JPH0511806A - Process dynamic characteristic learning device - Google Patents

Abstract

PURPOSE:To improve the fixing ability for process dynamic characteristics and reverse dynamic characteristics and to secure excellent and stable control characteristics by determining the network constitution of an NN and the initial value of a total weight coefficient according to a specific primary delay recurrence formula. CONSTITUTION:When a process is estimated or approximated with the primary delay of y={K/(1+T.s)}.X previously and the dynamic characteristics of the process are learnt by using a neural network, the neural network is constituted with the primary delay recurrence formula based upon a calculation period DELTAt shown by an equation I. The coefficient of the equation I is given as the initial value of the coefficient of coupling weight between respective neuron elements when the dynamic characteristics of the process begin to be learnt in the same meaning. In the equation I, X is a process input, (y) a process output, K a process gain estimated value, T a process time constant estimated value, (s) a Laplacean, and a suffix (n) a variable at a point n.DELTAt of time.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a process dynamic characteristic learning device for identifying a dynamic characteristic or a reverse dynamic characteristic of a process by utilizing a learning ability of a neural network and reflecting the result in control, and particularly to a process dynamic characteristic learning device. The present invention relates to a process dynamic characteristic learning device that determines the optimal configuration of a neural network and the initial value of the coupling weight coefficient between each neuron element after grasping the state.

[0002]

2. Description of the Related Art Among process control methods, there are methods such as adaptive control and Smith-Muda time compensation control which identify the dynamic characteristics and reverse dynamic characteristics of a process and reflect the results in the control.

By the way, in recent years, in identifying the dynamic characteristics of a process, it is greatly expected to be realized by utilizing the learning ability of a neural network (hereinafter abbreviated as NN). This NN is a coefficient that determines a large number of neuron elements that have been modeled as a model of nerve cells (neurons) such as the human brain, and the coupling weight coefficient for each neuron element so that a desired output can be obtained from these neuron elements. It is configured by a learning means, and each neuron element realizes a higher-order function as a whole while performing a simple operation. In particular, the NN can perform learning while changing the connection weighting coefficient between each neuron element, and the learning ability can change the environment by N.
It can be flexibly adapted to N.

The configuration model of this NN is roughly divided into a hierarchical network and an interconnected network. The former hierarchical NN is, as shown in FIG.
The intermediate layer b and the output layer c are composed of three parts. Each of the layers a, b, and c is composed of a plurality of neuron elements, and two or more intermediate layers b may be provided as necessary. The signal path of this hierarchical NN is considered to be a feedforward type because it flows from the input side to the output side as shown by the arrow in the figure, and there is no coupling between neuron elements in the same layer. On the other hand, the latter interconnected NN
Unlike the hierarchical type NN, there is no distinction of layers, neuron elements can be freely coupled to each other, and a signal path may be input from an output side to an input side, and is therefore considered to be a feedback type.

Usually, this type of hierarchical NN learning algorithm operates so as to set the coupling weight coefficient between neuron elements to an appropriate value by repeatedly presenting an example of a desirable transformation to the NN. Among them, the algorithm of back propagation (back propagation) learning rule is widely used. This back-propagation learning rule is that after introducing an appropriate input signal into the input layer,
NN obtained by sequentially converting signals by the neuron elements of each layer
The difference between the output result and the desired output, that is, the processing for sequentially correcting the connection weighting coefficient to the neuron element of each layer so that the error information becomes small is performed.

Further, the back propagation learning rule will be specifically described with reference to FIG. This NN
The first stage is the input layer a, the second stage is the intermediate layer b, and the third stage is the output layer c. From the j-th neuron element of the first stage to the i-th neuron element of the second stage, The connection weight coefficient is w
_{i, j} ^2,1 , And the coupling weight coefficient from the i-th neuron element in the second stage to the k-th neuron element in the third stage is w
_{k, i} ^3,2 It is represented by. In the above NN, the back propagation learning rule is that the output of the k-th neuron element in the output layer is y _k , and the desired output (teacher signal) of the element k with respect to the input signal at that time is d _k. Then, the squared error E is

[0007]

[Equation 3]

It is represented by Therefore, the back-propagation learning rule is to minimize the squared error E, that is, to change the coupling weight coefficient to the neuron element so that the actual output is as close to the desired output as possible. The algorithm for correcting the coupling weight coefficient is performed by using a kind of steepest descent method regarding the squared error E. As described above, the learning of the coupling weight coefficient to the neuron element progresses from the output layer to the input layer, contrary to the propagation of the signal, and this is also the origin of the name of the back propagation (back propagation) learning rule.

A typical NN used in a process control system is a hierarchical NN, but this type of NN is limited to expressing a static space pattern.
As it is, the essential system dynamics cannot be directly expressed due to the process control problem. So usually,
As shown in the example of FIG. 13, there is a method of expressing dynamics by preparing n delay devices and inserting past n data strings (time series patterns) in parallel between the input layer neuron elements. It is taken. In the example of FIG. 13, n = 7
And the delay devices are 1 ₁ to ₁₇ . a1 to a8 are input layer neuron elements, b1 to b4 are intermediate layer neuron elements, and c1 is an output layer neuron element.

There are many concrete application methods of this NN to the process control system, and the following two examples can be given as examples. One is that if NN is used to obtain a complete inverse dynamics model Gp ^-1 (s) of the transfer function Gp (s) of the process, as shown in FIG. It becomes possible to apply NN2 to the side. In this case, since the entire transfer function becomes "1", the target value R (t) = control amount Y (t), and the control amount can be made to quickly follow the target value. But in practice,
Since it is difficult to make a perfect inverse model, a method of adding a feedback loop in some form to absorb imperfections of the NN inverse model is used. Although various methods are conceivable for this, the method shown in FIG. 15 is often used. In this method, as shown in the upper part of the figure, the process control amount as an input signal is taken into NN3, and the coupling weight coefficient of NN3 is changed based on the difference between the output obtained by the NN3 and the teacher signal which is the operation amount. While learning the inverse model (reverse dynamics) Gp ⁻¹ (s) of the process.
The learned NN3 is applied as a feedforward compensator 3'as shown in the lower part of the figure.

The other one is to apply the NN to the identification of the dynamic characteristic (forward dynamics model) Gp (s) of the process in addition to learning and applying the above inverse model of the process, and the result Attempts have been made to reflect this in control. For example, there is a method used for process identification in the Smith method of waste time compensation control as shown in FIG. 16 (Japanese Patent Application No. 1-321295). This control device comprises a main control unit 4 such as PID control and a Smith method process model unit 5 using an NN. Process control is performed so that the Smith condition is satisfied in this process model unit 5 and process control with a dead time is performed. This is a configuration in which the system is equivalently approximated to a process control system with no dead time, the entire process is simplified, and the actual process 6 is controlled.

[0012]

Therefore, in the above-mentioned NN, the learning ability of the NN for identifying the process becomes an important point in applying the process control. However, the following problems have been pointed out in the case of the back-propagation learning rule, which is often used as a learning algorithm for NNs. 1. It is necessary to repeat learning many times before the output result of the NN converges to the desired output.

2. Although the method of learning an appropriate coupling weight coefficient is used by using the steepest descent method regarding the squared error E,
In this learning process, there is a possibility of convergence in the state of local minimum before reaching the minimum value, and there is another feeling in the learning ability.

One of the causes of the above problem is that the initial value of the coupling weight coefficient of each neuron element must be given at random when starting learning using NN. That is, when the NN is applied to the process control system, even if the characteristics of the process can be estimated to some extent, there is no guideline for associating this estimated value with the NN coupling weight coefficient. This is because the initial value of the coefficient must be given randomly. As a result, the starting point of the multidimensional weighted phase space and the true state point to be reached are far apart.

This means that not only the number of learnings is increased, but also the point which is not the true state point (the point of minimum error), that is, the probability of falling into the local minimum is high. In order to prevent this, it is necessary to prepare many teacher patterns and repeat learning many times. By the way, Figure 1
In the case of the configuration with the number of neuron elements such as 3, usually, it is necessary to repeatedly perform learning about 200 times for one teacher signal for convergence.

Further, since there is no guideline as to what kind of NN structure should be adopted when applying the NN to various processes, it is unclear the appropriate number of neuron elements in the input layer and the intermediate layer. This also relates to the delay time. Therefore, the number of neuron elements must be decided by trial and error.

The present invention has been made in view of the above situation, and it is possible to determine the initial values of the network configuration and the coupling weight coefficient of the NN while grasping the characteristics of the process, and thereby the ability to identify the process dynamic characteristics and the reverse dynamic characteristics. Can increase
It is an object of the present invention to provide a process dynamic characteristic learning device capable of ensuring better and stable control characteristics and expanding the range of application to NN control.

[0018]

First, in the invention corresponding to claim 1, the process is performed in advance with y = {k / (1 + T · s)}.
In the process dynamic characteristic learning device that learns the dynamic characteristic of the process using a neural network when estimated or approximated by the first-order lag of x,

[0019]

[Equation 4]

The neural network is constructed on the basis of the first-order lag recurrence formula based on the calculation period of Δt, and as the initial value of the coupling weight coefficient between the neuron elements when learning the dynamic characteristics of the process is started. It is a process dynamic characteristic learning device that gives the coefficient of the first-order lag recurrence formula. In the above equation, x is a process input, y is a process output, k is a process gain estimated value, T is a process time constant estimated value, and s is a Laplace operator. The suffix n represents a variable at the time point n · Δt.

Next, in the invention corresponding to claim 2, the process is estimated or approximated by the first-order delay + the dead time, and in this case, the neural network of claim 1 has a new waste time. This is the addition of a neuron element for use.

Furthermore, the invention corresponding to claim 3 is a professional
Set y = {k / (T² ・ S² + 2τ ・ T ・ s +
1)} · x is estimated or approximated by the secondary delay of x
The dynamics of the process using a neural network
, But with the neural network at this time
Is z_{n + 1}= Z_n+ △ z y_{n + 1}= Y_n+ △ y Δz = (k · x_n-2, τ, T, z_n-Y_n) ・ (△ t
/ T² ) △ y = z_n・ △ t

The quadratic value is used as the initial value of the connection weighting coefficient between the neuron elements when the learning is performed based on the quadratic delay recurrence formula based on the calculation period of Δt. It is to be a process dynamic characteristic learning device which gives a coefficient of a delay recurrence formula. In the above equation, τ
Is the process damping ratio.

Further, the invention corresponding to claim 4 is a case where the process is estimated or approximated by a second-order lag + waste time. In this case, the neural network of claim 3 has a new waste time. This is a configuration in which a dedicated neuron element is added. Further, in the invention corresponding to claim 5, when the process is estimated or approximated in advance with a secondary delay or a temporary delay, the inverse dynamic characteristic y = k '(1
+ T ₁ · s + T ₂ · s ² ) .X in a process dynamic characteristic learning device for learning using a neural network,

[0025]

[Equation 5]

The neural network is constructed on the basis of the recurrence formula of the calculation cycle of Δt, and the initial value of the connection weighting coefficient between the neuron elements is set as the initial value when learning the inverse dynamic characteristic of the process is started. 1 is a configuration of a process dynamic characteristic learning device that gives a coefficient of a recurrence formula. In the above equation, x is the process output, y is the process input, k ′
= 1 / k, k is the process gain estimation value, T ₁ and T ₂ are the process parameter estimation values, and in the case of first-order delay, T ₂ = 0, T ₁ : time constant, and in the case of second-order delay, T ₁ = 2
・ Τ ・ T, T ₂ = T ² , S are Laplace operators.

Further, in the invention corresponding to claim 6, the process dynamic characteristic (reverse dynamic characteristic) is approximated in advance by a recurrence formula (excluding the recurrence formulas of claims 1 to 5) according to a calculation cycle of Δt. If it can be expressed in a similar manner, a neural network is constructed by using the recurrence formula, and the weight of each neuron element when the learning of the dynamic characteristic (reverse dynamic characteristic) of the process is started by using the coefficient of the recursive equation. This is a configuration in which initial values of coefficients are used.

[0028]

Therefore, by taking the above means, the present invention determines the configuration of the NN based on the recurrence formula by the calculation of the sampling period in the process control system, while learning the coefficient of the recurrence formula. By setting the initial value of the connection weight coefficient at the start, the number of neuron elements in each layer of the NN can be easily determined from the above recurrence formula, and unlike the case where the learning is started from the random initial value of the connection weight coefficient. Since learning is started in the vicinity of the coupling weight coefficient close to the optimum convergence value, the types of teacher patterns to be given for learning and the number of times of learning can be significantly reduced, and the probability of falling into a local minimum can be reduced.

Furthermore, since the connection between the neuron elements which is not expressed in the recurrence formula is also added (however, the value of the connection weight at the start of learning is set to 0), the real process and the pre-estimated characteristic It is also possible to correct the error between the two by learning the NN, and it is possible to enhance the identification ability and to identify the dynamic (reverse) characteristic of the process satisfactorily. That is, there is redundancy. Further, by incorporating such an NN in the control device, good and stable control can be realized.

[0030]

First, an embodiment of the invention according to claims 1 and 2 will be described with reference to FIG. In the figure, 11 is a process to be controlled, and an NN 12 is provided in parallel with this process 11. This NN12 is
When the process 11 is estimated or approximated by the first-order lag + waste time, it is configured by the recurrence formula of the first-order lag by the calculation period described later, and the coefficient of the recurrence formula is given as the coupling weight coefficient to the neuron element of each layer. It is configured to start learning. Reference numeral 13 denotes an error calculating means, which is a control signal of the process 11 which is a teacher signal and NN12.
It has a function of obtaining a squared error from the output of and the variable connection weight coefficient between the neuron elements so as to minimize the squared error. Reference numeral 14 is a delay means.

However, the NN 12 as shown in FIG. 1 is constructed on the basis of the following. now,
It is assumed that the process 11 is estimated or approximated in advance by the first-order delay + waste time defined by the following equation (2). y = {k / (1 + Ts)} e ^-Ls ・ X …… (2)

In this equation, x is a process input, y is a process output, k is a process gain, T is a process time constant, L is a process waste time (L may be zero), and s is a Laplace operator. When there is no dead time in the equation (2), it is represented by y = {k / (1 + Ts)} x (3). This first-order lag process 11 is applied to the invention according to claim 1. Further, when the differential equation is solved based on the equation (3), T (dy / dt) + y = k · x (4)

Therefore, the equation (4) will be obtained by a recurrence equation based on the sampling calculation period Δt. Now, the calculation cycle Δt
When the output change amount is Δy, the above equation (4) can be replaced with the following equation (5). T (Δy / Δt) + y _n = k · x _n (5) Here, the suffix n is a variable at the time point n · Δt. Further, when the equation (5) is modified to obtain Δy, Δy = (Δt / T) · (k · x _n −y _n ) ... (6) Here, by substituting the relational expression of y _{n + 1} = y _n + Δy (7) into the equation (6), the following recurrence equation is obtained. y _{n + 1} = y _n + (Δt / T) · (k · x _n −y _n ) ... (8) Furthermore, the calculation is repeated for each calculation period Δt using the formula (8) to obtain a solution. To seek. Therefore, the recurrence formula of the formula (8) can be decomposed into the following formula.

[0034]

[Equation 6]

Therefore, if the above first-order lag recurrence formula is obtained, the configuration of the NN 12 is embodied based on this recurrence formula. In the configuration example of FIG. 1, the waste time L of the process 11 is
Is estimated to be approximately 2Δt, the two delay elements inserted between the neuron elements a _{1 to} a _{3 in} the input layer are for this dead time 2Δt. Further, as is clear from the equation (9), the NN output y at the current time is
_{Since the n + 1} calculation expression includes the history value of the output y in addition to the history value of the x, the output of the output layer is delayed by the delay means 14 and 14 of Δt via the feedback line 15. After that, the data is input to the input side of the NN 12.

When learning the process identification by the back propagation method, the initial value of the coupling weight coefficient between the neuron elements is the following value from the equation in the last stage of equation (9), that is, the first-order delay recurrence. The coefficients of the equation shall be given. w ₁₁ ²¹ = 0, w ₁₁ ³² = 0 w ₂₂ ²¹ = 0, w ₁₂ ³² = 0 w ₃₃ ²¹ = (Δt / T) k, w ₁₃ ³² = 1 w ₄₄ ²¹ = (1-Δt / T ) · (Δt / T) · k, w ₁₄ ³² = 1 w ₅₅ ²¹ = (1−Δt / T) ² ・ (Δt / T) ・ k, w ₁₅ ³² = 1 w ₆₆ ²¹ = (1-Δt / T) ³ , W ₁₆ ³² = 1

Note that the mutual connection between each neuron element in the input layer and each neuron element in the intermediate layer is not limited to that shown in FIG. 1, but the connection weight coefficient of these initial values is used. The bond line is omitted because the value is given as zero.

Further, a concrete example of the invention according to claim 1 and its response characteristic will be described with reference to FIGS. First, FIG. 2 is a configuration diagram showing a specific example of the process dynamic characteristic learning device. In this learning device, the process 11 to be applied has no dead time, and {1 / (1 + 20 · s)} (10)

A system with only a first-order delay represented by On the other hand, N
The configuration of N12 has a smaller number of neuron elements than that of FIG. 1, and for example, the input layer has three neuron elements and the intermediate layer has three.
Each neuron element, and the output layer consists of one neuron element. This is based on the above equation (9),

[0040]

[Equation 7]

This is the case when the decomposition formula of is adopted. And
When learning the process 11 represented by the equation (10), the following procedure is performed. Now, assuming that the estimated transfer function of the process 11 represented by the equation (10) is {1 / (1 + 50 · s)} (12) in advance, the initial weighting coefficient of FIG. The values are as follows. In addition, 1t is set as Δt
c was used. w ₁₁ ²¹ = (Δt / T) k = 1/50, w ₁₁ ³² = 1 w ₂₂ ²¹ = {1- (Δt / T)} · (Δt / T) · k = {1- (1 / 50)} · (1/50) , w 12 32 = 1 w 33 21 = {1- (△ t / T)} 2 = {1- (1/50)} ² , W ₁₃ ³² = 1 The initial values of the other coupling weight coefficients are all zero.

Next, let's look at the response characteristics of the learning device realized as described above by actually inserting a signal into the NN 12. 3 and 4 are response characteristics at the learning stage. First, in FIG. 3, it is assumed that signals of 0.5 level steps are input to the process 11 and the NN 12. In addition,
FIG. 7A shows the input / output response characteristics of process 11,
FIG. 7B shows input / output response characteristics of the NN 12 before learning. This is the calculation cycle Δt = 1s for the response characteristic of equation (12).
This corresponds to an approximate calculation with the accuracy of ec. The sigmoid function is

[0043]

[Equation 8] Was used. The learning method uses a back propagation learning rule, and learning is performed according to the following two stages.
(I) The manipulated variable MV shown in FIG.
Input to N12 and learn 50 times repeatedly. (II)
Next, the sine wave input shown in FIG. 4 is used as an initial value with the coupling weight coefficient obtained by learning of (I), that is, 0.5 sin {(2π / 40) · t} (14)

To process 11 and NN12,
As in (I), the learning is performed 50 times using back propagation. 4 (a) is the response characteristic of the process 11 as in FIG. 3 (a), and FIG. 4 (b) is the case where the coefficient of equation (11) is given as the coupling weight coefficient of NN12, that is, equation (12). Shows the response characteristics of.

As a result, when the weight of the learning result of (II) is given as the coupling weight coefficient of the NN 12,
The response characteristics as shown in FIG. 7 were obtained. In addition, (a) of these figures is the input / output response characteristic of the process 11, and (b) is N.
It is the input / output response characteristic of N12. That is, FIG.
5 level steps, FIG. 6 shows that when a sine wave is input, process 11 and NN 12 ′ both have equal and rapidly converging response characteristics, and FIG. 7 inputs a 1.0 level step signal. As an example, even in this case, the correct response is obtained as in FIGS. 5 and 6, although the 1.0 level step is not learned at all.

On the other hand, what about the case where learning is performed by giving a random connection weighting coefficient to the NN as in the conventional case? First, the required number of elements in the NN input layer will be examined. Since the step input changes only at the moment of Δt and is constant thereafter, Δt is 1 sec and the process time constant (see the above equation (10)) is 20 sec. 20 or more are required. Because usually N
This is because N does not have a feedback input, and therefore, if the same input is input to each element of the input layer of NN, the output of NN also becomes constant.

Furthermore, if the number of elements in the input layer is 20 or more, the number of elements in the intermediate layer must be increased accordingly. When this happens, many more complications arise. For example, the calculation time and the number of learning increases due to the explosive increase of the coupling weight coefficient.

In this respect, since the number of elements can be easily found from the first-order lag recurrence formula in the present apparatus, the number of elements can be minimized to construct the NN, and moreover, the initial value of the coupling weighting coefficient between the elements is also gradually increased. Since it is determined and given from the chemical formula, the process 11 can be identified with a very small number of learnings. next,
An embodiment of the invention according to claims 3 and 4 will be described. In the device of this embodiment, the process is performed by y = {k / (T ² ・ S ² + 2τ ・ T ・ s + 1)} e ^-Ls ・ X… (15)

When it is estimated or approximated by the second-order lag + waste time, this process is identified by using NN, and specifically, the configuration of NN12 as shown in FIG. 8 is adopted. In the equation (15), k is a process gain, T is a process parameter (time constant),
τ is the process damping ratio. However, when the dead time L = 0 in the equation (15), y = {k / (T ² ・ S ² + 2τ · T · s + 1)} · x (16)

Therefore, the case of the dead time L = 0 will be described below. However, when L ≠ 0, the neuron element for the dead time (for example, FIG. 1) is used in accordance with the magnitude of the estimated value of L as in FIG. A ₁ , a ₂ ) of the above may be appropriately added. First,
Solving equation (16) with a differential equation gives T ² (D ² y / dt ² ) + 2τ · T (dy / dt) + y = kx (17), where z = dy / dt (18) gives T ² (Dz / dt) + 2τ · T · z + y = kx (19) can be substituted. Further, when the equation (19) is modified by using minute amounts Δz, Δy, Δt, etc., Δz = (kx _n −2τ · T · z _n −y _n ) Δt / T ² (20) Δy = z _n · Δt (21) is obtained, and further, z _{n + 1} = z _n + Δz (22) y _{n + 1} = y _n + Δy (23)

There is a relationship of So, these (20)
If the NN is constructed using the second-order lag recurrence formula consisting of the formulas (23), and y is repeatedly calculated in the same manner as the first-order lag described above, y is obtained.

Therefore, the learning device shown in FIG.
It is configured based on the equations (0) to (23). When learning for process identification is performed using this learning device, the initial value of the connection weighting coefficient between the neuron elements is given by the coefficient of the second-order lag recurrence formula as described below. w ₁₁ ²¹ = k · Δt / T ² , W ₂₁ ³² = 1 w ₂₂ ²¹ = 1, w ₁₂ ³² = 1 w ₃₂ ²¹ = − (Δt / T ² ), W ₂₃ ³² = 1 w ₄₃ ²¹ =-(2τ · T · Δt / T ² ), W ₂₄ ³² = 1 w ₅₃ ²¹ = 1, w ₂₅ ³² = 1 w ₁₅ ³² = Δt

The initial values of the other coupling weight coefficients are all zero. The connections between the neuron elements of the input layer and the intermediate layer, and between the intermediate layer and the output layer are all connected in addition to those shown in the figure, but those that give initial weights as zero omit the connecting line. There is. If there is a dead time, the number of elements in the input layer may be appropriately increased, and the input x _n may be delayed as in FIG. Further, the control amount which is the output of the process 11 is used for the learning teacher signal of the output layer c ₁ output,
Differentiating means 16 is provided for the learning teacher signal of the output layer c _{2 and} the differential ΔPV / Δt of the control amount is used.

Next, an embodiment of the invention according to claim 5 will be described. In this embodiment, when the process is estimated or approximated by a second-order lag or a first-order lag in advance, its inverse dynamic characteristic is identified using NN. Now, the transfer function of the reverse dynamic characteristic is expressed as y = k ′ (1 + T ₁ s + T ₂ s ² ) .X ... (24) However, k '= 1 / k, and k is a process gain. In the following description, if the process is a first-order delay, T ₂ = 0 may be set. Then, when the differential equation is solved for the equation (24), y = k'x + k'T ₁ (dx / dt) + k'T ₂ (d ² x / dt ² ) …… (25) Therefore, if this equation (25) is rewritten into an equation using the calculation cycle of Δt and the suffix n,

[0055]

[Equation 9] The following recurrence formula is obtained.

FIG. 9 is a block diagram of a learning device using such a recurrence formula. In this apparatus, from the viewpoint of learning the identification of the reverse dynamic characteristic, the control amount of the process 11 is input to the NN 12, and the operation amount is given as a teacher signal of the NN 12. Further, the dotted line frame portion is a neuron element added for correction when the process 11 is deviated from the second-order lag or the first-order lag, and if the correction is not required, the dotted line frame portion may be omitted. Good.

When learning for process identification is performed using this learning device, the coefficient of the recurrence formula shown in the formula (26) is given to the initial value of the coupling weight coefficient of each neuron element. w ₁₁ ²¹ = k '+ (k'T ₁ / Δt) + (k'T ₂ / Δt ² ), W ₁₁ ³² = 1 w ₂₂ ²¹ = − {(k′T ₁ / Δt) + (2k′T ₂ / Δt ² )}, W ₁₂ ³² = 1 w ₃₃ ²¹ = (k′T ₂ / Δt ² ), W ₁₃ ³² = 1

The initial values of the other coupling weight coefficients are all zero. In addition, all the connections between the neuron elements in the input layer and the intermediate layer are connected in addition to those shown in FIG. 9. However, since the weight of the initial value is given as 0, their connecting lines are omitted.

Therefore, also in the configuration of this embodiment, since the number of elements can be easily found from the recurrence formula, the number of elements can be minimized to construct the NN, and the initial value of the coupling weight coefficient between the elements is also set. Since it is determined and given from the recurrence formula, the process 11 can be identified with a very small number of learnings.

The present invention is not limited to the above embodiment. That is, in FIG. 1, FIG. 2, FIG. 8, FIG. 9, etc., when the process is approximated by a first-order lag or a second-order lag, or a waste time is included therein, its dynamic characteristic or reverse dynamic characteristic is NN. However, if the dynamic characteristics of the process are represented by a past input / output data string and a recurrence formula based on their differentiation, it is obvious that the process can be easily applied.

In the intermediate layer shown in FIG. 1, the number of neuron elements is reduced to, for example, b ₁ and b ₂ , b ₃ and b ₄ , b _5.
It is also possible to combine b ₆ and b ₆ into three. Further, although only y _n-2 is adopted as the feedback input based on the last formula of the formula (9) in FIG. 1, as the correction input, for example, as shown in FIG. 10, y _n , y _n-1 , That is, the correction neuron elements a ₇ and a ₈ may be added. The coupling weight coefficient of these added correction neuron elements a ₇ and a ₈ is set to 0.

Further, the apparatus of FIG. 1 shows the case where the estimated approximate value L of the dead time is 2Δt, but the larger the dead time, the more the waste time input layer neuron elements. For example, in the case of 20Δt, 20 are required for the input layer neuron elements for waste time only. Therefore,
In such a case, only the delay time (a ₁ and a ₂ , a ₂ and a ₃ in FIG. 1) to the waste time input layer neuron element is made larger than the other delay times, so that the waste time neuron The number of elements can be reduced.

On the other hand, the effect of correcting the deviation from the estimation can be enhanced by increasing the number of neuron elements. For example, in the apparatus of FIG. 8, the number of neuron elements is reduced based on the recurrence formulas of formulas (20) to (23). △
By decomposing up to t, (n−2) Δt, ...} And increasing the number of neuron elements accordingly, it is possible to satisfactorily correct the deviation from the pre-estimated characteristic and perform learning. In addition, the present invention can be modified in various ways without departing from the scope of the invention.

[0064]

As described above, according to the present invention, while grasping the state of the process in advance, the composition of the NN is determined to some extent from the recurrence formula of process identification, and the coefficient of the recurrence formula is determined. Since the initial value of the coupling weight coefficient of the NN configuration is given based on the above, the learning ability and speed for identifying the process can be greatly improved, and thus the control performance can be improved, and moreover, various control systems can be provided. N
The applicable range of N can be expanded.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an apparatus when a process is approximated by a first-order delay + a waste time, as an embodiment of the invention according to claims 1 and 2.

FIG. 2 is a specific configuration diagram of an apparatus when a process is approximated by a first-order delay.

FIG. 3 is a response characteristic diagram when the device of the present invention is applied to an actual process for learning.

FIG. 4 is a response characteristic diagram when the device of the present invention is similarly applied to an actual process for learning.

FIG. 5 is a response characteristic diagram when the device of the present invention is similarly applied to an actual process for learning.

FIG. 6 is a response characteristic diagram when the device of the present invention is also applied to an actual process for learning.

FIG. 7 is a response characteristic diagram when the device of the present invention is also applied to an actual process for learning.

FIG. 8 is a block diagram of an apparatus when a process is approximated by a second-order delay as an embodiment of the invention according to claims 3 and 4.

FIG. 9 is a block diagram of a learning device for identifying a reverse dynamic characteristic when a process is approximated by a second-order lag or a first-order lag, as an embodiment of the invention according to claim 5;

FIG. 10 is a configuration diagram illustrating another embodiment of the device of the present invention.

FIG. 11 is a configuration diagram showing a conventional general hierarchical NN.

FIG. 12 is a diagram illustrating a back propagation learning rule.

FIG. 13 is a configuration diagram of a conventional NN that performs time series processing.

FIG. 14 is a view for explaining a concrete application example of NN control in a conventional device and the device of the present invention.

FIG. 15 is also a NN in the conventional device and the device of the present invention.
FIG. 6 is a diagram illustrating a specific application example of the control of FIG.

FIG. 16 is also a NN in the conventional device and the device of the present invention.
FIG. 6 is a diagram illustrating a specific application example of the control of FIG.

[Explanation of symbols]

11 ... Process (control target), 12 ... NN (neural network), 13, 13a ... Error calculation means, 14 ...
Delay means, 16 ... Differentiation means.

Claims (6)

[Claims] 1. A process in which y = {k / (1 + T ·
s)} · x is estimated or approximated by a first-order lag, in a process dynamic-characteristic learning device that learns the dynamic characteristics of the process using a neural network, the following first-order lag recurrence formula with a calculation cycle of Δt Number 1] The neural network is configured based on the above, and the coefficient of the first-order lag recurrence formula is given as an initial value of the coupling weight coefficient between the neuron elements when learning the dynamic characteristics of the process is started. Process dynamics learning device. In the above equation, x is a process input, y is a process output, k is a process gain estimated value, T
Is the process time constant estimate and s is the Laplace operator.
The suffix n represents a variable at the time point n · Δt. 2. A process in which y = {k / (1 + T ·
s)} e^-Ls ・ Estimated by first order delay of x + waste time or
Or approximate the dynamics of the process.
Process dynamics learning device that learns using a virtual network
, The first-order lag recurrence formula according to the calculation period of Δt
When the neural network is constructed based on
At the same time, when starting learning the dynamics of the process,
As the initial value of the coupling weight coefficient between neuron elements,
The coefficient of the second-order lag recurrence formula is given, and the neural network
A special feature is that a dead time neuron element is provided in the network.
A device for learning process dynamic characteristics. However, in the above formula
x is process input, y is process output, k is process
In estimate, T is process time constant estimate, s is Laplace
Operator, L is process waste time. 3. The process is performed by y = {k / (T ² ・ S ²
+ 2τ · T · s + 1)} · x, the process dynamic characteristic learning device that learns the dynamic characteristic of the process using a neural network can calculate Next-lag recurrence formula z _{n + 1} = z _n + Δz y _{n + 1} = y _n + Δy Δz = (k · x _n −2 · τ · T · z _n −y _n ) · (Δt
/ T ² ) The neural network is constructed based on Δy = z _n · Δt, and the second-order lag recurrence is used as an initial value of a connection weighting coefficient between the neuron elements when learning the dynamic characteristics of the process is started. A process dynamic characteristic learning device characterized by giving a coefficient of an equation. In the above equation, x is a process input, y is a process output, k is a process gain estimated value, T is a process parameter estimated value, τ is a process damping ratio estimated value, and s is a Laplace operator. The suffix n is the time point n
・ Indicates the variable in Δt. 4. The process is performed with y = {k / (T ² ・ S ²
+ 2τ ・ T ・ s + 1)} e ^-Ls In a process dynamic characteristic learning device that learns the dynamic characteristic of the process using a neural network when estimated or approximated by the secondary delay of x + waste time, a quadratic delay recurrence formula based on the calculation cycle of Δt Based on the above, the neural network is configured based on, and the coefficient of the second-order lag recurrence formula is given as an initial value of the coupling weight coefficient between the neuron elements when learning of the dynamic characteristics of the process is started, and A process dynamic characteristic learning device characterized in that a dead time neuron element is provided in a neural network. Where x is the process input,
y is the process output, k is the process gain estimate, T is the process parameter estimate, τ is the process damping ratio estimate,
s is a Laplace operator and L is a process waste time. 5. The process has a secondary delay or a primary delay in advance.
The inverse of the process, if estimated or approximated by
Dynamic characteristics y = k '(1 + T₁・ S + T₂・ S² ) X Dynamics of learning neural networks using neural networks
In the characteristic learning device, Recurrence formula based on Δt calculation cycle [Equation 2] When the neural network is constructed based on
If we start learning the reverse characteristic of the process,
As the initial value of the connection weight coefficient between the neuron elements,
Process dynamics characterized by giving recurrence coefficients
Learning device. In the above equation, x is the process output and y is the process
Input, k '= 1 / k, k is the process gain estimate, T
₁, T₂Is the process parameter estimate,
In this case T₂= 0, T₁: Time constant, secondary delay
In case of T₁= 2 · τ · T, T₂= T² , S is Laplace
It is an arithmetic. Also, the suffix n is at the time point n / Δt.
Represents a variable. 6. The process dynamic characteristic (reverse dynamic characteristic) is previously Δ.
When it can be approximately represented by a recurrence formula (excluding the recurrence formulas of claims 1 to 5) according to the calculation cycle of t, the process dynamics that learns the dynamics (reverse dynamics) of the process using a neural network. In the learning device, the neural network is configured based on the recurrence formula (excluding the recurrence formulas of claims 1 to 5), and learning of the dynamic characteristic (reverse dynamic characteristic) of the process is started. A process dynamic characteristic learning device, wherein a coefficient of the recurrence formula (excluding the recurrence formula of claims 1 to 5) is given as an initial value of a coupling weight coefficient between neuron elements.