JP7679238B2 - Optimum value search control device, optimum value search control method, and optimum value search control program - Google Patents

Description

Embodiments of the present invention relate to an optimum value search control device, an optimum value search control method, and an optimum value search control program.

In recent years, a technology called extreme value control has been attracting attention as a method of plant control. Extreme value control is a model-free real-time optimal control technology that does not use a complex model of the plant. It calculates the value of the evaluation function (cost function, performance index) to be optimized (minimized or maximized) from online sensor information that can directly measure it, and adaptively searches for the evaluation function value while changing the manipulated variable so as to maintain it at the (local) optimum value (local minimum value = minimum value or local maximum value = maximum value).

That is, the evaluation function value, which represents the plant performance index that you want to maximize or minimize, is measured online. The evaluation function is defined, for example, as a unified index obtained by multiplying multiple output values of the plant by a specified coefficient and then adding them up, and searching for a direction to reduce the evaluation function value is equivalent to automatically searching for a direction of manipulated variables that will reduce all of the outputs (balance at a low value). Extreme value control is generally characterized by a single input/output, and is primarily applicable to plants with a single manipulated variable to be optimized.

JP 2017-33140 A

On the other hand, depending on the plant, operations may be determined by multiple manipulated variables. Applying extreme value control to such cases has the potential to realize control that optimizes the entire plant operation (enabling the exploration of manipulated variables). However, because conventional extreme value control optimizes a single manipulated variable, there is no guarantee that appropriate control can be obtained even if the configuration is rearranged to optimize multiple manipulated variables, and there is a possibility that the plant may end up being partially optimized rather than optimized as a whole. This is because issues arise that are specific to extreme value control, which optimizes multiple manipulated variables, that did not occur with conventional extreme value control, and unless these issues are resolved, appropriate behavior cannot be obtained with extreme value control, which optimizes multiple manipulated variables.

The embodiments of the present invention have been made in consideration of the above circumstances, and aim to provide an optimum value search control device, an optimum value search control method, and an optimum value search control program that properly operate extreme value control that optimizes multiple manipulated variables.

An optimum value search control device according to an embodiment is a device that is applied to an arbitrary target process, which receives an input of a plurality of operation variables including at least a first operation variable and a second operation variable, and outputs one evaluation function value. The optimum value search control device includes: a first signal generating unit that receives an input of the evaluation function value measured in real time and generates a first signal that oscillates the first operation variable; and a first normalization processing unit that normalizes a gradient of the evaluation function value. The optimum value search control device includes: a first extreme value searching unit that searches for an operating point of the first operation variable at which the evaluation function value is an optimal value, based on a response of the evaluation function value caused by vibrating the first operation variable by the first signal; a second signal generating unit that receives an input of the evaluation function value and generates a second signal that is different from the first signal and that vibrates the second operation variable; and a second normalization processing unit that normalizes a gradient of the evaluation function value. The optimum value search control device includes: a first extreme value searching unit that receives an input of the evaluation function value and generates a second signal that oscillates the second operation variable, the second signal being different from the first signal;

FIG. 1 is a block diagram showing an outline of an example of the configuration of an optimum value search control device according to the first embodiment. FIG. 2 is a diagram illustrating an example of a dither signal used in the optimum value search control device of the first embodiment. FIG. 3 shows an example of the results of simulating the optimization of a plurality of manipulated variables using the optimum value search control device of the first embodiment. FIG. 4 shows an example of the results of a simulation of the optimization of a plurality of manipulated variables using the optimum value search control device of the first embodiment. FIG. 5 is a block diagram illustrating an example of the configuration of an optimum value search control device of a first comparative example. FIG. 6 is a block diagram illustrating an example of the configuration of an optimum value search control device of a second comparative example. FIG. 7 shows an example of the results of a simulation of the optimization of a plurality of manipulated variables using the optimum value search control device of the second comparative example. FIG. 8 shows an example of the results of a simulation of the optimization of a plurality of manipulated variables using the optimum value search control device of the second comparative example. FIG. 9 is a block diagram illustrating an example of the configuration of an optimum value search control device according to the second embodiment. FIG. 10 is a diagram illustrating an example of a dither signal used in the optimum value search control device of the first embodiment. FIG. 11 shows an example of the results of a simulation of the optimization of a plurality of manipulated variables using the optimum value search control device of the second embodiment. FIG. 12 shows an example of the results of a simulation of the optimization of a plurality of manipulated variables using the optimum value search control device of the second embodiment. FIG. 13 is a block diagram illustrating an example of the configuration of an optimum value search control device according to the third embodiment. FIG. 14 is a diagram illustrating an example of a dither signal used in the optimum value search control device of the third embodiment. FIG. 15 shows an example of the results of a simulation of the optimization of a plurality of manipulated variables using the optimum value search control device of the third embodiment. FIG. 16 shows an example of the results of simulating the optimization of a plurality of manipulated variables using the optimum value search control device of the third embodiment.

An optimum value search control device, an optimum value search control method, and an optimum value search control program according to an embodiment will be described below with reference to the drawings.
FIG. 1 is a block diagram showing an outline of an example of the configuration of an optimum value search control device according to the first embodiment.
The optimum value search control device of this embodiment optimizes a plurality of manipulated variables u1, u2 to be input to the plant 101 by using the value of the evaluation function y measured in real time from the plant 101.

As an optimization control technique, extreme value control is a control technique that adaptively searches for the optimal value of an evaluation function based on the change in the evaluation function value in response to a change in the manipulated variable. The evaluation function value is an index value that is determined based on the controlled variable of the controlled process, and the relationship between the evaluation variable and the controlled variable is expressed by a predetermined function. This evaluation function may be set based on any evaluation criterion based on the controlled variable. The evaluation variable may also be the controlled variable itself. In a controlled process in extreme value control, the evaluation function may be an unknown function of the manipulated variable.

In extreme value control, for example, the manipulated variable is continuously oscillated by a dither signal, and the change (increase or decrease) in the evaluation function value is observed. The dither signal is often given as a sine wave, but is not limited to a sine wave. Based on the change in the evaluation function value in response to the change in the manipulated variable, the manipulated variable is changed in a direction that brings the evaluation function value closer to the optimal value. The optimal value of the evaluation function is searched for by repeating such changes in the manipulated variable.

The optimum value search control device includes a first search control unit 200 that optimizes a first manipulated variable u1, and a second search control unit 300 that optimizes a second manipulated variable u2.
The first search control unit 200 includes a high-pass filter (HPF) 201, a first signal generation unit, a multiplier 203, a low-pass filter (LPF) 204, an integrator 205, a multiplier 207, an adder 208, and a normalization processing unit (first normalization processing unit) 209. The first signal generation unit includes a first dither signal generator 206 and a second dither signal generator 202.

The high-pass filter 201 can remove a constant bias corresponding to the optimal value of the evaluation function y fed back from the plant 101 .
The first dither signal generator 206 and the second dither signal generator 202 output perturbation signals (dither signals) to be added to the operation amounts u1 and u2. The first dither signal generator 206 and the second dither signal generator 202 can vary the evaluation function value.

The signal output from the first dither signal generator 206 is multiplied by the amplitude a in the multiplier 207 and then input to the adder 208. The signal output from the second dither signal generator 202 is input to the multiplier 203. The multiplier 203 outputs the product of the evaluation function y passed through the high-pass filter 201 and the dither signal.

In this embodiment, the first dither signal generator 206 and the second dither signal generator 202 output a dither signal including a first oscillation period and a first pause period. The dither signal output by the first dither signal generator 206 and the second dither signal generator 202 is, for example, a signal that alternates between a first oscillation period that outputs a sine wave of one wavelength and a first pause period of one wavelength. This method of applying a signal is called a "cyclic" method, and in the following description, the dither signal in this embodiment is referred to as a cyclic sine wave.

The waveform of the signal output from the first dither signal generator 206 and the waveform of the signal output from the second dither signal generator 202 can be set to the same waveform if the time delay and dead time in the plant 101 can be ignored. If the plant 101 has a time delay and dead time, the waveform of the sine wave output from the first dither signal generator 206 and the waveform of the sine wave output from the second dither signal generator 202 may be different by incorporating a predetermined phase compensation or dead time compensation. If the first dither signal generator 206 and the second dither signal generator 202 output the same dither signal, the optimum value search control device may be configured to include one dither signal generator and output a dither signal from the one dither signal generator to the multiplier 203 and the multiplier 207.

The low-pass filter 204 has the role of extracting low-frequency components from the output of the multiplier 203, which makes it possible to know whether the evaluation function value has increased or decreased.
The normalization processor 209 normalizes the gradient of the low-frequency components of the evaluation function y extracted by the low-pass filter 204. The normalization process in the normalization processor 209 will be described in detail later.

The integrator 205 integrates the evaluation function y after normalization processing by the normalization processing unit 209, thereby functioning as an estimator that estimates the direction of the manipulated variable that should be moved in order to bring the evaluation function value closer to the optimal value.
The adder 208 adds the value output from the integrator 205 and the dither signal output from the multiplier 207, and outputs the result to the plant 101 as a first manipulated variable u1.

The second search control unit 300 includes a high-pass filter 301, a second signal generation unit, a multiplier 303, a low-pass filter 304, an integrator 205, a multiplier 307, an adder 308, and a normalization processing unit (second normalization processing unit) 309. The second signal generation unit includes a third dither signal generator 306 and a fourth dither signal generator 302.

In this embodiment, the second search control unit 300 is configured such that the dither signals (second signals) output from the third dither signal generator 306 and the fourth dither signal generator 302 are different from the dither signals (first signals) output from the first dither signal generator 202 and the second dither signal generator 206, and the second signals are signals whose timing of oscillation (or timing of pause) is shifted relative to the first signals, for example.
The high-pass filter 301 can remove a constant bias corresponding to the optimal value of the evaluation function y fed back from the plant 101 .

The third dither signal generator 306 and the third dither signal generator 302 output a perturbation signal (dither signal) to be added to the manipulated variable u2. The signal output from the third dither signal generator 306 is multiplied by the amplitude a in the multiplier 307 and then input to the adder 308. The signal output from the fourth dither signal generator 302 is input to the multiplier 303. The multiplier 303 outputs the product of the evaluation function y passed through the high-pass filter 301 and the teaser signal.

The third dither signal generator 306 and the fourth dither signal generator 302 output a dither signal including a second oscillation period and a second pause period. The dither signal output by the first dither signal generator 206 and the second dither signal generator 202 is, for example, a cyclic sine wave that alternates between a second oscillation period in which a sine wave of one wavelength is output and a second pause period of one wavelength.

The waveform of the signal output from the third dither signal generator 302 and the waveform of the signal output from the fourth dither signal generator 306 can be set to the same waveform if the time delay and dead time in the plant can be ignored. If the plant has a time delay or dead time, the waveform of the signal output from the third dither signal generator 306 and the waveform of the signal output from the fourth dither signal generator 302 may be different by incorporating a predetermined phase compensation or dead time compensation. If the third dither signal generator 306 and the fourth dither signal generator 302 output the same dither signal, the optimum value search control device may be configured to include one dither signal generator and output dither signals from the one dither signal generator to the multipliers 203 and 207.

FIG. 2 is a diagram illustrating an example of a dither signal used in the optimum value search control device of the first embodiment.
Here, the waveform of the dither signal (first signal) applied to the first manipulated variable u1 and the waveform of the dither signal (second signal) applied to the second manipulated variable u2 are shown. In this embodiment, the timing of the oscillating period and the pause period of the first signal applied to the first manipulated variable u1 and the timing of the oscillating period and the pause period of the second signal applied to the second manipulated variable u2 are shifted.

That is, the first signal includes a first pause period T1, and the second signal includes a second pause period T2. The waveform of the first signal is sinusoidal during at least a portion of the second pause period T2, and is a constant value (=0) that does not oscillate during the first pause period T1. The waveform of the second signal is sinusoidal during at least a portion of the first pause period T1, and is a constant value (=0) that does not oscillate during the second pause period T2.

In this embodiment, the first pause period T1 of the first signal is a period of one wavelength of the sine wave (oscillating waveform) of the second signal. The second pause period T2 of the second signal is a period of one wavelength of the sine wave (oscillating waveform) of the first signal. Therefore, even if sine waves of the same cycle are set as the oscillatory waveforms of the first and second signals, the application timing is different, so that it is possible to eliminate the change due to the operation amount of one at the timing of extracting the change in the operation amount of the other.

The application time (oscillation period) and rest time of the dither signal applied to each of the manipulated variables u1 and u2 may be determined arbitrarily. For example, the period during which the dither signal applied to one manipulated variable oscillates may be set to be the rest period of the dither signal applied to the other manipulated variable.

In addition, in this embodiment, the vibration waveforms of the first dither signal and the second signal are sine waves, but the waveforms that can be used as dither signals are not limited to sine waves. For example, any signal can be used as long as it is an excitation signal that generates periodic vibration, such as a cosine wave, a square wave, a triangular wave, or a sawtooth wave. Then, the periods during which the first signal used to optimize the first manipulated variable u1 and the second signal used to optimize the second manipulated variable u2 oscillate can be set to be different.

Furthermore, in this embodiment, an extreme value control algorithm for optimizing two manipulated variables u1 and u2 has been described. Similarly, when optimizing three or more manipulated variables, the periods during which the dither signals used for each optimization oscillate can be shifted from each other to avoid interference between the responses caused by moving the two manipulated variables u1 and u2. In this case, it is necessary to set the oscillation period and the pause period during which the dither signal applied to each of the multiple manipulated variables oscillates according to the number of manipulated variables.

The low-pass filter 304 has the role of extracting low-frequency components from the output of the multiplier 303, which makes it possible to know whether the evaluation function value has increased or decreased.
The normalization processor 309 normalizes the gradient of the low-frequency components of the evaluation function y extracted by the low-pass filter 304. The normalization process in the normalization processor 309 will be described in detail later.

The integrator 305 integrates the evaluation function y after normalization processing by the normalization processing unit 309, thereby functioning as an estimator that estimates the direction of the manipulated variable that should be moved in order to bring the evaluation function value closer to the optimum value.
The adder 308 adds the value output from the integrator 305 and the dither signal output from the multiplier 307, and outputs the result to the plant 101 as a second manipulated variable u2.

Next, the normalization process performed by the normalization processors 209 and 309 will be described.
The normalization processing units 209 and 309 perform normalization processing for normalizing the gradient of the low-frequency components of the evaluation function y extracted by the low-pass filters 204 and 304. In this embodiment, for example, if the gradient information of the low-frequency components of the evaluation function is G[t], the gradient information Gn[t] of the normalized evaluation function is expressed as Gn[t]=G[t]/|G[t]|.
As a result of this operation, the normalized gradient information Gn[t] is always greater than or equal to -1 and less than or equal to 1 (-1≦Gn[t]≦1), making it possible to prevent the gradient of the evaluation function from taking on an excessively large value.

In addition, in the normalization processing units 209 and 309, if it is possible to prevent the gradient magnitude from becoming excessive when the gradient magnitude changes depending on the situation and to prevent the gradient magnitude from exceeding a certain level, a similar effect can be obtained even with normalization processing other than the above. For example, the normalization processing units 209 and 309 perform normalization processing to make the magnitude of the absolute value of the signal less than 1, but other than this, it may also be configured to set an arbitrary constant α and perform normalization processing to make the absolute value of the signal less than that value (constant α).

If the gradient information component of the extreme value control algorithm that optimizes the manipulated variable u1 is G1[t] (the component of the gradient of the evaluation function due to the manipulated variable u1), the gradient information of the normalized evaluation function is Gn1[t], the gradient information of the extreme value control algorithm that optimizes the manipulated variable u2 is G2[t] (the component of the gradient of the evaluation function due to the manipulated variable u2), and the gradient information of the normalized evaluation function is Gn2[t], then the respective normalized signals are expressed as follows.

Gn1[t]=k1G1[t]/|G1[t]+ε1|
Gn2[t]=k2G2[t]/|G2[t]+ε2|
Here, ε1 and ε2 are values provided to prevent division by zero, and k1 and k2 are arbitrary coefficients that satisfy the following relational expression:
k1 ² +k2 ² =1

This is a process for generating a normalized signal of the gradient of the evaluation function y when the evaluation function y is composed of the sum of the squares of the manipulated variable u1 component and the manipulated variable u2 component. As described above, when the gradient of the evaluation function is to be within a predetermined constant α, the right-hand side of the above equation is set to α (k1 ² + k2 ² = α).

In extreme value control of an evaluation function using three or more manipulated variables, the same effect can be obtained by calculating and applying the components of the gradient information of the evaluation function for the number of optimized manipulated variables. Note that in the extreme value control algorithm that optimizes the manipulated variables, the sum of the squares of each coefficient is set to 1 (or a constant α).

Furthermore, the above normalization method is a normalization process using a so-called soft sine function, but other normalization processes may also be performed using, for example, a sigmoid function, a Gompertz function, an arctangent function (arctan), a Gutermann function, or a function x/(ε+x ^p ) ^(1/p) .

These functions are functions that change smoothly between -1 and 1 (functions that change between -α and α if kept within a constant α), and any function with similar characteristics can be used for normalization. For example, the above normalization can be achieved by using a function that adds upper and lower limits to any monotonically increasing function.

Next, the effects of the optimum value search control device, the optimum value search control method, and the optimum value search control program of the present embodiment will be described.
3 and 4 show an example of the results of simulating the optimization of a plurality of manipulated variables using the optimum value search control device of the first embodiment.

Here, we will explain an example in which the minimum value of the evaluation function y is the optimal operating point, the optimal value of the first operation amount u1 is 2, and the optimal value of the operation amount u2 is 5. Figure 3 shows an example of the optimal value search result for the first operation amount u1 along the time axis. Figure 4 shows an example of the optimal value search result for the second operation amount u2 along the time axis.

According to the simulation results, it can be seen that the optimum value search control device of the first embodiment can converge to an optimum value within a predetermined time without causing the first manipulated variable u1 and the second manipulated variable u2 to interfere with each other and without causing a sudden, transient change in the manipulated variable. From the above, by applying the algorithm of this embodiment, it is possible to appropriately control the convergence speed in the multivariable extreme value control.
That is, according to the optimum value search control device, the optimum value search control method, and the optimum value search control program of the present embodiment, it is possible to appropriately operate extreme value control that optimizes a plurality of manipulated variables.

The effects of this embodiment will be further explained below in comparison with the operation of an optimum value search control device of a comparative example.
FIG. 5 is a block diagram illustrating an example of the configuration of an optimum value search control device of a first comparative example.

The optimum value search control device of this comparative example has one evaluation function value, which is determined by the operating conditions of multiple operation amounts u1 and u2. The optimum value search control device of this comparative example differs from the optimum value search control device of the first embodiment described above in that the dither signal applied to the first operation amount u1 and the dither signal applied to the second operation amount u2 have the same waveform, and in that it does not have normalization processing units 209 and 309.

The optimal value search control device of this comparative example includes a first search control unit 200 that optimizes the manipulated variable u1, and a second search control unit 300 that has the same function as the first search control unit 200, and the first search control unit 200 and the second search control unit 300 are connected in parallel to the plant 101.

First, in extreme value control that optimizes one manipulated variable, we will explain an example of searching for the optimal manipulated variable (the manipulated variable that minimizes the evaluation function) when the relationship between the unknown evaluation function and the manipulated variable has a downward convex quadratic function shape.

In the optimization process, in which the manipulated variable is driven by a dither signal and the manipulated variable that minimizes the evaluation function is searched for, the behavior differs depending on the operating point of the manipulated variable. Since the relationship between the manipulated variable and the evaluation function is unknown, in order to find the direction in which to move the manipulated variable next, it is necessary to add a dither signal to the manipulated variable and know the response of the evaluation function that results.

For example, when the manipulated variable is at a certain operating point A, the evaluation function value decreases when the manipulated variable is increased. Therefore, when the manipulated variable is driven by a sine wave at this operating point A, the resulting response of the evaluation function is in the opposite phase to the movement of the manipulated variable. On the other hand, when the manipulated variable is at another operating point B, the evaluation function value increases when the manipulated variable is increased. Therefore, when the manipulated variable is driven by a sine wave at this operating point B, the resulting response of the evaluation function is in the same phase as the movement of the manipulated variable.

Information on the response of the evaluation function to the movement of the manipulated variable is obtained by multiplying the value of the evaluation function by a dither signal in multipliers 203 and 303, and passing the result through low-pass filters 204 and 304. This part of the signal mainly corresponds to gradient information of the evaluation function, and extracts information on the degree (magnitude) of how much the manipulated variable should be moved, as well as information on the direction in which the manipulated variable should be moved.
In this way, information on the direction and magnitude of the manipulated variable to be moved next is obtained from the dither signal added to the manipulated variable and the shape of the evaluation function, which is its response waveform.

On the other hand, in the optimum value search control device of the comparative example, two manipulated variables u1 and u2 are input to the plant 101, and one evaluation function value is output from the plant 101. For example, consider a case where the optimum value search control device of this comparative example aims for the minimum value of the evaluation function formed by the two manipulated variables u1 and u2. Note that even in this example, the relationship between the first manipulated variable u1, the second manipulated variable u2, and the evaluation function cannot be known in advance, so the minimum manipulated variables u1 and u2 are searched for based on the response waveform of the evaluation function associated with each of the manipulated variables u1 and u2.

For example, when the operating points of the first manipulated variable u1 and the second manipulated variable u2 are at any position, the evaluation function is in phase with the first manipulated variable u1, and in phase with the second manipulated variable u2. In this case, if the waveforms of the dither signals of the first manipulated variable u1 and the second manipulated variable u2 are set to be the same waveform and sinusoidal, if there is no time delay, the response of the evaluation function to the first manipulated variable u1 and the response of the evaluation function to the second manipulated variable u2 will be inverse responses, and the responses of the evaluation function due to the movements of the multiple manipulated variables u1 and u2 may cancel each other out, causing the response waveform to disappear.

When this situation occurs, it becomes impossible to extract information about the direction and magnitude in which the first manipulated variable u1 and the second manipulated variable u2 should be moved next from the response of the evaluation function. As a result, depending on the operating point, there is a possibility that an event will occur in which the optimal value stops. This phenomenon occurs because of interference in the response between the manipulated variables u1 and u2, and the interference can be avoided by changing the waveform of the dither signal added to the manipulated variables u1 and u2. For example, by setting the waveform of the dither signal added to the manipulated variables u1 and u2 as in the first embodiment, the occurrence of the above event can be avoided.

FIG. 6 is a block diagram illustrating an example of the configuration of an optimum value search control device of a second comparative example.
The optimum value search control device of this comparative example has one evaluation function value, which is determined by the operating conditions of multiple manipulated variables u1 and u2. The optimum value search control device of this comparative example differs from the optimum value search control device of the above-mentioned first embodiment in that it does not include the normalization processing units 209 and 309.

The optimal value search control device of this comparative example includes a first search control unit 200 that optimizes the manipulated variable u1, and a second search control unit 300 that has the same function as the first search control unit 200, and the first search control unit 200 and the second search control unit 300 are connected in parallel to the plant 101.

In the optimal value search control device of this comparative example, as in the first embodiment, the vibration period and the pause period are set to be offset from each other in the first signal and the second signal. Therefore, in this embodiment, it is possible to avoid interference between the response of the first operation amount u1 and the response of the second operation amount u2.

On the other hand, for example, when the optimum value search control device of this comparative example aims for the minimum value of the evaluation function formed by the two manipulated variables u1 and u2, there are cases where the manipulated variables u1 and u2 change suddenly immediately after control starts.

7 and 8 show an example of the results of simulating the optimization of a plurality of manipulated variables using the optimum value search control device of the second comparative example.
Here, an example will be described in which the minimum value of the evaluation function y is the optimal operating point, the optimal value of the first manipulated variable u1 is 2, and the optimal value of the manipulated variable u2 is 5. Fig. 7 shows an example of the optimal value search result for the first manipulated variable u1 along the time axis. Fig. 8 shows an example of the optimal value search result for the second manipulated variable u2 along the time axis.

According to the results of this simulation, the optimal value search control device of the second comparative example allowed the first operation amount u1 and the second operation amount u2 to ultimately converge to the optimal value without interfering with each other. However, the first operation amount u1 and the second operation amount u2 tended to change rapidly immediately after control started (e.g., A1, A3 in Figures 7 and 8) and then gradually converge to the optimal value (e.g., A2, A4 in Figures 7 and 8). If it is assumed that the optimal value search control device of this comparative example is actually applied to a plant, such a large change in the operation amount immediately after control started is not desirable for operating the plant, and it takes a long time to converge to the optimal operating conditions.

This tendency occurs because the extreme value control operates the first manipulated variable u1 and the second manipulated variable u2 based on the gradient information of the evaluation function, and therefore the gradient of the evaluation function in the multivariable extreme value control is affected by changes in the operating point. In the extreme value control, it is theoretically possible to smooth out the sudden change in the manipulated variable immediately after the start of control, for example, by changing the magnitude of the integral gain. However, if the magnitude of the integral gain is changed, the behavior after the operating points of the first manipulated variable u1 and the second manipulated variable u2 move to a region with a small value (gradient) of the evaluation function changes even more slowly, resulting in a significant amount of time being required for convergence to the vicinity of the optimal manipulated variable. In addition, in the extreme value control, there is no guide for changing the integral gain in real time, so it was difficult to appropriately control the convergence speed of the extreme value control.

In contrast, for example, in the optimum value search control device of the first embodiment, the magnitude of the gradient of the evaluation function can be normalized by the normalization processing units 209 and 309. As a result, the optimum value search control device of the first embodiment can appropriately control the convergence speed without changing the integral gain in real time.

As described above, the optimal value search control device, optimal value search control method, and optimal value search control program of the first embodiment can appropriately operate extreme value control that optimizes multiple manipulated variables.

Next, the optimal value search control device, optimal value search control method, and optimal value search control program of the second embodiment will be described in detail with reference to the drawings. Note that in the following description, the same components as those in the first embodiment described above will be denoted by the same reference numerals and will not be described.

FIG. 9 is a block diagram illustrating an example of the configuration of an optimum value search control device according to the second embodiment.
The optimum value search control device of this embodiment includes a first search control unit 200 that optimizes a first manipulated variable u1 and a second search control unit 300 that optimizes a second manipulated variable u2, and optimizes a plurality of manipulated variables u1, u2 to be input to the plant 101 using an evaluation function y obtained from the plant 101. The optimum value search control device of this embodiment differs from the above-described first embodiment in the first signal applied to the first manipulated variable u1 and the second signal applied to the second manipulated variable u2.

The first dither signal generator 206 and the second dither signal generator 202 output perturbation signals (dither signals) to be added to the manipulated variables u1 and u2. The signal output from the first dither signal generator 206 is input to the multiplier 203. The signal output from the second dither signal generator 202 is multiplied by the amplitude a in the multiplier 207 and then input to the adder 208. The multiplier 203 outputs the product of the evaluation function y passed through the high-pass filter 201 multiplied by the dither signal.

In this embodiment, the first dither signal generator 206 and the second dither signal generator 202 output a sine wave as a dither signal. The waveform of the sine wave output from the first dither signal generator 206 and the waveform of the sine wave output from the second dither signal generator 202 can be set to the same waveform if the time delay and dead time in the plant 101 can be ignored. If the plant 101 has a time delay and dead time, the waveform of the sine wave output from the first dither signal generator 206 and the waveform of the sine wave output from the second dither signal generator 202 may be different by incorporating a predetermined phase compensation or dead time compensation. When the first dither signal generator 206 and the second dither signal generator 202 output the same dither signal, the optimum value search control device may be configured to include one dither signal generator and output a dither signal to the multiplier 203 and the multiplier 207 from the one dither signal generator.

In this embodiment, the second search control unit 300 outputs a dither signal (second signal) from the third dither signal generator 306 and the fourth dither signal generator 302 that is different from the dither signal (first signal) from the first dither signal generator 202 and the second dither signal generator 206, and the second signal is, for example, a signal that is out of phase with respect to the first signal.

The third dither signal generator 306 and the third dither signal generator 302 output a perturbation signal (dither signal) to be added to the manipulated variable u2. The signal output from the third dither signal generator 306 is input to the multiplier 303. The signal output from the fourth dither signal generator 302 is multiplied by the amplitude a in the multiplier 307 and then input to the adder 308. The multiplier 303 outputs the evaluation function y passed through the high-pass filter 301 multiplied by the dither signal.

The third dither signal generator 306 and the fourth dither signal generator 302 output a cosine wave as a dither signal. The waveform of the cosine wave output from the third dither signal generator 302 and the waveform of the cosine wave output from the fourth dither signal generator 306 can be set to the same waveform if the time delay and dead time in the plant can be ignored. If the plant has a time delay or dead time, the waveform of the cosine wave output from the third dither signal generator 306 and the waveform of the cosine wave output from the fourth dither signal generator 302 may be different by incorporating a predetermined phase compensation or dead time compensation. When the third dither signal generator 306 and the fourth dither signal generator 302 output the same dither signal, the optimum value search control device may be configured to include one dither signal generator and output a dither signal to the multiplier 203 and the multiplier 207 from the one dither signal generator.

FIG. 10 is a diagram illustrating an example of a dither signal used in the optimum value search control device of the first embodiment.
In the optimum value search control device of this embodiment, the phase difference between the dither signal used to optimize the first manipulated variable u1 and the dither signal used to optimize the second manipulated variable u2 is 90°. In this embodiment, the phase of the dither signal for each of the multiple manipulated variables is shifted to suppress interference between the responses of the multiple manipulated variables u1 and u2. Note that the phase difference of the dither signal for each of the multiple manipulated variables is not limited to 90°, and any phase difference may be set as long as it can effectively prevent interference.

In addition, in this embodiment, the dither signal used to optimize the first manipulated variable u1 is a sine wave, and the dither signal used to optimize the second manipulated variable u2 is a cosine wave, but the dither signal used to optimize the second manipulated variable u2 may be set to a sine wave, in which case the dither signal used to optimize the first manipulated variable u1 may be set to a cosine wave.

Furthermore, the waveforms that can be used as the dither signal are not limited to sine waves and cosine waves. For example, any excitation signal that generates periodic vibrations, such as a square wave, a triangular wave, or a sawtooth wave, can be used. The dither signal used to optimize the first manipulated variable u1 and the dither signal used to optimize the second manipulated variable u2 can be set so that a phase difference occurs.

Furthermore, in this embodiment, the extremum control algorithm for optimizing two manipulated variables u1 and u2 has been described, but in the case of optimizing three or more manipulated variables, the phases of the dither signals used for each optimization can be shifted from each other to avoid interference between the responses caused by the movement of each of the two manipulated variables u1 and u2. In this case, the phase difference of the dither signals applied to the multiple manipulated variables must be set according to the number of the manipulated variables.
Next, the effects of the optimum value search control device, the optimum value search control method, and the optimum value search control program of the present embodiment will be described.

11 and 12 show an example of the results of simulating the optimization of a plurality of manipulated variables using the optimum value search control device of the second embodiment.
Here, an example will be described in which the minimum value of the evaluation function y is the optimal operating point, the optimal value of the first manipulated variable u1 is 2, and the optimal value of the manipulated variable u2 is 5. Fig. 11 shows an example of the optimal value search result for the first manipulated variable u1 along the time axis. Fig. 12 shows an example of the optimal value search result for the second manipulated variable u2 along the time axis.

According to the simulation results, it can be seen that the optimum value search control device of the second embodiment can converge to an optimum value within a predetermined time without causing the first manipulated variable u1 and the second manipulated variable u2 to interfere with each other and without causing a sudden, transient change in the manipulated variable. From the above, by applying the algorithm of this embodiment, it is possible to appropriately control the convergence speed in the multivariable extreme value control.
That is, according to the optimum value search control device, the optimum value search control method, and the optimum value search control program of the present embodiment, it is possible to appropriately operate extreme value control that optimizes a plurality of manipulated variables.

Next, the optimal value search control device, optimal value search control method, and optimal value search control program of the third embodiment will be described in detail with reference to the drawings. Note that in the following description, the same components as those in the first embodiment described above will be denoted by the same reference numerals and will not be described.

FIG. 13 is a block diagram illustrating an example of the configuration of an optimum value search control device according to the third embodiment.
The optimum value search control device of this embodiment includes a first search control unit 200 that optimizes a first manipulated variable u1 and a second search control unit 300 that optimizes a second manipulated variable u2, and optimizes a plurality of manipulated variables u1, u2 to be input to the plant 101 using an evaluation function y obtained from the plant 101. The optimum value search control device of this embodiment differs from the first and second embodiments described above in that a first signal applied to the first manipulated variable u1 and a second signal applied to the second manipulated variable u2.

The first dither signal generator 206 and the second dither signal generator 202 output perturbation signals (dither signals) to be added to the manipulated variables u1 and u2. The signal output from the first dither signal generator 206 is input to the multiplier 203. The signal output from the second dither signal generator 202 is multiplied by the amplitude a in the multiplier 207 and then input to the adder 208. The multiplier 203 outputs the product of the evaluation function y passed through the high-pass filter 201 multiplied by the dither signal.

In this embodiment, the first dither signal generator 206 and the second dither signal generator 202 output a sine wave as a dither signal. The waveform of the sine wave output from the first dither signal generator 206 and the waveform of the sine wave output from the second dither signal generator 202 can be set to the same waveform if the time delay and dead time in the plant 101 can be ignored. If the plant 101 has a time delay and dead time, the waveform of the sine wave output from the first dither signal generator 206 and the waveform of the sine wave output from the second dither signal generator 202 may be different by incorporating a predetermined phase compensation or dead time compensation. When the first dither signal generator 206 and the second dither signal generator 202 output the same dither signal, the optimum value search control device may be configured to include one dither signal generator and output a dither signal to the multiplier 203 and the multiplier 207 from the one dither signal generator.

In this embodiment, the second search control unit 300 outputs a dither signal (second signal) from the third dither signal generator 306 and the fourth dither signal generator 302 that is different from the dither signal (first signal) from the first dither signal generator 202 and the second dither signal generator 206, and the second signal is, for example, a sine wave having a different period from the first signal.

The third dither signal generator 306 and the third dither signal generator 302 output a perturbation signal (dither signal) to be added to the manipulated variable u2. The signal output from the third dither signal generator 306 is input to the multiplier 303. The signal output from the fourth dither signal generator 302 is multiplied by the amplitude a in the multiplier 307 and then input to the adder 308. The multiplier 303 outputs the evaluation function y passed through the high-pass filter 301 multiplied by the dither signal.

The third dither signal generator 306 and the fourth dither signal generator 302 output a sine wave as a dither signal. The waveform of the sine wave output from the third dither signal generator 302 and the waveform of the sine wave output from the fourth dither signal generator 306 can be set to the same waveform if the time delay and dead time in the plant can be ignored. If the plant has a time delay or dead time, the waveform of the sine wave output from the third dither signal generator 306 and the waveform of the sine wave output from the fourth dither signal generator 302 may be different by incorporating a predetermined phase compensation or dead time compensation. When the third dither signal generator 306 and the fourth dither signal generator 302 output the same dither signal, the optimum value search control device may be configured to include one dither signal generator and output a dither signal to the multiplier 203 and the multiplier 207 from the one dither signal generator.

FIG. 14 is a diagram illustrating an example of a dither signal used in the optimum value search control device of the third embodiment.
In the optimum value search control device of this embodiment, the period of the dither signal used to optimize the first manipulated variable u1 is set shorter than the period of the dither signal used to optimize the second manipulated variable u2. In this embodiment, by applying dither signals with different periods to each of the multiple manipulated variables u1 and u2, interference between the responses of the multiple manipulated variables u1 and u2 is suppressed. When the first signal and the second signal are set to have the same waveform, interference between the responses of the multiple manipulated variables u1 and u2 occurs as in the first comparative example described above. However, for example, by setting the period of one dither signal longer than the other, the interference between the responses can be avoided. For example, when the period of the sine wave of the first signal is TA and the period of the sine wave of the second signal is TB, the interference between the responses of the multiple manipulated variables u1 and u2 can be avoided by setting TB = 10 x TA.

In addition, in this embodiment, the dither signal applied to the first operation amount u1 and the dither signal applied to the second operation amount u2 are sine waves, but this is not limited to this. The waveform that can be used as the dither signal is not limited to a sine wave, and any signal can be used as long as it is an excitation signal that generates periodic vibration, such as a cosine wave, a square wave, a triangular wave, or a sawtooth wave. The periods of the dither signal applied to the first operation amount u1 and the dither signal applied to the second operation amount u2 can be set to be different.

In addition, in this embodiment, an extreme value control algorithm that optimizes two manipulated variables u1 and u2 has been described. Similarly, when optimizing three or more manipulated variables, the periods of the dither signals used for each optimization can be shifted from each other to avoid interference between the responses caused by moving each of the two manipulated variables u1 and u2. In this case, the period of the dither signal applied to the multiple manipulated variables must be set according to the number of manipulated variables.

Next, the effects of the optimum value search control device, the optimum value search control method, and the optimum value search control program of the present embodiment will be described.
15 and 16 show an example of the results of simulating the optimization of a plurality of manipulated variables using the optimum value search control device of the third embodiment.

Here, we will explain an example in which the minimum value of the evaluation function y is the optimal operating point, the optimal value of the first operation amount u1 is 2, and the optimal value of the operation amount u2 is 5. Figure 15 shows an example of the optimal value search result for the first operation amount u1 along the time axis. Figure 16 shows an example of the optimal value search result for the second operation amount u2 along the time axis.

According to the simulation results, it can be seen that the optimum value search control device of the third embodiment can converge to an optimum value within a predetermined time without causing the first manipulated variable u1 and the second manipulated variable u2 to interfere with each other and without causing a sudden, transient change in the manipulated variable. From the above, by applying the algorithm of this embodiment, it is possible to appropriately control the convergence speed in the multivariable extreme value control.
That is, according to the optimum value search control device, the optimum value search control method, and the optimum value search control program of the present embodiment, it is possible to appropriately operate extreme value control that optimizes a plurality of manipulated variables.

In the optimal value search control device of each of the above-mentioned embodiments, the above-mentioned functions may be realized by hardware or software. In addition, when the functions of the optimal value search control device are realized by software, the optimal value search control device may include at least one processor and a memory in which a program executed by the processor is recorded. In addition, the optimal value search control device may be configured to store setting information such as a dither signal and a normalization processing function applied to each of the multiple operation amounts in a storage device such as a magnetic hard disk device or a semiconductor storage device, and to acquire the setting information from the storage device. In addition, in this case, the optimal value search control device may be configured to include an input unit that accepts a change operation of the dither signal and the normalization processing, and a setting update unit that updates the setting information of the dither signal and the normalization processing in response to the input. Furthermore, in this case, the optimal value search control device may be configured to include a display unit that displays the contents of the setting information.

Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are included in the scope of the invention and its equivalents described in the claims.

101...plant, 200...first search control unit, 201...high pass filter, 202...dither signal generator, 203...multiplier, 204...low pass filter, 205...integrator, 206...dither signal generator, 207...multiplier, 208...adder, 209...normalization processing unit, 300...search control unit, 301...high pass filter, 302...dither signal generator, 303...multiplier, 304...low pass filter, 305...integrator, 306...dither signal generator, 307...multiplier, 308...adder, 309...normalization processing unit

Claims (7)

An apparatus applicable to any target process, which receives a plurality of operation variables including at least a first operation variable and a second operation variable, and outputs one evaluation function value,
a first signal generating unit that receives the evaluation function value measured in real time as an input and generates a first signal for oscillating the first manipulated variable, and a first normalization processing unit that normalizes a gradient of the evaluation function value, and a first search control unit that searches for an operating point of the first manipulated variable at which the evaluation function value becomes an optimal value, based on a response of the evaluation function value caused by oscillating the first manipulated variable by the first signal;
a second signal generating unit configured to receive the evaluation function value as an input and generate a second signal that is different from the first signal and that vibrates the second manipulated variable, and a second normalization processing unit configured to normalize a gradient of the evaluation function value, and a second search control unit configured to search for an operating point of the second manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value resulting from vibrating the second manipulated variable by the second signal,
An optimum value search control device, wherein the first signal and the second signal are periodically oscillating signals having mutually different phases. An apparatus applicable to any target process, which receives a plurality of operation variables including at least a first operation variable and a second operation variable, and outputs one evaluation function value,
a first signal generating unit that receives the evaluation function value measured in real time as an input and generates a first signal for oscillating the first manipulated variable, and a first normalization processing unit that normalizes a gradient of the evaluation function value, and a first search control unit that searches for an operating point of the first manipulated variable at which the evaluation function value becomes an optimal value, based on a response of the evaluation function value caused by oscillating the first manipulated variable by the first signal;
a second signal generating unit configured to receive the evaluation function value as an input and generate a second signal that is different from the first signal and that vibrates the second manipulated variable, and a second normalization processing unit configured to normalize a gradient of the evaluation function value, and a second search control unit configured to search for an operating point of the second manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value resulting from vibrating the second manipulated variable by the second signal,
An optimum value search control device, wherein the first signal and the second signal are periodically oscillating signals having different periods. An apparatus applicable to any target process, which receives a plurality of operation variables including at least a first operation variable and a second operation variable, and outputs one evaluation function value,
a first signal generating unit that receives the evaluation function value measured in real time as an input and generates a first signal for oscillating the first manipulated variable, and a first normalization processing unit that normalizes a gradient of the evaluation function value, and a first search control unit that searches for an operating point of the first manipulated variable at which the evaluation function value becomes an optimal value, based on a response of the evaluation function value caused by oscillating the first manipulated variable by the first signal;
a second signal generating unit configured to receive the evaluation function value as an input and generate a second signal that is different from the first signal and that vibrates the second manipulated variable, and a second normalization processing unit configured to normalize a gradient of the evaluation function value, and a second search control unit configured to search for an operating point of the second manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value resulting from vibrating the second manipulated variable by the second signal,
The first signal is a signal that oscillates during a first vibration period and has a constant value during a first pause period,
The second signal is a signal that oscillates during a second vibration period and has a constant value during a second pause period,
An optimum value search control device, wherein a first vibration period is included in the second rest period, and the second vibration period is included in the first rest period . The method is applied to an arbitrary target process that receives a plurality of operation variables including at least a first operation variable and a second operation variable as inputs and outputs one evaluation function value,
Acquire the evaluation function value measured in real time;
A method for searching for an operating point of the first manipulated variable and the second manipulated variable at which the evaluation function value becomes an optimal value, comprising the steps of:
The searching for the operating point of the first manipulated variable includes generating a first signal for oscillating the first manipulated variable, normalizing a gradient of the evaluation function value, and searching for an operating point of the first manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value caused by oscillating the first manipulated variable by the first signal;
Searching for an operating point of the second manipulated variable includes generating a second signal that is a signal for oscillating the second manipulated variable and is different from the first signal, normalizing a gradient of the evaluation function value, and searching for an operating point of the second manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value caused by vibrating the second manipulated variable by the second signal ;
An optimum value search control method , wherein the first signal and the second signal are periodically oscillating signals having different periods . The method is applied to an arbitrary target process that receives a plurality of operation variables including at least a first operation variable and a second operation variable as inputs and outputs one evaluation function value,
Acquire the evaluation function value measured in real time;
A method for searching for an operating point of the first manipulated variable and the second manipulated variable at which the evaluation function value becomes an optimal value, comprising the steps of:
The searching for the operating point of the first manipulated variable includes generating a first signal for oscillating the first manipulated variable, normalizing a gradient of the evaluation function value, and searching for an operating point of the first manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value caused by oscillating the first manipulated variable by the first signal;
Searching for an operating point of the second manipulated variable includes generating a second signal that is a signal for oscillating the second manipulated variable and is different from the first signal, normalizing a gradient of the evaluation function value, and searching for an operating point of the second manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value caused by vibrating the second manipulated variable by the second signal;
An optimum value search control method, wherein the first signal and the second signal are periodically oscillating signals having different periods. The method is applied to an arbitrary target process that receives a plurality of operation variables including at least a first operation variable and a second operation variable as inputs and outputs one evaluation function value,
Acquire the evaluation function value measured in real time;
A method for searching for an operating point of the first manipulated variable and the second manipulated variable at which the evaluation function value becomes an optimal value, comprising the steps of:
The searching for the operating point of the first manipulated variable includes generating a first signal for oscillating the first manipulated variable, normalizing a gradient of the evaluation function value, and searching for an operating point of the first manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value caused by oscillating the first manipulated variable by the first signal;
Searching for an operating point of the second manipulated variable includes generating a second signal that is a signal for oscillating the second manipulated variable and is different from the first signal, normalizing a gradient of the evaluation function value, and searching for an operating point of the second manipulated variable at which the evaluation function value is an optimal value based on a response of the evaluation function value caused by vibrating the second manipulated variable by the second signal;
The first signal is a signal that oscillates during a first vibration period and has a constant value during a first pause period,
The second signal is a signal that oscillates during a second vibration period and has a constant value during a second pause period,
An optimum value search control method, wherein a first vibration period is included in the second rest period, and the second vibration period is included in the first rest period. 7. An optimum value search control program for causing a computer to execute the optimum value search control method according to claim 4 .