JP7573825B2 - Flow rate control method and flow rate control device - Google Patents

Description

The present invention relates to a flow control method and a flow control device.

Conventionally, database-driven control has been developed that uses a database that stores multiple data sets (information vectors) that include system parameters that represent the state of a control system and control parameters that correspond to the system parameters (for example, non-patent document 1).

ERIT (Estimated Response Iterative Tuning) has also been proposed as a method for predicting input/output data for various control parameters using a set of input/output data, without obtaining a large amount of operational data through experiments, when generating information vectors to be stored in a database (Non-Patent Document 2).

Kenji Takao, Toru Yamamoto, Takao Hinamoto, "Design of Memory-Based PID Controller", Transactions of the Society of Instrument and Control Engineers, Vol. 40, No. 9, pp. 898-905, 2004 Takeo Nakamura, Osamu Kaneko, "Estimated Response Iterative Tuning of Feedforward for Two-Degree-of-Freedom Control Systems Using One-Degree-of-Freedom Control System Data," 60th Japan Joint Conference on Automatic Control, 2017.

When creating a database from input/output data predicted using ERIT, the accuracy of the set of input/output data that forms the basis of the prediction affects the accuracy of the prediction by ERIT. In other words, if the accuracy of the operational data, which is the set of input/output data that forms the basis of the prediction, is low, the accuracy of the predicted input/output data will decrease, and the control performance of database-driven control will decrease. Not only when ERIT is used, but also when generating an information vector based on operational data, the accuracy of the operational data will affect the accuracy of the information vector, so creating a database using noisy operational data will decrease the control performance.

For example, when using the above database-driven control to measure granular grains, powdered flour, highly viscous fluids, etc. based on flow rate, large noise may occur in the output of the sensor used to control the flow rate due to avalanches that occur when the fluid breaks down inside the weighing machine, bridge collapse, vibrations caused by resonance of the load cell, etc. If a database is created using information vectors generated based on operational data that contains a lot of such noise, control will be performed based on parameters that are affected by the noise, which may degrade the control performance.

The present invention was made in consideration of the above circumstances, and aims to provide a flow control method and flow control device with high control performance using a database created by reducing the effects of noise from initial data, which is operational data.

In order to achieve the above object, a flow rate control method according to a first aspect of the present invention comprises:
an initial data acquisition step of acquiring step response data, which is operational data of a control system for controlling a flow rate of a fluid, and smoothing the data by filtering;
a database creation step of creating a database for storing a plurality of information vectors including a system parameter representing a state of the control system and a control parameter representing a characteristic of a controller, the information vectors being generated by data prediction using ERIT based on the step response data smoothed in the initial data acquisition step;
and a control process for controlling the flow rate using database-driven control that updates control parameters based on the similarity between an information vector stored in the database and a demand point that represents the current state of the control system.

The filtering process includes:
smoothing the step response data by linear regression and multiple weighted linear regressions;
This may also be the case.

The filtering process is further
smoothing the step response data with an IIR filter and a zero phase filter;
This may also be the case.

Further, in the control step,
filtering the measured flow rate by a linear regression and a plurality of weighted linear regressions to calculate the demand point;
This may also be the case.

In addition, in the database creation process,
The control parameters are optimized using Adam.
This may also be the case.

In addition, in the database creation step,
and reducing the size of the database by deleting any of the generated information vectors so that the mutual similarity of the plurality of information vectors becomes equal to or less than a predetermined threshold.
This may also be the case.

In addition, in the initial data acquisition step,
The step response data is smoothed by a filtering process that averages a rising flow rate and a falling flow rate corresponding to a gate opening degree that controls the flow rate,
In the database creation step,
Calculating the control parameters based on the gate opening and the average flow rate to generate the plurality of information vectors.
This may also be the case.

A flow rate control device according to a second aspect of the present invention comprises:
a database storing a plurality of information vectors including a system parameter representing a state of a control system for controlling a flow rate of a fluid and a control parameter representing a characteristic of a controller;
a control unit that performs database-driven control to update a control parameter and control the flow rate based on a similarity between the information vector stored in the database and a demand point that represents a current state of the control system;
The information vectors stored in the database include information vectors created by data prediction using ERIT based on data smoothed by filtering step response data, which is operational data.

Also, the step response data is smoothed by linear regression and multiple weighted linear regressions .
This may also be the case .

The step response data is smoothed by an IIR filter and a zero-phase filter .
This may also be the case .

The step response data is smoothed by filtering to average a rise time flow rate and a fall time flow rate corresponding to a gate opening degree that controls the flow rate,
The plurality of information vectors are generated by calculating the control parameters based on the gate opening degree and the average flow rate.
This may also be the case.

The flow control method and flow control device of the present invention generate a database based on filtered step response data, which is operational data, making it possible to reduce the effects of noise in the operational data and improve control performance.

1 is a diagram showing a configuration of a flow rate control device according to a first embodiment of the present invention; FIG. 2 is a block diagram of a control unit according to the first embodiment. 4 is a flowchart showing a flow rate control procedure according to the first embodiment. 4 is a flowchart showing a flow of a filtering process according to the first embodiment. 13 is a graph of operational data showing the change over time in the weight of rice in the weighing tank, where (A) is a graph showing the measured weight and the moving average, and (B) is a graph showing the measured weight and data after filtering by linear regression. 11 is a graph showing an example of a change in flow rate over time before application of an IIR filter, after application of an IIR filter, and after application of a zero phase filter. 11 is a graph showing an example of step response data in which a maximum flow rate is a target flow rate. 13 is a graph showing an example of rice flow rate control using a fixed PID, where (A) is a graph when the target flow rate is 2 t/h, (B) is a graph when the target flow rate is 5 t/h, and (C) is a graph when the target flow rate is 10 t/h. FIG. 1 is a block diagram of a one-degree-of-freedom I-PD control system. FIG. 1 is a block diagram of a two-degree-of-freedom I-PD control system. FIG. 1 is a block diagram of a two-degree-of-freedom I-PD control system having a reference model. 1 is a learning curve showing an example of optimization using Adam. 13 is a graph showing optimized input and output, in which (A) is a graph showing the change over time in the opening degree, which is the control input, and (B) is a graph showing the change over time in the flow rate, which is the output. FIG. 1 is a block diagram of a one-degree-of-freedom control system for optimizing a PID gain. 6A and 6B are graphs relating to optimization of PID gains, in which (A) is a graph showing a learning curve of an error function _J2 , and (B) is a graph showing a change in opening degree over time. 1 is a graph showing an example of rice flow rate control in accordance with embodiment 1, where (A) is a graph when the target flow rate is 2 t/h, (B) is a graph when the target flow rate is 5 t/h, and (C) is a graph when the target flow rate is 10 t/h. 10 is a flowchart showing a flow rate control procedure according to the second embodiment. FIG. 11 is a conceptual diagram showing a method for calculating an opening-specific flow rate according to the second embodiment. 13 is an example of a database according to the second embodiment. 11 is a graph showing an example of rice flow rate control in accordance with embodiment 2, where (A) is a graph when the target flow rate is 2 t/h, (B) is a graph when the target flow rate is 5 t/h, (C) is a graph when the target flow rate is 10 t/h, and (D) is a graph when the target flow rate is 16 t/h.

The flow control method and flow control device according to the embodiment of the present invention will be described below with reference to the drawings.

(Embodiment 1)
1, a flow control device 1 according to this embodiment includes a weighing machine main body 10 that contains rice, which is a fluid, and weighs the outflowing rice, and a control unit 20 that controls the flow rate of the outflowing rice by adjusting the opening of a discharge gate 15 of the weighing machine main body 10. The control system of the flow control device 1 includes a controller and a database DB configured in the control unit 20, and the database DB stores a plurality of sets (information vectors) of system parameters that represent the state of the control system and control parameters that correspond to the system parameters. The control system is a database-driven control system that adjusts the control parameters to control the flow rate of rice based on the current state of the control system and the information vectors in the database DB.

In this embodiment, we will explain an example of a control system that uses a database created by generating multiple information vectors based on operational data using ERIT (Estimated Response Iterative Tuning).

As shown in FIG. 1, the weighing machine main body 10 includes a supply gate 11 that supplies the rice to be weighed to the weighing machine main body 10, a weighing tank 12 that stores the rice, a load cell 13 that measures the weight of the weighing tank 12, a weighing amplifier 14 that amplifies the output signal of the load cell 13 and transmits it to the control unit 20, a discharge gate 15 that allows the rice to flow out of the weighing tank 12, and a motor unit 16 that opens and closes the discharge gate 15.

The supply gate 11 is a supply port that supplies rice to the weighing tank 12 of the weighing machine main body 10. The supply gate 11 may be automatic or manual. By supplying rice appropriately from the supply gate 11, the amount of rice inside the weighing machine main body 10 is maintained at an appropriate amount. Note that in this embodiment, the fluid that is the subject of flow rate control is rice, but is not limited to this and may be granular matter such as grains, powder such as flour, or liquid such as water.

The measuring tank 12 is installed inside the housing of the measuring machine main body 10 and is a tank for storing rice. A discharge gate 15 is provided at the bottom of the measuring tank 12, and rice flows out through the discharge gate 15.

The measuring tank 12 is attached to the housing of the measuring machine main body 10 via a load cell 13. The load cell 13 is also connected to the control unit 20 via a measuring amplifier 14. The output signal of the load cell 13 is amplified by the measuring amplifier 14 and transmitted to the control unit 20. This makes it possible to measure the weight of the rice in the measuring tank 12. The flow control device 1 measures the flow rate of rice and controls the flow rate based on the change over time in the weight of the rice measured using the load cell 13.

The discharge gate 15 is equipped with an on-off valve (not shown). A motor unit 16 is connected to the on-off valve, and the on-off valve opens and closes due to the rotational movement of the motor unit 16. This changes the flow rate of rice, i.e., the amount of rice discharged. Hereinafter, opening and closing of the on-off valve is also referred to as opening and closing of the discharge gate 15. The angle related to the opening and closing movement of the on-off valve is also referred to as the opening degree of the discharge gate 15. The opening degree of the discharge gate 15 is the operation amount of the control system according to this embodiment. There are no particular limitations on the type of motor of the motor unit 16, and it is sufficient if it can perform opening and closing operations while receiving pressure from a fluid such as rice.

The control unit 20 is, for example, a computer device, and includes a control unit 21, a memory unit 22, a display unit 23, and an input unit 24, as shown in the block diagram of FIG. 2.

The control unit 21 is composed of a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), a crystal oscillator, etc., and controls the operation of the flow control device 1. The control unit 21 also determines control parameters based on data (request points) including the opening degree and flow rate of the discharge gate 15, which are parameters that represent the current state of the weighing machine main body 10, and information vectors previously stored in the memory unit 22, and controls the flow rate of rice. The control unit 21 loads various operation programs and data stored in the ROM, memory unit 22, etc. of the control unit 21 into the RAM and operates the CPU, thereby realizing each function of the control unit 21 shown in FIG. 2. As a result, the control unit 21 operates as a flow rate information acquisition unit 211, a calculation unit 212, and a discharge gate control unit 213.

The flow rate information acquisition unit 211 acquires the output signal of the load cell 13 amplified by the measuring amplifier 14. As a result, the flow rate information acquisition unit 211 measures the amount (weight) of rice contained in the measuring tank 12. In addition, the flow rate information acquisition unit 211 calculates the rice flow rate, i.e., the change in the weight of rice per unit time, based on the information acquired from the load cell 13.

The calculation unit 212 calculates control parameters for operating the on-off valve based on the information vector stored in the database DB of the memory unit 22 and the current rice flow rate calculated by the flow rate information acquisition unit 211, etc.

The discharge gate control unit 213 uses the control parameters calculated by the calculation unit 212 to control the opening of the discharge gate 15. More specifically, the calculated control parameters are used to modify the PID gain of the controller to calculate the operation amount, and control the operation of the motor unit 16 that drives the opening and closing valve of the discharge gate 15. This allows the discharge gate 15 to be adjusted to an appropriate opening, and the flow rate of rice is controlled.

The storage unit 22 is a non-volatile memory such as a hard disk or flash memory, and stores the rice flow rate calculated by the flow rate information acquisition unit 211, an algorithm for determining control parameters, and the like. The storage unit 22 also includes a database DB that stores multiple information vectors for determining control parameters. The information vectors in this embodiment include the rice flow rate representing the state of the control system, system parameters such as the opening degree of the discharge gate 15, and control parameters such as PID gains representing the characteristics of the controller.

The display unit 23 is a display device that displays information such as the current rice flow rate, the target flow rate, control parameters, and the amount of rice remaining in the measuring tank 12, and is, for example, a liquid crystal panel attached to the housing of the measuring machine main body 10.

The input unit 24 is an input device such as a keyboard or a mouse that inputs instructions to start and end flow control, and data on each information vector to be stored in the database DB. The input unit 24 in this embodiment is a touch panel disposed on the display unit 23.

Next, a method for controlling the flow rate of a fluid using the flow control device 1 will be described in detail with reference to the flowchart in FIG. 3.

As shown in FIG. 3, in the initial data acquisition process, initial data, which is operational data of the control system, is acquired. More specifically, first, the gate opening is manually adjusted while checking the gate opening at which the flow rate becomes the maximum target flow rate (step S1). The maximum target flow rate is the maximum target flow rate within a range of target flow rates that have been set in advance. For example, in the flow control device 1 according to this embodiment, the gate opening is 61 degrees when the maximum target flow rate is set to 10 t/h.

Based on the gate opening confirmed in step S1, step response data is obtained with the maximum target flow rate as the target value (step S2).

The step response data acquired in step S2 is smoothed by a filtering process (step S3). FIG. 4 is a flowchart showing the details of the filtering process in step S3. As shown in FIG. 4, the filtering process in step S3 includes a first filtering process and a second filtering process. The filtering process will be described in detail below.

In the filtering process according to the present embodiment, first, as a first filtering process, the following linear regression is applied to the step response data acquired in step S2 (step S21). Specifically, α ₁ (t) and β ₁ (t) that minimize the error sum of squares J ₁ (t) shown in the following formula are obtained.
where y ₁ (t) is the predicted weight, f ₁ (t) is the flow rate, and the above linear regression is applied in intervals of one second.

In addition, a weight _w1 is calculated using a prediction error _e1 between the predicted weight _y1 by the linear regression and the acquired data _y0 .
However, thresh ₁ is a constant and can be set arbitrarily by the user.

Furthermore, the calculated weight _w1 is used to apply the weighted linear regression shown below (step S22). Specifically, the weight _w1 is used to find α ₂ (t) and β ₂ (t) that minimize the error sum of squares J ₂ (t) shown in the following formula:
Here, y ₂ (t) is the predicted weight, f ₂ (t) is the flow rate, and the above weighted linear regression is applied in intervals of one second.

In addition, a weight _w2 is calculated using a prediction error _e2 between the predicted weight _y2 by the weighted linear regression and the acquired data _y0 .
However, thresh ₂ is a constant and can be set by the user.

Furthermore, the calculated weight _w2 and predicted weight _y2 are used to apply the weighted linear regression shown below (step S23). Specifically, the weight _w2 is used to find _α3 (t) and _β3 (t) that minimize the error sum of squares _J3 (t) shown in the following formula.
where y ₃ (t) is the predicted weight, f ₃ (t) is the flow rate, and the weighted linear regression is applied to an interval of T(t) seconds. T(t) is the flow rate measurement time shown below, for example, and is determined based on the time of day.

From the above formula, the weighted linear regression is applied with 1 second in the early stage where the flow rate change is large and 3 seconds in the later stage where the flow rate change is small.
This makes it possible to reduce the amount of calculation while maintaining the effect of smoothing by linear regression.

Figures 5 (A) and (B) are graphs showing an example of the change in flow rate over time. As shown in Figure 5 (A), the measured weight contains a large amount of noise, and it is clear that the effects of the noise cannot be sufficiently eliminated by a simple moving average. In contrast, in the data after the first filtering process shown in Figure 5 (B), it is clear that the effects of noise in the measured data have been reduced, resulting in a significant smoothing effect.

In this embodiment, the initial data smoothed by the first filtering process (steps S21 to S23) related to the linear regression described above is further smoothed by a second filtering process to reduce the influence of noise. Specifically, filtering is performed by combining an infinite impulse response (IIR) filter (step S24) and a zero phase filter (step S25).

The transfer function of the zero-order hold is expressed as H(s)=(1-e ^-sTs )/s. The transfer function of the IIR filter is expressed as F(s)=1/(1+ _Tfs ). And the discretized IIR filter is expressed by the following equation.

For example, y _f (t) obtained by applying the above-mentioned IIR filter to the acquired data y ₀ (t) is expressed as follows.

As shown in the above formula, when an IIR filter is applied to time-series response data, a phase lag occurs. This phase lag affects the creation of a predicted response for generating an information vector, PID gain learning, and the like. Therefore, a zero-phase filter is applied to the data after application of the IIR filter to reduce the effect of the phase lag. Data y ₁ (t) to which the IIR filter has been applied is expressed as y ₁ (t) = F(z ^-1 ) y ₀ (t).

Then, y ₁ (t) is inverted with respect to time as shown in the following equation.
y ₂ (t)=y ₁ (N-t)
Here, N is the number of data.

In the data y ₂ (t) inverted with respect to time, the data is shifted to the negative side by the stationary value of the input data so that the data at time 0 becomes 0.
y ₃ (t) = y ₂ (t) - y _{0 stable}
where y _{0 stable} is the average steady value of y ₀ (t).

An IIR filter is further applied to y ₃ (t) as shown in the following equation.
y ₄ (t)=F(z ⁻¹ )y ₃ (t)

The data y ₄ (t) after application of the IIR filter is inverted with respect to time to calculate y ₅ (t). Then, the data at time 0 of y ₅ (t) is shifted to the positive side by the steady value of the input data so that it becomes 0.
y ₅ (t)=y ₄ (N-t)
y _f (t)=y ₅ (t)+y _{0 stable}

The above algorithm makes it possible to obtain a filter applied waveform _yf (t) without phase delay. Fig. 6 is a graph showing the flow rate before and after filtering processing, showing examples of data before application of the IIR filter, data after application of the IIR filter, and data after application of the zero phase filter. As shown in Fig. 6, it can be seen that by applying the IIR filter and the zero phase filter, smoothed data without phase delay and with high frequency components cut off is obtained.

Next, returning to the flowchart of FIG. 3, the filtered step response data is used to calculate control parameters for creating a fixed PID controller (step S4). Specifically, the system gain K and dead time L are calculated from the filtered step response data.

Figure 7 is a graph showing an example of step response data. From Figure 7, the tangent of the flow rate at 63.2% ± 20% of the steady-state flow rate is found, and the time it takes for the tangent to change from a flow rate of 0 to a flow rate of 63.2% is defined as the time constant T. The system gain K is the steady-state flow rate Y divided by the steady-state opening degree U. The dead time L is the time from the time the on-off valve is operated to start opening the discharge gate 15 to the time the tangent of the flow rate intersects with the flow rate of 0.

The PID gain is calculated using the CHR method shown in the following equation from the calculated time constant T, system gain K, and dead time L. In this embodiment, the overshoot in the CHR method is set to 20%.
K _c =0.6T/KL, T _i =T, K _p =K _c , K _i =K _c /T _i・T _s , K _d =0

In this embodiment, in order to compare the control performance, a medium flow rate y _mid and a small flow rate y _min are calculated, and the PID gain of the fixed PID is calculated for each. The medium flow rate y _mid is a flow rate where the time constant T _mid is T _mid = T _max × 0.632 when the time constant for the maximum flow _rate y max is T _max . The small flow rate y _min is a flow rate where the time constant T _min is T _min = T _max × 0.632 ² .

In step S5, step response data (initial data) is obtained using the fixed PID controller with the control parameters calculated in step S4 (Figures 8 (A) to (C)).

The initial data, which is the step response data acquired in step S5, contains various noises, as shown in the measured weight graph in FIG. 5(A). For example, in the fluid flow control device 1 according to this embodiment, the avalanche phenomenon that occurs when rice, which is a fluid, is discharged, the impact of bridge collapse, etc., affects the output signal of the load cell 13 as large noise.

When generating an information vector based on initial data, there is a risk that noise may affect the information vector, resulting in a decrease in control performance. Therefore, in this embodiment, a filtering process is applied to the initial data, which is the acquired operational data, to smooth the data (step S6).

The filtering process in step S6 is similar to the filtering process in step S3 (steps S21 to S25 in FIG. 4), and includes a first filtering process using linear regression and a second filtering process using an IIR filter and a zero-phase filter. This filtering process smoothes the initial data and reduces the effects of noise.

Then, the process proceeds to the database creation step, where information vectors are generated based on the data smoothed by the filtering process in step S6 (step S7). In this embodiment, by predicting the response using ERIT, multiple information vectors including system parameters and control parameters are generated from the filtered initial data, which is a set of input/output data. The method for generating information vectors using ERIT is described below.

In this embodiment, the controller realized by the control unit 21 is a controller using a proportional action leading PI control law (I-PD controller). This control system is expressed as a one-degree-of-freedom control system using the initial data manipulated variable (opening degree of the discharge gate 15) u ₀ and the output (flow rate of rice) y ₀ , as shown in the block diagram of FIG. 9 and the following equation.

Then, y _0f and u _0f to which the filter is applied are expressed by the following equations.

Moreover, a block diagram of a two-degree-of-freedom I-PD control system including the feedforward C _r and the filter F _z (z ⁻¹ ) is expressed as shown in FIG. 10, and the manipulated variable u and output y are expressed by the following equations.

Moreover, y _f and u _f to which the filter is applied are expressed as follows.

From the above formula, the following formula is obtained.

Therefore, the predicted output y _f ^* (t) and predicted input u _1f ^* (t) by ERIT are expressed as follows:

Here, if a model that realizes a desired response is G _m and the dead time is d, the desired response y _r is expressed by the following equation.

Furthermore, when the desired response is applied to a two-degree-of-freedom control system and the model of the controlled object estimated from the closed-loop transfer function of the two-degree-of-freedom control system and the reference model _Gm is denoted as G^, the block diagram is represented in FIG. 11 and the following equation holds:

Furthermore, the following desired manipulated variable u ₂ ^* (t) can be derived from the desired response y _r (t) and the reference model G _m .

Moreover, the input u ₂ ^* _f (t) and output y _rf (t) to which the filter is applied are expressed by the following equations.

By minimizing the error between the output of the above control system and a desired output, C _r ^* (t) can be optimized. The update equation for _{C r} ^* (t) and the error function J ₁ are expressed by the following equations.

In this embodiment, Adam is used as the optimization algorithm, and optimization is performed by successively updating C _r (t) so as to minimize the error function J _1. Specifically, the optimization algorithm is expressed by the following update formula.

The values of the parameters α, ε, β ₁ , and β ₂ are not particularly limited, and in this embodiment, α=0.001, ε=1 ⁻¹⁰ , β ₁ =0.9, and β ₂ =0.999.

By using Adam, it is possible to make the value of _J1 converge quickly in appropriate increments compared to when the learning rate η is set to a fixed value (FIG. 12).

Figure 13(A) is a graph showing examples of control inputs u ₁ ^* _f (t) and u ₂ ^* _f (t) optimized by learning, and Figure 13(B) is a graph showing examples of reference trajectory learning y _rf (t) and optimized y _f ^* (t).

As described above, _Cr for a desired response can be calculated using ERIT, and multiple input/output responses can be derived from a set of initial data. The following information vector φ(j) is generated from the system parameters ^φ- (j) including the derived input/output and the corresponding control parameters K(j).
Here, j=1, 2, . . . , N, and N indicates the number of data.

Next, the control unit 21 optimizes the PID gain, which is a control parameter of the generated information vector (step S8). When a one-degree-of-freedom control system with C _r =0 is constructed using the input/output responses u ₂ ^* _f (t) and y _rf (t), it is expressed as shown in the block diagram of Fig. 14. In this case, the manipulated variable u _PID (t) is expressed by the following equation.

The _PID gains _KP ^* (t), KI*( _{t), and KD*(t) can be optimized by minimizing the error function J2=f(u2*f} ^, _{uPID ) between uPID} ⁽ _{t )} and _u2 ^{* f} (t) in the above equation. The update equations for the PID gains _KP ^* (t), _KI ^* (t), and _KD ^* (t) are as follows:

In this embodiment, Adam is used as an optimization algorithm, and optimization is performed by successively updating _KP ^* (t), _KI ^* (t), and _KD ^* (t) so as to minimize the error function _J2 . By using Adam, the value of the error function _J2 can be converged quickly in appropriate increments compared to the case where the learning rate η ( _ηP , _ηI , _ηD ) is a fixed value.

Fig. 15(A) is a graph showing an example of a learning curve of the error function _J2 , and Fig. 15(B) is a graph showing an example of u _PID (t) after learning. From Figs. 15(A) and (B), it can be seen that the error function _J2 converges appropriately, and u _PID (t) that approximately coincides with u ₂ ^* _f (t) is calculated. By the above method, the PID gains K _P ^* (t), _KI ^* (t), and K _D ^* (t) can be optimized.

Next, the control parameters _KP ^* (t), _KI ^* (t), and _KD ^* (t) obtained by learning the PID gains are stored in the database DB as an updated information vector φDB(t) together with the system parameters target value (flow rate) r(t), input (opening) _u2 ^* _f (t), and reference trajectory _yrf (t), to create the database DB (step S9). The information _{vector φDB (} t) is expressed as follows using the system parameter ^φ- (t) and the control parameter K(t):

In this embodiment, the database DB created above is made lighter (step S10). Specifically, similarity is calculated between information vectors stored in the database DB, and those with high similarity are deleted from the database DB. For example, the similarity S between the j-th information vector φ _DB (j) and the j+1-th information vector φ _DB (j+1) stored in the database DB is defined by the following formula:
Here, h _i is the bandwidth, φ(j,i) is the i-th element of the j-th information vector, and n _y and n _u are the degrees of the output and input, respectively.

When the similarity S is high, i.e., when the similarity S is equal to or greater than a predetermined threshold, one of the information vectors (φ(j) or φ(j+1)) is deleted. In other words, the information vectors are deleted so that the mutual similarity between the information vectors is equal to or less than a predetermined threshold, thereby reducing the weight of the database DB. This makes it possible to delete information vectors that have little effect when performing flow control, thereby reducing the amount of calculations during control while suppressing deterioration in control performance.

Next, flow rate control is performed using the created database DB as the control process shown in the flowchart of Fig. 3. When the rice to be measured is placed in the measuring tank 12, parameters such as the target flow rate that is the target value r are input from the input unit 24, and then an instruction to start control is input, control of the flow rate control device 1 is started (step S11). The control unit 21 generates a request point (query) φ-( ^t ) that represents the current state of the control system based on the measurement data acquired from the load cell 13 and the opening sensor of the discharge gate 15.

More specifically, filtering is applied to the measurement data acquired from each sensor (step S12). The filtering in step S12 is the same as the first filtering in step S3, i.e., smoothing by linear regression. This makes it possible to reduce the influence of noise with a small amount of calculation. Then, the required point φ ⁻ (t) is calculated based on the data filtered in step S12 (step S13).

The calculation unit 212 calculates the similarity S between the information vector φ _DB and the request point φ ⁻ (t) stored in the database DB, more specifically, the similarity S between the system parameter φ _DB ⁻ of the information vector φ _DB and the request point φ ⁻ (t) (step S14). The similarity S (φ ⁻ (t), φ _DB ⁻ (j)) between the request point φ ⁻ (t) and the system parameter φ _DB ⁻ (j) of the j-th information vector φ _DB (j) stored in the database DB is the same as the similarity S used in the weight reduction of the database DB described above, and is expressed by the following formula.

Here, h _i is the bandwidth, and φ _DB ⁻ (j, i) represents the i-th element of the system parameter in the j-th information vector stored in the database DB.

There are various methods for determining the bandwidth h _i , but in this embodiment, the following plug-in method is used.

The standard deviation σ _i in the above formula is defined as follows using the number of data N:
Here, μ _i is the average value of φ ⁻ _i .

The similarity S is highest when there exists an information vector φ _DB (j) having the exact same system parameter φ _DB ⁻ (j) as the demand point φ ⁻ (t) representing the current system state. In this case, the similarity S is expressed by the following formula.

Furthermore, if the request point φ ⁻ (t) and the system parameter φ _DB ⁻ (j) of the information vector φ _DB (j) are not similar, the similarity S(φ ⁻ (t), φ _DB ⁻ (j)) approaches 0. Therefore, the calculation unit 212 according to this embodiment selects the system parameter φ _DB ⁻ (j) of the information vector φ _DB (j) that satisfies the following conditional expression based on the similarity S as the neighboring data (step S15).
Here, _Tth is a threshold value, which is set in the range of 0≦ _Tth ≦1.

Next, using the control parameter K(j) of the information vector φ _DB (j), which is the neighborhood data selected in step S15, the PID gains constituting the local controller are calculated by the following Linearly Weighted Average (LWA) method (step S16).
Here, n _k is the number of neighboring data selected in step S15, and is automatically determined from the above formula.

Furthermore, w _i is a weight for the PID gain K(i) included in the i-th information vector φ _DB (i) selected as the neighborhood data, and is given by the following equation.

By following the above steps, the PID gains at each time can be calculated based on the similarity S.

The discharge gate control unit 213 uses the control parameters calculated in step S16 to control the discharge gate 15 (step S17). Note that if there is no data similar to the request point φ ⁻ (t) in the system parameters φ _DB ⁻ (j) of the information vector φ _DB (j) stored in the database DB, that is, if n _k =0, K(t) cannot be calculated, and in this case, in this embodiment, the PID gains of the information vector φ _DB (j) with the highest similarity S are adopted.

Then, steps S12 to S17 are repeated for each time step until the control ends (NO in step S18). When a termination condition is met, such as the end of a predetermined time step or an end instruction from the operator (YES in step S18), the control ends.

As described above, according to the flow control method and flow control device of this embodiment, the initial data is filtered by linear regression and weighted linear regression to generate an information vector, and a database DB is created, so that it is possible to reduce the effects of noise and improve control performance. In addition, in this embodiment, the initial data is filtered by applying an IIR filter and a zero-phase filter, so that it is possible to further reduce the effects of noise and improve control performance without causing a time delay.

In addition, in this embodiment, the step response data at the maximum target flow rate for calculating the control parameters for the fixed PID controller is filtered by linear regression and weighted linear regression to calculate the control parameters for the fixed PID controller, so it is possible to reduce the effects of noise and improve control performance. In addition, the step response data at the maximum target flow rate is filtered by applying an IIR filter and a zero-phase filter, so it is possible to further reduce the effects of noise and improve control performance without causing a time delay.

Furthermore, in this embodiment, the flow rate data in the control process is filtered by linear regression and weighted linear regression to calculate the required point φ ⁻ (t), so that it is possible to further reduce the influence of noise and improve control performance.

In this embodiment, in generating the information vector φ by ERIT, the feedforward C _r and the PID gain are learned using Adam, which allows the learning coefficient to be automatically determined taking into account the filter, dead time, etc., and allows for more efficient parameter optimization.

In addition, in this embodiment, similar information vectors are deleted using the similarity S to reduce the size of the database DB, thereby efficiently reducing the size of the database DB while preventing a decrease in control performance, and reducing the amount of calculations required during control.

In this embodiment, an IIR filter and a zero-phase filter are applied to the initial data to which linear regression and weighted linear regression have been applied, but this is not limited to the above, and an IIR filter and a zero-phase filter may also be applied to the initial data. This simplifies the calculations in the filtering process, while reducing the effects of noise and improving control performance.

In addition, in this embodiment, weighted linear regression is performed twice, but this is not limited to this, and weighted linear regression may be performed once or three or more times.

(Numerical example)
Figures 16(A) to (C) are examples of rice flow rate control using the flow rate control device 1 of this embodiment, where Figure 16(A) is a graph when the target flow rate is a small flow rate of 2 t/h (tons per hour), Figure 16(B) is a graph when the target flow rate is a medium flow rate of 5 t/h, and Figure 16(C) is a graph when the target flow rate is a large flow rate of 10 t/h.

As shown in Figures 16 (A) to (C), the flow rate control method according to this embodiment allows the target flow rate to be reached in about 10 to 15 seconds for each flow rate. With the fixed PID control of Figures 8 (A) to (C), the time to reach the target flow rate is about 60 seconds for a small flow rate, about 45 seconds for a medium flow rate, and about 40 seconds for a large flow rate. In comparison, the flow rate control according to this embodiment allows the target flow rate to be reached more quickly, demonstrating improved control performance.

(Embodiment 2)
In the first embodiment, the initial measurement data and the like are filtered using linear regression, weighted linear regression, an IIR filter, and a zero-phase filter, but the smoothing of the operational data in the filtering process can be realized by various filters. In the present embodiment, a case will be described in which the filtering process is an averaging process of data at the rise and fall of a step response.

The configuration of the flow control device 1 in this embodiment is the same as in embodiment 1, so the same reference numerals are used and detailed description is omitted.

The flow rate control method according to this embodiment will be described below with reference to the flowchart in FIG. 17. As shown in FIG. 17, in the initial data acquisition step, initial data, which is the operating data of the flow rate control device 1, is acquired. The initial data is step response data when the discharge gate 15 is opened to the maximum degree, and includes data at the time of rising and falling. In other words, the opening and closing valve of the discharge gate 15 is opened to the maximum degree, and after the flow rate has stabilized, the data until the opening and closing valve is closed and the flow rate becomes 0 is acquired as the initial data (step S31).

Next, as a filtering process, the initial data acquired in step S31 is smoothed to estimate the flow rate by opening degree (step S32). In the flow control device 1 according to this embodiment, a flap-type on-off valve driven by a motor unit 16 is used. In step response data when such an on-off valve is used, the flow rate at a certain opening during the rise is generally a smaller flow rate than the steady flow rate at the same opening. Also, the flow rate at a certain opening during the fall is a larger flow rate than the steady flow rate at the same opening.

Therefore, in this embodiment, the average flow rate at the rise and fall times for each opening is calculated at a predetermined interval and smoothed to estimate the flow rate for each opening (Figure 18).

Next, in the database creation process, a system gain and a control parameter are calculated based on the flow rate for each opening degree calculated in step S32 (step S33). The system gain SG is calculated by the following formula.
System gain SG = flow rate y / opening u

The control parameters are calculated using the above system gain SG. In this embodiment, a proportional gain _Kp and an integral gain _Ki are set as the control parameters. The proportional gain _Kp and the integral gain _Ki are calculated using the system gain SG by the following equations.
Here, Ratio is a constant, and _Ts is the sampling time.

Next, an information vector φ _DB is generated based on the calculated control parameters, and a database DB is created (step S34). Fig. 19 shows a part of an example of the information vector φ _DB stored in the database DB. As shown in Fig. 19, in this example, the constant Ratio is set to 0.3, and the information vector φ _DB includes the flow rate and the opening degree, which are the system parameters φ _DB ^- , and the proportional gain K _p and the integral gain K _i , which are the control parameters K.

Next, in the control process, the flow rate of rice is controlled by the flow control device 1. The flow rate control method in the control process is the same as steps S11 to S18 in the first embodiment. That is, when control is started (step S35), the control unit 21 measures the gate opening, the flow rate, etc., and generates the required point φ ⁻ (t).

More specifically, after applying filtering to the measurement data acquired from each sensor (step S36), the required point φ ⁻ (t) is calculated (step S37). The filtering process in step S36 is the same as the first filtering process in step S3 of the first embodiment, that is, smoothing by linear regression. This makes it possible to reduce the influence of noise with a small amount of calculation.

Then, the control unit 21 calculates the similarity S between the system parameter φ _DB ^- (j) of the information vector φ _DB (j) stored in the database DB and the demand point φ ^- (t) (step S38). Then, the control unit 21 extracts the information vector φ _DB (j) (neighborhood data) having a similarity S equal to or greater than a predetermined threshold (step S39), and calculates a control parameter using a weight based on the similarity S (step S40). Then, the control unit 21 controls the gate opening using the calculated control parameter to control the rice flow rate (step S41), and repeats the above control operation until a termination condition is satisfied (step S42).

As described above, in the flow rate control method according to the present embodiment, a filtering process is performed to average the rising and falling edges of the step response data, which is the initial data, to generate the information vector φ _DB , and to create the database DB. Therefore, it is possible to construct a database-driven control system that is less susceptible to noise by smoothing the initial data using a simpler method.

(Numerical example)
Figures 20(A) to (C) are examples of rice flow rate control by the flow rate control device 1 according to this embodiment, where Figure 20(A) is a graph when the target flow rate is a small flow rate of 2 t/h, Figure 20(B) is a graph when the target flow rate is a medium flow rate of 5 t/h, Figure 20(C) is a graph when the target flow rate is a large flow rate of 10 t/h, and Figure 20(D) is a graph when the target flow rate is an even larger flow rate of 16 t/h.

As shown in Figures 20 (A) to (D), it can be seen that the flow rate control method according to this embodiment allows the target flow rate to be reached in about 10 to 15 seconds for each flow rate. Compared to the fixed PID control of Figures 8 (A) to (C), the flow rate control according to this embodiment allows the target flow rate to be reached more quickly, and it can be seen that control performance has been improved. Furthermore, even in the case of a larger flow rate, such as a target flow rate of 16 t/h, it can be seen that the flow rate quickly converges to the target value and flow rate control is performed appropriately.

The present invention is suitable for flow control systems in which the system gain changes significantly with changes in input (opening). It is particularly suitable for flow control systems in which noise due to the characteristics of the fluid is significantly affected when acquiring operational data.

1 Flow control device, 10 Measuring machine body, 11 Supply gate, 12 Measuring tank, 13 Load cell, 14 Measuring amplifier, 15 Discharge gate, 16 Motor unit, 20 Control unit, 21 Control unit, 211 Flow information acquisition unit, 212 Calculation unit, 213 Discharge gate control unit, 22 Memory unit, 23 Display unit, 24 Input unit, DB Database

Claims (11)

an initial data acquisition step of acquiring step response data, which is operational data of a control system for controlling a flow rate of a fluid, and smoothing the data by filtering;
a database creation step of creating a database for storing a plurality of information vectors including a system parameter representing a state of the control system and a control parameter representing a characteristic of a controller, the information vectors being generated by data prediction using ERIT based on the step response data smoothed in the initial data acquisition step;
and a control step of controlling the flow rate by a database-driven control method that updates a control parameter based on a similarity between an information vector stored in the database and a demand point that represents a current state of the control system.
A flow rate control method comprising: The filtering process includes:
smoothing the step response data by linear regression and multiple weighted linear regressions;
2. The method of claim 1, wherein the flow rate is controlled by the first and second electrodes. The filtering process includes:
smoothing the step response data with an IIR filter and a zero phase filter;
3. The flow rate control method according to claim 1 or 2 . In the control step,
filtering the measured flow rate by a linear regression and a plurality of weighted linear regressions to calculate the demand point;
4. The flow rate control method according to claim 1, wherein the flow rate control method is a flow rate control method for controlling a flow rate of the fluid. In the database creation process,
The control parameters are optimized using Adam.
5. The flow rate control method according to claim 1, wherein the flow rate control method is a flow rate control method for controlling a flow rate of the fluid. In the database creation step,
and reducing the size of the database by deleting any of the generated information vectors so that the mutual similarity of the plurality of information vectors becomes equal to or less than a predetermined threshold.
6. The flow rate control method according to claim 1, In the initial data acquisition step,
The step response data is smoothed by a filtering process that averages a rising flow rate and a falling flow rate corresponding to a gate opening degree that controls the flow rate,
In the database creation step,
Calculating the control parameters based on the gate opening and the average flow rate to generate the plurality of information vectors.
2. The method of claim 1, wherein the flow rate is controlled by the first and second electrodes. a database storing a plurality of information vectors including a system parameter representing a state of a control system for controlling a flow rate of a fluid and a control parameter representing a characteristic of a controller;
a control unit that performs database-driven control to update a control parameter and control the flow rate based on a similarity between the information vector stored in the database and a demand point that represents a current state of the control system;
The information vector stored in the database includes an information vector created by data prediction using ERIT based on data smoothed by filtering step response data, which is operation data.
A flow rate control device. The step response data is smoothed by linear regression and multiple weighted linear regressions .
The flow control device according to claim 8 . The step response data is smoothed by an IIR filter and a zero phase filter .
10. The flow rate control device according to claim 8 or 9. The step response data is smoothed by filtering to average a rise time flow rate and a fall time flow rate corresponding to a gate opening degree that controls the flow rate,
The plurality of information vectors are generated by calculating the control parameters based on the gate opening degree and the average flow rate.
The flow control device according to claim 8 .