and the downstream function of the upstream functions, the.

In the learning device and the learning method of the setting calculation system according to the present invention, even when a part of the actual value necessary for learning the mathematical model is not obtained, the learning using the actual value for the obtained amount is performed. It is possible to appropriately update the learning term of each function constituting the mathematical model, and it is possible to suppress a decrease in accuracy of the mathematical model.

It is a figure explaining the whole structure of the setting calculation system which concerns on Embodiment 1 of this invention. It is a flowchart explaining operation | movement of the learning apparatus of the setting calculation system which concerns on Embodiment 1 of this invention. It is a figure explaining the detailed structure centering on the learning apparatus with which the setting calculation system which concerns on Embodiment 1 of this invention is provided.

The present invention will be described with reference to the accompanying drawings. Throughout the drawings, the same reference numerals indicate the same or corresponding parts, and redundant description thereof will be simplified or omitted as appropriate.

Embodiment 1 FIG.
FIGS. 1 to 3 relate to Embodiment 1 of the present invention, FIG. 1 is a diagram for explaining the overall configuration of the setting calculation system, and FIG. 2 is a flowchart for explaining the operation of the learning device of the setting calculation system. FIG. 3 is a diagram illustrating a detailed configuration centering on a learning device provided in the setting calculation system.

FIG. 1 shows an overall configuration of a setting calculation system 2 that calculates set values of mechanical equipment 1 constituting a process line of a rolling plant, for example. The setting calculation system 2 includes a setting calculation device 3, an actual measurement value collection device 4, and a learning device 5.

In the setting calculation device 3, a mathematical model in which a physical phenomenon generated by the operation of the machine facility 1 is modeled by a mathematical formula is registered in advance. The setting calculation device 3 uses the mathematical model to simulate the result of the operation of the machine facility 1, thereby calculating the set value of the machine facility 1 so that the operation result of the machine facility 1 is closer to the target value. . The set value calculated by the setting calculation device 3 is output to the machine facility 1. The mechanical equipment 1 operates according to the set value calculated by the setting calculation system 2.

The measured value collection device 4 collects measured values of predetermined types of physical quantities necessary for learning the mathematical model in the learning device 5 among the physical quantities in the machine equipment 1 and the environment in which the machine equipment 1 operates. It is. In the environment in which the mechanical equipment 1 or the mechanical equipment 1 is installed, a measuring instrument or the like for measuring the physical quantity is installed in advance. The actual measurement value collection device 4 collects the actual measurement values of the physical quantities measured by a measuring instrument or the like.

The learning device 5 performs learning of the mathematical model used by the setting calculation device 3 based on the actual measurement values collected by the actual measurement value collection device 4. The learning device 5 includes a mathematical model learning calculation unit 6, an actual value determination unit 7, and a complementary learning calculation unit 8.

The formula model learning calculation unit 6 calculates the actual calculation value using the same formula model as that used in the setting calculation device 3. Here, the actual calculation value is an output from the mathematical model when the actual measurement value collected by the actual measurement value collection device 4 is input. Then, the mathematical model learning calculation unit 6 compares the actual calculation value calculated from the mathematical model with the actual value obtained directly from the actual measurement value collected by the actual measurement value collection device 4, and the difference between these values is reduced. The learning term of the mathematical model is calculated as follows.

The actual measurement value determination unit 7 determines whether the actual measurement value collected by the actual measurement value collection device 4 is normal or abnormal. Here, the state where the actual measurement value is abnormal is a state where the actual measurement value deviates from the normal range (this range is determined in advance for each measurement target), and the actual measurement value itself does not exist. State is also included.

When the actual measurement value is determined to be abnormal by the actual measurement value determination unit 7, the learning device 5 replaces the actual measurement value when the actual calculation value corresponding to the actual measurement value determined to be abnormal exists. The learning term is calculated using the actual calculation value.

Here, “the actual calculation value corresponding to the actual measurement value” will be further described. First, as described above, when a physical phenomenon to be reproduced with a mathematical model is complicated, a mathematical model that reproduces the physical phenomenon may be configured by combining a plurality of simpler functions. . In such a case, an output from a certain function (hereinafter referred to as “upstream function”) is input to another function (hereinafter referred to as “downstream function”), and calculation of the mathematical model proceeds.

When there is an upstream function and a downstream function having such a relationship, the update of the learning term in the downstream function is normally performed by inputting the actual value collected by the actual value collection device 4 to the downstream function. This is done using the actual calculation value output from the downstream function. However, if the actual measurement value to be input to the downstream function is abnormal, the actual measurement value cannot be input to the downstream function, so that the downstream function cannot be learned.

Therefore, learning in the downstream function is advanced by inputting the actual calculation value output from the upstream function to the downstream function instead of the actual measurement value. That is, the above-mentioned “actual calculation value corresponding to the actual measurement value” is the actual calculation value of the upstream function used as the input of the downstream function in the mathematical model.

In this way, when learning in the downstream function is advanced by inputting the actual calculation value output from the upstream function to the downstream function instead of the actual measurement value, if it is natural (that is, input to the downstream function) If a normal actual measurement value to be obtained has been obtained), errors that should be absorbed by the upstream function learning term are also absorbed by the downstream learning term. Then, when the correct measured value used for the input of the downstream function can be obtained after that, when the learning in the upstream function and the downstream function is resumed, the upstream function and the downstream function are The situation that the learning accuracy in each learning term of the side function is lowered falls into the situation.

Therefore, in the learning device 5 according to the present invention, the downstream function output calculated using the actual calculation value output from the upstream function instead of the actual measurement value as an input, and the actual value corresponding to this output, Is provided to the learning term of both the upstream function and the downstream function with the distribution ratio obtained by a predetermined procedure.

When the actual value to be input to the downstream function is determined by the actual value determination unit 7 to be abnormal, the complementary learning calculation unit 8 outputs the downstream function and the actual value corresponding to this output. First, the error is calculated. Then, this error is distributed to the learning terms of both the upstream function and the downstream function with the distribution ratio obtained by a predetermined procedure.

The procedure for obtaining the error distribution ratio will be explained. First, the error between the output of the mathematical model and the actual value becomes smaller as the learning of the mathematical model progresses. Therefore, in a state where learning of the mathematical model is sufficiently advanced, a change when the learning term is updated becomes small. In such a state where the learning term is stable, the relative size of the learning term of each function can be evaluated using the size of each function as a guide. Here, the size of a function is the size of an output from the function based on the input to the function.

Therefore, the complementary learning calculation unit 8 obtains the magnitude of the output from the function based on the input to the function for each of the upstream function and the downstream function. Then, the ratio between the magnitude of the upstream function and the magnitude of the downstream function thus obtained is set as the distribution ratio. Here, as a value input to the function when obtaining the size of each function, the measured value collected by the measured value collecting device 4 is used in principle. However, when the actual measurement value to be input is abnormal and cannot be used, another value such as a result calculation value may be substituted.

The complementary learning calculation unit 8 distributes the error to the learning terms of both the upstream function and the downstream function with the distribution ratio thus obtained. Then, the mathematical model learning calculation unit 6 updates each learning term of the upstream function and the downstream function based on the error distributed by the complementary learning calculation unit 8.

Next, the setting calculation device 3 and the learning device 5 will be described in more detail with reference to FIG. Here, for convenience of explanation, it is assumed that the mathematical model is composed of two functions, that is, an upstream function and a downstream function, and as shown in FIG. It is assumed that the downstream function is further composed of a model expression 2a and a model expression 2b.

The setting calculation device 3 calculates the set value of the machine facility 1 so that the output of the mathematical model representing the result of the operation of the machine facility 1 is closer to the target value. The calculation of the output of the mathematical model will be described step by step. First, regarding the upstream function of the mathematical model, the function of the model formula 1a is f, the input physical quantity is V ⁰ , the set value of the mechanical equipment 1 is X ¹ _i (i = 1, 2,...), And other condition inputs Assuming that a ¹ _j (j = 1, 2,...), the intermediate result output Y ¹ from the model equation 1a is expressed by the following equation (1).

Y ¹ _Z , which is the result of correcting the halfway result output Y ¹ by the learning term expressed by the following equation (2), is obtained. Then, V ¹ that is an output from the upstream function is obtained by calculating the following equation (3) using the obtained Y ¹ _Z as input of the model equation 1b. Here, in Equation (2), H is an error correction function, Z ¹ is a coefficient of a learning term, and in Equation (3), g is a function of model equation 1b, W ¹ _l (l = 1, 2). ,... Are other variable inputs, and b ¹ _k (k = 1, 2,...) Are other condition inputs.

Thus, V ¹ output from the upstream function is input to the downstream function, and the calculation of the mathematical model proceeds. For the downstream function of the mathematical model, the function of the model formula 2a is f, the set value of the mechanical equipment 1 is X ² _i (i = 1, 2,...), And other condition inputs are a ² _j (j = 1, 2). ,..., The intermediate result output Y ² from the model equation 2a is expressed by the following equation (4).

Then, the following equation (6) is input with Y ² _{Z obtained} as a result of correcting the intermediate result output Y ² by the learning term expressed by the following equation (5) as the input of the model equation 2b. By calculating, V ² that is an output from the downstream function is obtained. The V ² is the final result output from the mathematical model. Here, in equation (5), H is an error correction function, Z ² is a coefficient of a learning term, and in equation (6), g is a function of model equation 2b, W ² _l (l = 1, 2). ,... Are other variable inputs, and b ² _k (k = 1, 2,...) Are other condition inputs.

The setting calculation device 3 determines the set value of the mechanical equipment 1 by obtaining X ¹ _i and X ² _i such that the final result output V ² obtained in this way is equal to the target value V ^AIM . That is, X ¹ _i and X ² _i satisfying the expressions (Expression 1) to (Expression 7) are obtained after placing them as the following Expression (7).

The mathematical model learning calculation unit 6 included in the learning device 5 includes an actual calculation value calculation unit 6a and a learning term calculation unit 6b. The actual calculation value calculation unit 6 a calculates the actual calculation value using the same mathematical model as that used in the setting calculation device 3. The learning term calculation unit 6b compares the actual calculation value calculated by the actual calculation value calculation unit 6a with the actual value obtained by the actual measurement value collected by the actual measurement value collection device 4 so that the difference between these values is reduced. Calculate the learning term of the mathematical model.

Regarding calculation of the learning term in the mathematical model learning calculation unit 6, the calculation of the actual calculation value in the actual calculation value calculation unit 6a will be described first. Since the calculation of the actual calculation value is basically the same for the upstream function and the downstream function, the subscripts “1” and “2” representing the upstream function and the downstream function are omitted. The function of the model formula a is f, the actual value of the input physical quantity is V ^ACT , the actual value of the set value of the mechanical equipment 1 is X ^ACT _i (i = 1, 2,...), And other condition inputs are a _j ( j = 1, 2,...), the intermediate result output (actual calculation value) Y ^ACAL from the model equation a is expressed by the following equation (8).

Y ^ACAL _Z , which is the result of correcting the halfway result output Y ^ACAL by the learning term expressed by the following equation (9), is obtained. Then, using the obtained Y ^ACAL _Z as an input of the model formula b, the following formula (Equation 10) is calculated to obtain the actual calculated value V ^ACAL . Here, in Equation (9), H is an error correction function, Z is a coefficient of a learning term, and in Equation (10), g is a function of model equation b, W _l (l = 1, 2,... ) Are other variable inputs, and b _k (k = 1, 2,...) Are other condition inputs.

Next, calculation of learning terms in the learning term calculation unit 6b will be described. The learning term calculation is the same for the upstream function and the downstream function in principle, as in the case of the actual calculation value. In the example described here, the learning term is applied to the intermediate result output from the model equation a. Therefore, the comparison between the actual calculation value and the actual value for learning is performed in the intermediate result output from the model formula a.

Therefore, first, on the basis of the measured values V ^ACT, the actual value Y ^ACT corresponding to intermediate results output from the model formula a is calculated by the following equation (11). In the equation (11), g ⁻¹ is an inverse function of the model equation b, W ^ACT _l (l = 1, 2,...) Is an actual measurement value of other variable inputs, b _k (k = 1, 2,. ) Is another condition input.

The error Z ^CUR is calculated by the following equation (12) from the actual value Y ^ACT thus obtained and Y ^ACAL obtained by the equation (8). In the equation (12), h is an error calculation function.

The specific form of the error calculation function h may be, for ^example, as taking the difference between the ^{Y ACT} and ^{Y ACAL} ^{^{^{(i.e., Z CUR = Y ACT -Y ACAL}}} ), Y ACT and ^{Y ACAL} (Ie, Z ^CUR = Y ^ACT / Y ^ACAL ). Depending on the specific form of the error calculation function h, the specific form of the error correction function H in the expressions (Equation 2), (Equation 5), and (Equation 9) changes. That is, when the error calculation function h takes the difference of the input variables, the error correction function H takes the sum of the input variables, and when the error calculation function h takes the ratio of the input variables, The error correction function H is a product of input variables.

Then, this error is smoothed by the following equation (13) and reflected in the learning term to calculate a new learning term Z ^NEW . In this equation (13), Z ^NEW is a learning term used in the next setting calculation in the setting calculation device 3, Z ^OLD is a learning term used in the previous setting calculation in the setting calculation device 3, and β is a smoothing. It is a coefficient.

The above is the calculation method of the learning term when the actual measurement value necessary for the calculation of the learning term can be normally obtained. On the other hand, as described above, when the actual measurement value input to the downstream function is abnormal in the calculation of the learning term of the downstream function, the actual calculation value output from the upstream function is converted to the downstream function. By inputting and proceeding with the calculation, the actual calculation value necessary for learning in the downstream function is obtained. In the following, calculation of the actual calculation value necessary for learning in the downstream function when the actual measurement value input to the downstream function is abnormal in the calculation of the learning term of the downstream function will be described.

In this case, the input to the downstream function model equation 2a is the V ^ACAL calculated using the equation (10) in the upstream function. Therefore, the function of the model function 2a of the downstream function is f, the measured value of the set value of the mechanical equipment 1 is X ^2ACT _i (i = 1, 2,...), And other condition inputs are a ² _j (j = 1, 2), the intermediate result output Y ^2ACAL from the model equation 2a is expressed by the following equation (14). In the equation (14), V ^1ACAL is a result calculation value calculated using the equation (10) in the upstream function.

Then, thus performing the calculated ^{Y 2ACAL,} from the equation (11) actual values obtained by substituting the measured value ^{^V} ACT ⁼ ^V ^2ACT in formula ^{Y 2ACT,} the calculation of the learning term. However, as described above, since Y 2 ^ACAL includes errors of both the upstream function and the downstream function, the complementary learning calculation unit 8 ^{converts the} error (Y ^ACT −Y 2 ^ACAL ) into the upstream function. And to the learning terms of both the downstream function.

The calculation method of the error distribution ratio in the complementary learning calculation unit 8 varies depending on the form of the error calculation function h in the equation (12). Specifically, when the error calculation function h takes the difference of the input variables, the error distribution in the complementary learning calculation unit 8 is proportional distribution based on the following equations (15) and (16). Is done.

Further, when the error calculation function h takes a ratio of input variables, the error distribution in the complementary learning calculation unit 8 is performed by proportional distribution based on the following equations (17) and (18). .

In these equations (15) to (18), Z ^1CUR is an error distributed to the learning term of the upstream function, Z ^2CUR is an error distributed to the learning term of the upstream function, and f is It is a function of the model formula a.

After the error is proportionally distributed to the learning terms of the upstream function and the downstream function in this way, the error distributed to each learning term of the upstream function and the downstream function is smoothed by the equation (13). And then reflected in the learning terms. The updated new learning term is used for setting value calculation in the setting calculation device 3 at the next setting timing.

The flow of the operation in the learning device 5 configured as described above will be described once again with reference to FIG. First, in step S1, the actual measurement value collection device 4 collects actual measurement values. Next, it progresses to step S2, and the measured value determination part 7 confirms whether ^V2ACT is abnormal among the measured values which the measured value collection device 4 collected. If ^V2ACT is not abnormal, the process proceeds to step S3. In step S ^<b> 3, the actual measurement value determination unit 7 confirms whether or not V ^1ACT is abnormal among the actual measurement values collected by the actual measurement value collection device 4. If V ^1ACT is not abnormal, the process proceeds to step S4.

In step S4, the result calculation value calculation unit 6a calculates the result calculation value ^Y1ACAL of the upstream function using the equation (8). In subsequent step S5, the learning term calculation unit 6b calculates the actual value ^Y1ACT of the upstream function using the equation (11). Next, the process proceeds to step S6, where the learning term calculation unit 6b calculates the error ^Z1CUR of the upstream function using the equation (12).

After step S6, the process proceeds to step S7. In this step S7, the actual calculated value calculating section 6a calculates the actual calculated value ^{Y 2ACAL} downstream function by substituting the measured values ^{V 1ACT} the ^{V ACT} of (number 8). In subsequent step S8, the learning term calculation unit 6b calculates the actual value ^Y2ACT of the downstream function using the equation (11). Next, the process proceeds to step S9, where the learning term calculation unit 6b calculates the error Z ^2CUR of the downstream function using the equation (12).

After step S9, the process proceeds to step S10. In this step S10, the learning term calculation unit 6b (number 13) of the by substituting ^{Z 1CUR} and ^{Z 2CUR} respectively ^{Z CUR,} the learning term of learning term ^{Z 1 NEW} and downstream function of the upstream function ^{Z 2NEW} Calculate Then, the process proceeds to step S11, and updates the learning term of the calculated learning term Z ^{1 NEW} and mathematical models of setting calculation device 3 in Z ^2NEW, a series of learning term updating process ends.

On the other hand, if ^V1ACT is abnormal in step S3, the ^process proceeds to step S20. In step S20, the actual calculation value calculation unit 6a calculates the actual calculation value ^Y1ACAL of the upstream function using the equation (8). In subsequent step S21, the actual calculation value calculation unit 6a calculates the actual function calculation value ^V1ACAL of the upstream function using Equation (9) and Equation (10).

After step S21, the process proceeds to step S22. In this step S22, actual calculated value calculating section 6a calculates the actual calculated value ^{Y 2ACAL} downstream functions using ^{V 1ACAL} by equation (14) below. In subsequent step S23, the learning term calculation unit 6b calculates the actual value ^Y2ACT of the downstream function using the equation (11).

Then, the process proceeds to step S24, the complementary learning calculation unit 8 (number 15) and (Expression 16), or by using the equation (17) and (Equation ^18), calculates the ^{Z 1CUR} and ^{Z 2CUR} To do. After step S24, the process proceeds to step S10 described above.

In step S2, if ^V2ACT is abnormal, the process proceeds to step S30. In step S30, the actual measurement value determination unit 7 confirms whether or not ^V1ACT is abnormal among the actual measurement values collected by the actual measurement value collection device 4. If V ^{1ACT is} also abnormal, the learning term is not updated in both the upstream function and the downstream function.

On the other hand, if V ^1ACT is not abnormal in step S30, the ^process proceeds to step S31. In step S31, the actual calculation value calculation unit 6a calculates the actual function calculation value ^Y1ACAL of the upstream function using the equation (8). In subsequent step S32, the learning term calculation unit 6b calculates the actual value ^Y1ACT of the upstream function using the equation (11). Next, the process proceeds to step S33, where the learning term calculation unit 6b calculates the error Z ^1CUR of the upstream function using the equation (12).

After step S33, the process proceeds to step S34. In step S34, the learning term calculation unit 6b calculates the learning term ^Z1NEW of the upstream function using the equation (13). And it progresses to step S35, the learning term of the numerical formula model of the setting calculation apparatus 3 is updated by the calculated learning term ^Z1NEW , and a series of learning term update processes are complete ^| finished.

The learning device of the setting calculation system configured as described above includes a learning term calculation unit that calculates the learning terms of the upstream function and the downstream function that constitute the mathematical model using actual measurement values, and a learning term calculation unit. In the calculation of the learning term of the downstream function, an actual value determination unit that determines whether or not the first actual value input to the downstream function is abnormal is determined, and the first actual value is determined to be abnormal In the case where the output from the upstream function is input to the downstream function, the actual calculation value calculation unit for calculating the actual calculation value output from the downstream function, and the second corresponding to the output from the downstream function A complementary learning calculation unit that distributes an error of the actual calculation value with respect to the actual measurement value to the learning term of the upstream function and the learning term of the downstream function.

Therefore, even if the actual measurement value to be input to the downstream function for learning is not obtained, the actual calculation value output from the upstream function is input to the downstream function instead of the actual measurement value. As the learning proceeds, the accumulated error is distributed to the learning term of the upstream function, thereby realizing appropriate learning in both the upstream function and the downstream function.

That is, even if some of the actual values necessary for learning the mathematical model are not obtained, learning using the actual values for the obtained amount is performed, and the learning terms of each function constituting the mathematical model are set. It is possible to update appropriately, and it is possible to suppress a decrease in accuracy of the mathematical model. For this reason, it can contribute to the improvement of the calculation accuracy of the set value of the mechanical equipment.

Note that the learning device of the setting calculation system configured as described above includes, for example, functions performed by each unit of the mathematical model learning calculation unit, the actual calculation value calculation unit, the learning term calculation unit, the actual measurement value determination unit, and the complementary learning calculation unit. It is also possible to construct information processing for realizing the above by causing a hardware resource having a central processing unit, a storage device, and the like to execute.

The present invention can be used in a learning apparatus and a learning method for updating a learning term of a mathematical model using an actual measurement value in a setting calculation system that calculates a setting value of a mechanical facility using a mathematical model.

DESCRIPTION OF SYMBOLS 1 Mechanical equipment 2 Setting calculation system 3 Setting calculation apparatus 4 Actual value collection apparatus 5 Learning apparatus 6 Formula model learning calculation part 6a Actual calculation value calculation part 6b Learning term calculation part 7 Actual value determination part 8 Complementary learning calculation part

Claims

In a setting calculation system that calculates a setting value of a mechanical facility using a mathematical model, a learning device that updates a learning term of the mathematical model using an actual measurement value,
A learning term calculation unit that calculates learning terms for each of the upstream function and the downstream function constituting the mathematical model using the measured values;
An actual value determination unit that determines whether or not the first actual measurement value that is the actual measurement value input to the downstream function in the learning term calculation of the downstream function in the learning term calculation unit is abnormal; ,
When it is determined that the first actual measurement value is abnormal, the actual calculation value for calculating the actual calculation value output from the downstream function by inputting the output from the upstream function to the downstream function A calculation unit;
Complementary learning calculation that distributes the error of the actual calculation value with respect to the second actual measurement value corresponding to the output from the downstream function to the learning term of the upstream function and the learning term of the downstream function And a learning device for a setting calculation system.
The complementary learning unit is a ratio of an output magnitude from the upstream function based on an input to the upstream function and an output magnitude from the downstream function based on an input to the downstream function. The learning apparatus of the setting calculation system according to claim 1, wherein the error is distributed to the learning term of the upstream function and the learning term of the downstream function based on the equation.
The complementary learning unit
The error is obtained from the difference between the second actual measurement value and the actual calculation value,
The obtained error is stored in the upstream function learning term and the downstream function learning term, the absolute value of the difference between the input to the upstream function and the output from the upstream function, and the downstream function. 3. The learning apparatus for a setting calculation system according to claim 2, wherein proportional distribution is performed based on an absolute value of a difference between an input to the output and an output from the downstream function.
The complementary learning unit
While obtaining the error by the ratio of the second actual measurement value and the actual calculation value,
The obtained error is stored in the upstream function learning term and the downstream function learning term, the absolute value of the ratio between the input to the upstream function and the output from the upstream function, and the downstream function. 3. The learning apparatus for a setting calculation system according to claim 2, wherein proportional distribution is performed based on an absolute value of a ratio between an input to the output and an output from the downstream function.
In a setting calculation system for calculating a setting value of mechanical equipment using a mathematical model, a learning method for updating a learning term of the mathematical model using an actual measurement value,
A first step of calculating a learning term of each of the upstream function and the downstream function constituting the mathematical model using the measured value;
A second step of determining whether or not the first actual measurement value that is the actual measurement value input to the downstream function in the calculation of the learning term of the downstream function in the first step is abnormal; ,
When it is determined that the first actually measured value is abnormal, the output from the upstream function is input to the downstream function, and the actual calculation value output from the downstream function is calculated. Steps,
An error of the actual calculation value with respect to the second actual measurement value that is the actual measurement value corresponding to the output from the downstream function is distributed to the learning term of the upstream function and the learning term of the downstream function. And a learning method for the setting calculation system.