JPH03260807A - Fuzzy controller of hierarchical network constitution - Google Patents

Abstract

PURPOSE:To adaptively cope with the adjustment, the change, etc., of a control algorithm by making use of a data converting function of a hierarchical network structure which is adaptive to a fuzzy control rule. CONSTITUTION:A hierarchical network part 11 contains a calculation layer consisting of a processing unit 13 which calculates the truth value of a membership function with the control state value used as an input and a rule layer consisting of a rule unit 14 which calculates the grade value with the multiplication value of the output value of the unit 13 and the internal state value as an input. Then a rule layer consists of a rule unit 15 which calculates the grade value with the multiplication value of the output value of the unit 14 and the interval state value used as an input. Furthermore an output layer consists of an output unit 16 which calculates a control manipulated variable from the membership function. That is, a desired control manipulated variable is outputted from the output layer when the control state value is given to the calculation layer as the internal state value. Thus, the controller adaptively cope with the adjustment, the change, etc., to the fuzzy control rule.

Description

[Detailed Description of the Invention] [Summary] Regarding a fuzzy controller with a hierarchical network configuration for realizing a fuzzy controller, the present invention also aims to enable the construction of an adaptive fuzzy controller. A calculation layer in which the hierarchical network part is configured of processing units that calculate the truth value of the antecedent membership function using control state quantities as input, and corresponds to the output value of the corresponding processing unit of the calculation layer. A first rule layer consists of a rule unit that calculates a grade value according to arithmetic processing by inputting a multiplication value with an internal state value of A second rule layer includes a rule unit that calculates a grade value according to arithmetic processing using the multiplication value with the state value as input, and the output value and consequent membership of the corresponding rule unit of the second rule layer. and an output layer composed of an output unit that calculates the control operation amount etc. from the function, and when the lIM control state amount is given to the calculation layer as the internal state value, the desired control operation amount etc. is output from the output layer. Configure to set the value that will be used.

[Industrial application field]

The present invention relates to a hierarchical network configured fuzzy controller that functions as an adaptive fuzzy controller.

Fuzzy control is becoming popular as a new control processing method. This fuzzy control expresses a control algorithm that includes ambiguity such as human judgment in a 1f-then format, and executes this control algorithm according to fuzzy reasoning to calculate the control operation amount from the detected control state amount. It calculates and controls the controlled object. Since the fuzzy controller that performs this fuzzy control handles uncertain things, it is necessary to have a configuration that can adaptively respond to adjustments and changes in the control algorithm.

[Conventional technology]

The fuzzy control rule is if Xt is big and Xt is ss
+all then y+ is big (
The IF part is called the antecedent part and is a part that describes the conditions for control state quantities such as temperature data, and the THEN part is called the consequent part and describes the conditions for control manipulated variables such as operating terminals. The fuzzy controller describes the control logic according to the control object vIi1g
In addition to managing such multiple fuzzy control rules prepared as logic, the meaning of ambiguous linguistic expressions such as "large" and "small" written in each fuzzy control rule is quantified as a membership function. A configuration will be adopted in which the information will be managed by

When the fuzzy controller is given control state quantities such as temperature data and water level data from the controlled object, it first calculates the membership function (antecedent membership function) of the control state quantity that it is managing. The membership function [(truth value) of the control state quantity given from 0
Next, in accordance with antecedent computations such as selecting the minimum value, a process is executed to determine the applicable value for the consequent in each fuzzy control rule. That is, to explain using the example of the fuzzy control rule mentioned above, rx, isbig J
The truth value of is “0.8”, rx, is 5sall”
When the truth value of is "o, s", this "0.5" is determined as the applied value to the consequent part of the fuzzy control rule according to the antecedent part operation that selects the minimum value. This is how it is processed.

Subsequently, the fuzzy controller calculates each fuzzy control rule given for the same membership function (which becomes the consequent membership function) of the same control manipulated variable according to the consequent operation, such as selecting the maximum value. A process is executed to determine the applied value for the membership function from the applied value for the consequent. In other words, in the fuzzy control rule described above, for the consequent 'V+ is big'
o, s” as the applied value, and on the other hand, when using another fuzzy control rule to set “0.6” as the applied value for 'L is biH, according to the consequent operation that selects the maximum value, Processing is performed to determine this "0.6" as the applied value of the control operation amount y1 to rbtgJ.

Subsequently, the fuzzy controller reduces the membership function of the control manipulated variable according to the determined application value, and
Execute the process of calculating the control operation amount, which is a fuzzy inference value, according to the process such as finding the center of gravity of the figure of the function sum of membership functions for the same control operation amount, and output it to the operation terminal etc. I will do it.

Conventional fuzzy controllers have adopted a configuration in which execution of such fuzzy control rules is achieved by executing them programmatically. That is, according to the software means of the computer system that performs sequential processing, first, the truth value of the control state quantity is calculated, then the applied value for the consequent is determined according to the antecedent part operation, and then Therefore, the application value of the control operation amount to the membership function is determined according to the consequent part calculation, and the control operation amount is calculated and output by center of gravity calculation or the like.

[Problem to be solved by the invention]

However, fuzzy control rules are generated according to the knowledge of the target process of an operator who is familiar with controlling the target.

The reality is that it is not possible to generate fuzzy control rules that can achieve the desired control from the beginning, and the generated fuzzy control rules must be tuned by trial and error while being evaluated through simulators and on-site tests. By doing so, we have no choice but to complete the fuzzy control rules that are suitable for the control target.

However, in the conventional method of configuring a fuzzy controller using software means, the fuzzy controller simply executes the control logic according to the description procedure of the generated fuzzy control rule, and does not have any adaptability. No, from now on, conventional fuzzy controllers will not exhibit a function that reduces the burden of fuzzy control rule tuning on the operator.

The present invention has been made in view of the above circumstances, and is a new system that adopts a configuration for realizing fuzzy control rules and also adopts a configuration that can adaptively respond to adjustments and changes to fuzzy control rules. The purpose of the present invention is to provide a fuzzy controller with a hierarchical network configuration that forms a fuzzy controller.

[Means to solve the problem]

FIG. 1 is a diagram showing the principle configuration of the present invention.

In the figure, reference numeral 10 denotes a hierarchical network configuration fuzzy controller according to the present invention, which has a configuration that realizes fuzzy control rules and can also adaptively respond to adjustments and changes of fuzzy control rules. The fuzzy controller 11 is a hierarchical network unit included in the fuzzy controller 10 having a hierarchical network structure, which performs data conversion defined by a hierarchical network structure and internal state values assigned to internal connections of the hierarchical network structure. those that execute fuzzy control rules according to functions, 17
is an internal state value management unit, and the hierarchical network unit 11
20 is a learning signal presentation device that presents a group of control state quantity data for learning to the input layer of the hierarchical network section 11, and A group of control operation amount data for learning that is paired with a control state amount data group for learning (when the control state amount data for learning is given, the control operation amount to realize the desired control state) ) is presented to a learning processing device 30, which will be described next, and 30 is a learning processing device, which is a hierarchical network unit that is output in response to presentation of a control state quantity data group for learning by the learning signal presentation device 20. The error value between the control operation amount data group from 11 and the learning control operation amount data group presented from the learning signal presentation device 20 is calculated, and the internal state to be learned is determined based on the error value. By sequentially updating the values, the internal state value for which the error value falls within the allowable range is learned.

The hierarchical network unit 11 includes an input layer composed of a plurality of input units 12 provided in a format corresponding to input control state quantities, and is located as a subsequent layer of the input layer.
A calculation layer composed of a plurality of processing units 13 provided in a format corresponding to the antecedent membership function of a fuzzy control rule, and a calculation layer located as a subsequent layer of the calculation layer, for example, a number of processing units 13 corresponding to the number of fuzzy control rules. A first rule layer composed of a plurality of rule units 14, and a plurality of rule units positioned as a subsequent layer of the first rule layer, for example, the number of rule units is equal to the number of consequent membership functions of a fuzzy control rule. 15, and a second rule layer located as a subsequent layer of the second rule layer,
It is configured to have a hierarchical network structure called an output layer composed of one or more output units 16.

The input unit 12 of this input layer receives the control state quantity of the controlled object as input and processes the input control state quantity to be distributed to the corresponding processing unit 13 of the calculation layer. 13 calculates the truth value of the antecedent membership function of the fuzzy control rule using one or more control state quantities notified from the input unit 12 as input, and calculates the truth value of the membership function of the antecedent part of the fuzzy control rule, and calculates the truth value of the membership function of the antecedent part of the fuzzy control rule, and calculates the truth value of the membership function of the antecedent part of the fuzzy control rule.
The rule unit (14) of the first rule layer outputs an internal state assigned to the internal connection between one or more truth values notified from the processing unit 13 and those processing units 13. The multiplication value with the value is input, and the grade value is calculated by executing the assigned operation and output to the corresponding rule unit 15 of the second rule layer, and the rule unit 1 of the second rule layer
5 executes the assigned operation using as input the multiplication value of one or more grade values notified from the rule unit 14 and the internal state value assigned to the internal connection between those rule units 14; The output unit 16 of the output layer is configured with, for example, a two-stage configuration, and calculates the grade value and outputs it to the corresponding output unit 16 of the output layer. From the consequent membership function of the fuzzy control rule associated with the grade value input, the control operation amount or its corresponding value that becomes the fuzzy inference value is calculated and output.

One or more rule layers may be provided between the first rule layer and the second rule layer, each of which is composed of one or more rule units that execute assigned operations. The rule unit of the rule layer provided in this way multiplies one or more grade values notified from the rule unit of the previous rule layer by the internal state value assigned to the internal connection between those rule units. By taking the value as input and executing the assigned operation, the grade value is calculated and processed to be output to the corresponding rule unit in the subsequent rule layer.

[Effect]

The hierarchical network unit 11 of the present invention includes an input unit I2
When the control state quantity is given via The truth value of the function is calculated, and the calculated truth value is distributed and output to the rule unit 14 of the first rule layer that is internally coupled. When the truth value of the membership function for the control state quantity is received from the corresponding one or more processing units 7113, the first
The rule unit 14 of the rule layer calculates a grade value that is a calculation result value by executing calculation processing such as an assigned function conversion function, and a second rule that internally combines the calculated grade value. It is distributed and output to the rule unit 15 of the layer. Here, the first rule layer and the second
When further hierarchical rule layers are provided between the second rule layer, the rule units of those rule layers calculate the grade value and the rule unit 15 of the second rule layer calculates the grade value.
It will be processed so that it is propagated to.

When the grade value is received from the corresponding one or more rule units in the previous layer, the rule unit 15 in the second rule layer calculates the calculation result value by executing calculation processing such as the assigned function conversion function. At the same time, the calculated grade value is outputted to the output unit 16 of the output layer that is internally coupled. Upon receiving this grade value, the output unit 16 calculates the control operation amount that becomes the fuzzy inference value or the value associated with it by applying the function conversion function assigned to this grade value. It operates to output as follows.

In this way, the hierarchical network unit 11 of the present invention operates to execute the fuzzy control rules according to the data conversion function of the hierarchical network structure structured to conform to the fuzzy control rules.

In the present invention, a hierarchical network configuration fuzzy controller lO
When attempting to construct a desired fuzzy controller, the operator first asks the processing unit 13 to
A function for calculating the truth value of the membership function for the generated control state quantity is assigned, and the function for calculating the truth value of the membership function for the generated control manipulated variable is assigned to the output unit 16. (necessary for the calculation process), as well as registering the first
and the rule unit of the second rule layer) 14°15 (the first rule unit)
When a rule layer is provided between the first rule layer and the second rule layer, the function of rule calculation is assigned based on the knowledge of fuzzy control rules obtained for the rule units (including those rule units). Here, the number of units (14.15) is set based on the number of fuzzy control rules obtained. For example, a fuzzy The internal connection is constructed in a format that follows the control rules, and the initial values of the internal state values are set and stored in the internal state value management section 17. Furthermore, when a rule layer is provided between the first and second rule layers, internal connections are configured in an appropriate format between those rule units, and initial values of internal state values are set. Processing is performed so as to store it in the internal state value management unit 17.

Subsequently, the operator starts the learning signal presentation device 20 and the learning processing device 30. When activated in this manner, the learning signal presentation device 20 presents control state quantity data for learning to the input layer of the hierarchical network unit 11, and performs control operations for learning to the learning processing device 30. Present quantitative data. In response to the presentation of this learning signal, the learning processing device 30 determines that the control operation amount data from the hierarchical network unit 11 output in response to the presentation of the control state amount data for learning is used to control the presented learning control. The internal state value of the internal connection between the processing unit 13 and the rule unit 14, the internal state value of the internal connection between the rule unit 14 and the rule unit 15, and , when a rule layer is provided between the first and second rule layers, in addition to this, learn internal state values of internal connections between those rule units,
A process of setting the learned value of the determined internal state value in the internal state value management section 17 is executed.

In addition, in this internal state value learning process, the internal state m4il of the internal connection set between the input crab unit 12 and the processing unit 13, and the internal state m4il set between the rule unit 15 and the output unit 16 are The internal state values of the connections, and furthermore, the internal state values of the internal connections within the processing unit 13 itself or the output unit 16 itself, can also be used as learning targets. If these are the learning targets, for example, when the processing unit 13 is configured to calculate the truth value of the membership function of the control state quantity based on the internal state value of the internal connection with the input unit 12, this learning process With the new internal state value found, it becomes possible to tune the membership function of the control state quantity assumed in advance to one that is more suitable for the controlled object. In addition, the output unit 16
When adopting ** in which the truth value of the membership function of the control operation amount is managed by the internal state value of the internal connection with the rule unit 15, the new internal state value found by this learning process is used to control the control assumed in advance. It becomes possible to tune the membership function of the manipulated variable to be more suitable for the controlled object.

Once the learning value of this internal state value is set, the hierarchical network configured fuzzy controller IO will function as a fuzzy controller that executes the desired control logic.

As described above, the present invention is characterized in that a fuzzy controller is constructed according to the hierarchical network unit 11 having a hierarchical network structure structured to conform to fuzzy control rules.

Well known as a neurocomputer,
A data processing device that follows the data conversion function of a hierarchical network structure will be able to handle it even if an unexpected input signal may be input once the internal state value assigned to the internal connection of the hierarchical network structure is set. The present invention realizes "soft" data processing that outputs a unique output signal.8 The present invention adapts the hierarchical network structure of such a hierarchical network-configured data processing device to a fuzzy control rule and structures it to create a hierarchical network unit. 11. Separately from membership functions for control state quantities and control operation quantities generated in advance, the internal state values of internal connections in the rule portion of fuzzy control rules are configured to be learning targets.

By using such a hierarchical network configuration data processing device, it is possible to mechanically determine the relationships among fuzzy control rules using the learning processing device 30, which reduces the burden of fuzzy control rule tuning on the operator. can be significantly reduced. and,
Since the hierarchical network configuration data processing device is not used as is, but the hierarchical network unit 11 is structured in a format that conforms to the fuzzy control rules, the fuzzy control rules mapped to the hierarchical network unit 11 are, so to speak, It becomes visible from the outside, and processing such as tuning the membership function separately can be performed, and consistency with conventional fuzzy controllers can be maintained.

〔Example〕

Hereinafter, the present invention will be explained in detail according to examples.

FIG. 2 illustrates one embodiment for constructing the hierarchical networked fuzzy controller 10 of the present invention. In the figure, the same features as those explained in FIG. 1 are indicated by the same symbols. 1"-h is the input crab knit corresponding to the input crab knit 12 in Fig. 1, and 1-i is the processed uniso in Fig. 1)
13 and Rule Uni.

) 14. The basic unit that will constitute 15, 1-
j is a basic unit that constitutes the output unit 16 in FIG. 1, and 17a is a gradation value management unit corresponding to the internal state value management unit 17 in FIG.
It manages the weight values (corresponding to the internal state values in FIG. 1) assigned to the internal connections of 1.

21 is a learning signal storage unit, and the hierarchical network unit 1
22 is a learning signal presentation unit that stores a learning signal (the basic unit is a pair of control state quantity data and control operation amount data) used for learning the weight value of 1; The learning signal is read out from the learning signal unit 21, and the control state quantity data therein is presented to the hierarchical network unit 11, and the other pair of control state quantity data is sent to the weight value changing unit 30a, which will be described later, and the learning signal, which will be described next. The convergence determination unit 23 is a learning convergence determination unit that receives the control operation amount data output from the hierarchical network unit 11 and the control operation amount data for learning from the learning signal presentation unit 22. Then, it is determined whether the error in the data processing function of the hierarchical network section 11 is within an allowable range, and the learning signal presentation section 22 is notified of the determination result.

Reference numeral 30a denotes a weight value changing unit corresponding to the learning processing device 30 in FIG. By calculating the update amount of the weight value according to the back propagation method and updating the weight value, learning is performed so that the weight value converges to a convergence value.

As explained in FIG. 1, the hierarchical network unit 11 of the present invention has a hierarchical network structure structured to have the same structure as the fuzzy control rule. Hierarchical network unit 1 characteristic of the present invention
Before going into the explanation of 1, we will explain the hierarchical network structure (hereinafter referred to as the known hierarchical network structure) provided in the hierarchical network configuration data processing device known for boat fishing, and also explain the weight value changing unit. 30a
The back-propagation method, which is a learning algorithm implemented by Microsoft and which determines the weight value of internal connections in a known hierarchical network structure, will be described in detail.

A known hierarchical network structure basically consists of a type of node called a basic unit and internal connections having weight values. FIG. 3 shows the basic configuration of the basic unit 1.

, this basic unit 1 is a multi-car output system, and includes a multiplication processing section 2 that multiplies a plurality of inputs by weight values of respective internal connections, and an accumulation processing section 3 that adds up all the multiplication results. and a dark value processing unit 4 that performs nonlinear threshold processing on this accumulated value and outputs one final output.

Assuming that the h layer is the first layer and the i layer is the second layer, the summer processing section 3 executes the calculation of the following equation (1), and the threshold processing section 4
Now, execute the calculation of equation (2) below.

x28−Σ)'phWih (1
) formula y pi = 1 / (1+ exp(x pt
+θi)) Equation (2) where, h: Unit number of the h layer i: Unit number of the 1st layer p: Input signal pattern number 0.21 Darkness value of the 1st unit of the layer W4. : Weight value of internal connection between h-i layer : Output from the 1st unit of the i layer in response to the input signal of the p-th pattern.And in a known hierarchical network structure, a large number of basic units 1 with such a configuration distribute and output the input signal value as it is. With 1 piece of crab knit as input layer,
As shown in Figure 4, they are connected hierarchically to form a hierarchical network, and exhibit a parallel data processing function of converting an input signal into a corresponding output pattern. It turns out.

In the puncture propagation method, the weight value Wtb and the threshold value θ of the hierarchical network are adaptively and automatically adjusted and learned using error feedback. (1)
As is clear from equation (2), the weight value Wlk and the threshold value θ,
It is necessary to perform the adjustment at the same time, but this is a difficult task that interferes with each other. Therefore, the present applicant previously applied for the patent application No. 1984-333484.
As disclosed in "Network Configuration Data Processing Apparatus" (filed on December 28, 2013), a unit that always outputs 1" to the h layer on the input side and assigns a threshold value θ as a weight value to the output. By providing a threshold value θ, a weight value W,.

We proposed that the threshold value θ be incorporated into the weight value and treated as a weight value. By doing this, the above (1)
) Equation (2) is: )[,,=ΣVphWth (3)
Formula y□= 1 / (1+ exp(x j+))
It is expressed by equation (4).

Next, a weight value learning processing method using the bank propagation method will be explained according to the description format of equations (3) and (4). Here, this explanation will be given assuming that the hierarchical network unit 11 has a three-layered hierarchical network consisting of layers h, layer i, and layer j as shown in FIG.

(3) By analogy with equations (4), the following (5) (6
) formula is obtained. That is, Xpj-Σy□WJl(5)2)'tJ-1/(1+exp(-x,))
) Equation j: Unit number W of the j layer, , : Weight value of internal connection between the i-j layers The output weight value changing unit 30a from the fourth unit of the j layer changes the output pattern y from the output layer when the learning input pattern is presented
and the teacher pattern d□ (JIJ for the input signal of the p-th pattern), which is the output signal of the output pattern y1
When the teacher signal to the th unit is received, first, the difference value between the output pattern y□ and the teacher pattern d1 [
d□y□], then calculate α□=y□(1-y□)(d□-y□) (7), and then, ΔW, (t)-εΣαpj )' pi+ζΔWji(
t-1) The update amount ΔWj of the weight value between the i layer and the j layer according to (8) formula % formula %: :.

Calculate (1). Here, the reason for adding "ζΔWB(t-1)J, which is related to the update amount of the weight value determined in the previous update cycle, is to speed up the learning.

Next, the weight value changing unit 30a first uses the calculated α□ to calculate β2. =y□(l y, +t)Σα. W, 1(t-1
) (9) and then ΔWr h (t) = tΣβpi V ph+ζΔ
Wih(t-1) According to equation (10), the update amount ΔW3h(1
) is calculated.

Subsequently, the weight value changing unit 30a changes the weight value W7t(t) for the next update cycle according to the calculated update amount.
=WJt(t1)+ΔW,,(t)Wtb(t)
By repeating the method of determining = Wib (t-1) + ΔWih (t), the output cover turn y from the output layer that is output when the input cover turn for learning is presented
□ and the teacher pattern d, which is the signal to be taken by the output turn y2, match the weight 4rlt
Processing is performed to learn w, t 1w t b.

Then, the hierarchical network unit 11 creates layer g=h layer i layer−j
When you have a hierarchical network with a four-layer structure called layers,
The weight value changing unit 30a first sets r ph=3' a
h(1'J pk)ΣβptWth(t 1)(11
) formula, then ΔWk, (t)-aΣr, h3' ps+ζΔWh
The update amount ΔW of the weight value between the g-layer and the h-layer (t) is calculated according to the formula o(t-t) (12), and the update amount ΔW of the weight value between the eyebrows on the previous stage is Processing is performed to determine the value using the value obtained from the subsequent stage and the network output data.

Here, if the basic unit 1 does not include the threshold processing section 4, the above equation (7) becomes αpj=(dpj 31□)(7), and the above equation (9) becomes β, 8 -Σαp=WJt(t-1) (9)
The equation (11) above becomes γ, = ΣβpiW+b(t 1) (1
1) Formula Next, the hierarchical network unit 11 characteristic of the present invention
I will explain about it.

FIG. 5 shows an embodiment of the hierarchical network unit 11. In the diagram, 1 is the human unit shown in FIG. 4, 1 is the basic unit shown in FIG. 4, and 1a is 2 located in the previous stage. The weight values “1” and “-1” are applied to the output values of the basic unit 1.
The subtracters 1a and 1b, which are configured by basic units 1 without the dark value processing section 4, have as input the product multiplied by the value multiplied by a prescribed weight value. A truth value adder constituted by a basic unit 1 which takes an object as an input and does not include a dark value processing section 4,
Reference numeral c denotes a center-of-gravity derived value calculator which receives as input the output value of the truth value adder lb located at the previous stage multiplied by a prescribed weight value, and is constituted by a basic unit 1 without the threshold processing section 4.

Reference numeral 40 denotes an antecedent truth value calculation layer constituting the hierarchical network unit 11, which receives and receives input control state quantities and calculates truth values of membership functions for the control state quantities; The first rule layer that constitutes the hierarchical network section 11 is the antecedent truth value calculation layer 4.
The grade value is calculated by executing the function conversion process assigned to the truth value of the member knob function for the control state quantity calculated in A rule layer, which calculates a grade value by executing the interval conversion process assigned to the grade value calculated in the first rule layer 41; The truth value processing layer calculates the sum of membership functions for the control operation amount that is contracted and expanded based on the grade value calculated in the second rule layer 42, and 44 is a layer. A centroid derived value calculation layer constituting the network unit 11, which calculates the centroid derived value used when determining the centroid value, which is a fuzzy inference value, from the sum of functions calculated in the consequent truth value processing layer 43. It is.

In the embodiment shown in FIG. 5, the antecedent truth value calculation layer 40 is
The input layer (consisting of the human crab unit 12) explained in FIG. 1 and the calculation layer (consisting of the processing unit 13) are configured. Furthermore, the consequent truth value processing layer 43 and the centroid derived value calculation layer 44 constitute the output layer (constituted by the output unit 16) described in FIG. 1.

The antecedent truth value calculation layer 40, as shown in FIG.
The function is to calculate the truth value of the membership function of the input control state quantity, such as the truth value of Xl found by the membership function SA for the control state quantity.

FIGS. 6 and 7 illustrate an embodiment for realizing the functions of the antecedent truth value calculation layer 40.

In the embodiment shown in FIG. 6, when the output value y output from the basic unit 1 is ω×0 θ×0 as shown in FIG. 6(a), the “temperature is Note that when ω〉0 θ〉0, the function shape becomes similar to the membership function of ``J with high temperature'' shown in Figure 8. As shown in FIG. 6(b), this basic unit 1
By appropriately setting the weight value ω and threshold value θ assigned to the input of Discloses something.

Further, in the embodiment shown in FIG. 7, two basic units 1
The difference value y1+eXp (-ω2
As shown in Figure 7(a), x + θ2) is 1! Focusing on the similar function shapes, processing is performed to calculate the difference between the output values of two basic units 1 and these two basic units 1, as shown in FIG. 7(b). The weight values ω2 . By setting ω2 and the threshold value θ1□θ2 appropriately, a method is disclosed that realizes the process of calculating the truth value of a membership function having a function shape in which the temperature is normal j in FIG. 8.

The consequent truth value processing layer 43, as shown in FIG.
The rule layer 42 outputs grade values associated with the membership functions SA, MM, and LA for the control operation amount Y, so the corresponding membership functions for the same control operation amount are reduced according to these grade values ( ) and performs the function of calculating the function sum of the reduced membership functions. FIG. 9 illustrates an embodiment for realizing the functions of the consequent truth value processing layer 43. Although the embodiment shown in FIG. 5 is disclosed with no control operation amount,
When there are multiple types, such as the valve opening degree and the heating amount of the heater, 8! ability will be provided.

In the embodiment shown in FIG. 9, as shown in FIG. 9(a), the membership function for the same amount of control operation is finely partitioned at equal intervals, the truth value y of each partition is specified, and the following To,
As shown in FIG. 9(b), this specified truth value y,
The inputs (
It is disclosed that the function of the consequent truth value processing layer 43 is realized by assigning a weight value "y," to the output from the basic unit l of the second rule layer 42).

As shown in FIG. 5, the gravity center derived value calculation layer 44 outputs the function sum of the membership functions for the reduced control operation amount from the consequent part truth value processing layer 43, so It executes the function of calculating two centroid derived values Y, , Y, which are required when determining the centroid. FIG. 10 illustrates an embodiment for realizing the function of the center of gravity derived value calculation layer 44.

In general, in fuzzy controllers, 5 grade(y) is obtained by finding the center of gravity of the sum of functions of the reduced member jump functions for the same control operation amount according to the following formula.
dy A method is used to calculate the control operation amount, which is a fuzzy inference value. Therefore, in the embodiment shown in FIG. 10, two of the above-mentioned barycenter derived value calculators 1c are prepared, and the weight value of the internal connection with the truth value adder lb of the consequent truth value processing layer 43 is For one center of gravity derived value calculator 1e, the control operation amount corresponding to the truth value adder 1b is increased at equal intervals from 0 to l in increasing order starting from the minimum value. In addition to assigning a weight value to
For the other center of gravity derived value calculator 1c, the control operation amount associated with the truth value adder 1b is decreased at equal intervals from 0 to -l in descending order starting from the maximum value. The present invention discloses a process for calculating two centroid derived values y, , yt by assigning weight values. The final centroid value is determined using the output values of these two centroid derived value calculators 1c in accordance with arithmetic means prepared separately from the hierarchical network unit 11.

That is, with this configuration, one center of gravity derived value calculator 1
For example, if the truth value adder 1b has 6 units, c is Yl-0・C,+0.2・C2+0.4・C3+0.6
・c, +o, B ・C9+l・C6 However, C8 outputs the output value of the truth knowledge adder 1b, whereas the other center of gravity derived value calculator 1c outputs Y,=-1-C,- 0,8・cz-o, s・C1-0,
Since it outputs 4・C,−0,2・C5−0・C6, the output values Y + , Y z of this centroid derivation f direct calculator 1e
Using Yz I Yz , we get Y.

Yz C, +C2+C, 3+Ca+Cs+Ch, etc.
The center of gravity value explained using equation (13) above is determined.

Hierarchical network section]l configured as above, between the basic unit 1 of the antecedent truth value calculation layer 40 and the subtractor 1a and the basic unit 1 of the first rule layer 41,
Between the basic unit 1 of the first rule layer 41 and the basic unit 7) 1 of the second rule layer 42, there is a partial non-zero weight value, for example, in a format that follows the description of the fuzzy control rule. By being allocated, an internal link is formed. Note that between the input unit 1' and the basic unit 1 in the antecedent truth value calculation layer 40, the membership function for which the truth value is calculated is controlled differently, such as a discomfort index defined by temperature and humidity. When related to state quantities, a configuration is adopted in which multiple items are internally coupled.

In the present invention, undetermined weight values among the internal connection weight values of the hierarchical network unit 11 configured as described above are changed according to the weight value changing unit 30a using actual control data obtained from the controlled object. You will have to process it as you learn. That is, the operator calculates the truth value of the membership function of the control state quantity described in the fuzzy control rule for each internal combination of the inputs of the basic UNISO) 1 of the antecedent truth value calculation layer 40. At the same time, the membership function of the control operation amount described in the fuzzy control rule is set for the internal combination of each input of the truth value adder 1b of the consequent truth value processing layer 43. When weight values are set to express the function pattern of In order to learn the weight value,
The actual control data group obtained from the controlled object is registered in the learning signal storage unit 21 of the learning signal presentation device 20, and then,
By instructing the learning signal presentation device 20 to present the registered control data group to the hierarchical network unit 11 and the weight value changing unit 30a, the hierarchical network unit 11 is instructed to perform learning of the weight values. do.

When receiving control state quantity data for learning from the learning signal presentation device 20, the human crab unit l of the hierarchical network unit 11
”, this control state quantity data is used in the antecedent truth value calculation layer 40.
to the corresponding basic units 1. Upon receiving this control state quantity data, the basic unit 1 and subtractor 1a of the antecedent truth value calculation layer 40 calculate the truth value of the membership function for the control state quantity according to the above-mentioned configuration, and then The calculated truth value is output to the internally connected basic unit 1 of the first rule layer 41 of the stage. When this truth value is received, the basic unit l of the first rule layer 41 calculates the grade value according to the above-mentioned formula (2), and calculates the grade value according to the above-mentioned formula (2), and then calculates the grade value according to the above-mentioned formula (2), and then calculates the grade value according to the above-mentioned formula (2), and then calculates the grade value according to the above-mentioned formula (2). )
After outputting the calculated grade value for 1 and receiving this grade value, the basic unit 1 of the second rule layer 42 calculates the grade value according to the above equation (2) and calculates the grade value for the next stage. The calculated grade value is output to the truth value adder 1b which is internally connected in the consequent part truth value processing layer 43.

Then, when the grade value from the basic unit 1 of the second rule layer 42 is received, the truth value adder 1b of the consequent part truth value processing layer 43 calculates the member for the control operation amount to be reduced according to the grade value. Calculate the function of the 7-p function and output the calculated function to the center of gravity derived value calculator IC which is internally coupled in the next stage gravity center derived value calculation layer 44, and then take this function. , the centroid derived value calculator 1C of the centroid derived value calculation layer 44 calculates the two centroid derived values Y1. Y
z is calculated and output, and in response to this output, a calculation means (not shown) calculates control operation amount data, which is the final output, and outputs it to the weight value changing section 30a and the learning convergence determining section 23.

When the control operation amount data is received from the hierarchical network unit 11 in this way, the weight value changing unit 30a
7) according to formula α2. The update amount of the weight value of the internal connection between the gravity center derived value calculator 1c and the truth value adder 1b is calculated according to equation (8).

Since the weight value of this internal connection should be set fixedly, the process of updating the weight value according to the calculated update amount is not performed. In addition, at the start of this learning process, it is required to calculate the above-mentioned two Y+ and Yt used as teacher signals from the control operation amount data for learning. This is carried out by a method such as determining the intersection of a straight line with a specific slope passing through the quantitative data value and the left and right end points of the consequent membership function.

Subsequently, the weight value changing unit 30a uses the calculated α and J to calculate βp according to the above-mentioned formula (9), and obtains (1
0) truth value adder 1b and second rule layer 42 according to formula
The update amount of the weight value of the internal connection with the basic unit 1 is calculated. This weight value will be the learning target when tuning the membership function for the control operation amount, but it is not necessary to use it as the learning target when an approximately appropriate membership function is being obtained. Therefore, it is not necessary to perform the process of updating the weight value according to the calculated update amount.

Subsequently, the weight value changing unit 30a changes the calculated β2. Teh is calculated according to the above equation (11) using (
12) The basic unit l of the second rule layer 42 according to
The update amount of the weight value of the internal connection between the basic unit 1 of the first rule layer 41 and the basic unit 1 of the first rule layer 41 is calculated.

Since the weight value of this internal connection is to be learned, the weight value changing unit 30a calculates the weight value for the next update cycle according to the calculated update amount and stores it in the weight value management unit 17a. do.

Subsequently, the weight value changing unit 30a performs the same process as follows.
The update amount of the weight value of the internal connection between the basic unit 1 of the first rule layer 41 and the basic unit 1 of the antecedent truth value calculation layer 40 and the subtractor 1a is calculated. Since the weight value of this internal connection is also a learning target, the weight value changing unit 30a calculates the weight value for the next update cycle according to the calculated update amount and stores it in the weight value management unit 17a. do. Then, the weight changing unit 30a sets the weight values of internal connections in the antecedent truth value calculation layer 40 as learning targets when an approximately appropriate membership function of the control state quantity is required. Since this is unnecessary, the weight value updating process can be terminated at this point.

If the learning convergence determination unit 23 of the learning signal presentation device 20 does not determine learning convergence of the weight values, the learning signal presentation unit 22 will present the same learning control data group again. The weight value changing unit 30a changes the second rule layer 42 by repeating the same procedure as the above-described process.
The weight value of the internal connection between the basic unit 1 of the first rule layer 41 and the basic unit l of the first rule layer 41, the basic unit 1 of the first rule layer 41 and the basic unit l of the antecedent truth value calculation layer 40, and The process of updating the weight value of the internal connection with the subtractor 1a continues. Note that the initial values of the weight values of these internal connections at the start of learning are set, for example, as random numbers.

Through this learning process, the basic Unison) 1 and the first
The weight value of the internal connection between the basic unit 1 of the rule layer 41 and the basic unit 1 of the first rule layer 41 and the basic unit 1 of the antecedent truth value calculation layer 40 and the subtractor 1a. The weight value of the internal connection is determined, and this weight value represents that the hierarchical network unit 11 is a weight value that enables the control target to be accurately controlled according to the fuzzy control rule. By finally registering the learned weight values in the weight value management section 17a, the hierarchical network section 11 can function as a desired fuzzy controller.

The embodiment disclosed in FIG. 5 includes two rule layers, a first rule layer 41 and a second rule layer 42, corresponding to the antecedent part calculation and the consequent part calculation of the fuzzy control rule. The present invention was disclosed in the above, but this first rule layer 4
It is also possible to provide a rule layer between the first rule layer 42 and the second rule layer 42 that has one or more hierarchical network configurations. In this way, multi-stage reasoning becomes possible, so it becomes possible to construct a fuzzy controller that executes more complex control logic.

Further, in the embodiment shown in FIG. 5, the first rule layer 41 and the second rule layer 42 execute a function calculation process (rule calculation process) for determining a grade value according to one basic unit 1. However, the hierarchical network part 1
It is also possible to configure the grade value to be determined according to a method that utilizes a partial hierarchical network within 1. That is, a group of basic units for realizing rule calculations is secured in the hierarchical network section 11, and a fixed set of units learned to realize a prescribed rule calculation for the weight values of internal connections of this group of basic units is secured. The grade value may be determined by setting a value. In this way, more complex calculation functions can be realized as rule calculations.

In the embodiment shown in FIG. 5, the arithmetic processing of the basic unit 1 involves converting the sigmoid function expressed by equation (2). It is also possible to realize this by applying processing.

A method of configuring the hierarchical network section 11 using hardware components is described in the patent application filed by the applicant in 1983-21.
No. 6865 (filed on August 31, 1988, "Network Configuration Data Processing Apparatus")" can be used.

That is, the basic unit 1, as shown in FIG.
A multiplication type D/A converter 2a that multiplies the output from the previous layer input via the input switch unit 7 by the weight value held by the weight value holding unit 8; and a multiplication type D/A converter 2a.
An analog adder 3a calculates a new cumulative value by adding the output value of and the previous cumulative value, and a sample hold circuit 3b holds the addition result of the analog adder 3a. A nonlinear function generation circuit 4a that nonlinearly converts data held by the sample and hold circuit 3b, an output holding section 5 that holds an analog signal value of the nonlinear function generation circuit 4a that is output to a subsequent layer, and an output holding section. This is realized by including an output switch section 6 that outputs the held data of No. 5, and a control circuit 9 that controls each of these processing sections. Here, as for the subtracter 1a, the truth value adder 1b, the center of gravity derived value calculator 1c, and the average value calculator 1d, the data held by the sample and hold circuit 3b is outputted without going through the nonlinear number generation circuit 4a. A configuration in which the device is held by the holding section 5 will be adopted.

The hierarchical network unit 11 is realized in a configuration in which basic units 1 having this configuration are electrically connected by one common analog bus 70, as shown in FIG. Here, in the figure, 71 is a weight output circuit that gives a weight value to the weight holding unit 8 of the basic unit I, 72 is an initial signal output circuit corresponding to the input unit I, and 73 is a synchronization signal that is a control signal for data transfer. Control signal weight output circuit 71
, a synchronous control signal line 74 that transmits the synchronous control signal to the initial signal output circuit 72 and the control circuit 9, is a main control circuit that sends out the synchronous control signal.

In the hierarchical network unit 11 having this configuration, the main control circuit 74 sequentially selects the basic units 1 in the previous layer in chronological order, and in synchronization with this selection process, the output holding unit of the selected basic unit 1. 5 is processed so as to be outputted to the multiplication type D/A converter 2a of the basic unit 1 in the subsequent layer via the analog bus 70 in accordance with the time-division transmission format. When this input is received, the multiplication type D/A converter 2a of the basic unit 1 in the subsequent layer sequentially selects the corresponding weight value, multiplies the input value and the weight value, and connects the analog adder 3a. The accumulation processing section 3 constituted by the sample and hold circuit 3b sequentially accumulates the multiplied values.

Subsequently, when all the accumulation processing for the basic unit 1 in the previous layer is completed, the main control circuit 74 activates the nonlinear function generation circuit 4a of the basic unit 1 in the subsequent layer to calculate the final output. The output holding unit 5 performs processing to hold the final output of the conversion processing result. And the main control circuit 7
4, by repeating the same process for the next subsequent layer using this latter layer as a new previous layer, processing is performed so that an output cover turn corresponding to an input cover turn is output.

The hierarchical network unit 11 can also be configured by software means instead of being configured by such hardware components. 1111 in this way, the truth value of the membership function for the control state quantity is managed accurately using a table means, etc., and the antecedent part truth value calculation layer 40 is searched for the managed membership function.
Since it is possible to realize the calculation process of the truth value of the membership function for the control state quantity, which is the function executed by the system, it is no longer limited to the membership function having the function shape as explained in FIG. Further, as the rule calculation processing of the first rule layer 41 and the second rule layer 42, it becomes possible to employ more complicated rule calculation processing without being restricted by the calculation processing of the basic unit 1.

Although the illustrated embodiment has been described, the present invention is not limited thereto. For example, the applicant previously
Japanese Patent Application No. 1983-227825 (filed on September 12, 1988, "Learning Processing Method for Network Configuration Data Processing Device") aims to improve the back propagation method to learn weight values in a shorter time. Although the present invention has disclosed an invention that enables the processing to be realized, the present invention can also utilize other weight value learning methods other than such improved pack propagation method or bank propagation method. . The method of implementing the arithmetic processing of each unit constituting the hierarchical network unit 11 is not the essence of the present invention, and any configuration method may be used.

〔Effect of the invention〕

As explained above, according to the present invention, since a fuzzy controller is constructed according to a hierarchical network structure, a desired control can be created according to mechanical processing. I
You will be able to build fuzzy controllers that perform logic. From now on, the burden of fuzzy control rule tuning on the operator can be significantly reduced.

Rather than using the hierarchical network structure of a known hierarchical network data processing device as it is, we decided to use a structure that was structured in a format that conformed to fuzzy control rules. Since the fuzzy control rules mapped onto the structure can be seen from the outside, consistency with conventional fuzzy controllers can also be maintained.

[Brief explanation of drawings]

Fig. 1 is an illustration of the principle of the present invention, Fig. 2 is an embodiment of the present invention, Fig. 3 is a basic configuration diagram of a basic unit, Fig. 4 is a basic configuration diagram of a hierarchical network, and Fig. 5 is a hierarchy diagram. An example of the network part, Figures 6 and 7 are examples of the antecedent truth value calculation layer, Figure 8 is an explanatory diagram of the membership function, and Figure 9 is an example of the consequent truth value processing layer. FIG. 10 is an example of the gravity center derived value calculation layer, FIG. 11 is an example of the basic unit, and FIG. 12 is an example of the hierarchical network part.
is the basic unit, 1 is the input unit, 2 is the multiplication processing unit, 3 is the accumulation processing unit, 4 is the dark value processing unit, lO is the hierarchical network m * fuzzy controller, 11 is the hierarchical network unit, 12 is the input unit , 13 is a processing unit, 14 and 15 are rule units, 16 is an output unit, 17 is an internal state value management unit, 20 is a learning signal presentation device, 21 is a learning signal storage unit, 22 is a learning signal presentation unit, and 23 is a learning unit. Convergence determination unit, 30 is a learning processing device, 40 is an antecedent truth value calculation layer, 41 is a first rule layer, 42 is a second rule layer, 43
is a consequent truth value processing layer, 44 is a center of gravity derived value calculation layer, 45
is the average value calculation layer.

Claims (2)

[Claims] (1) A fuzzy controller with a hierarchical network configuration that calculates and outputs a control operation amount corresponding to an input control state amount by executing a fuzzy control rule according to the hierarchical network unit (11), a calculation layer configured by a plurality of processing units (13) in which the unit (11) calculates and outputs the truth value of the antecedent membership function of the fuzzy control rule using the control state quantity as input; A grade value is calculated and output according to the assigned operation, using as input the multiplication value of the output value of the corresponding processing unit (13) and the internal state value assigned to the internal connection between the processing unit (13). a first rule layer composed of a plurality of rule units (14); and an internal connection between the output value of the corresponding rule unit (14) of the first rule layer and the rule unit (14). A second rule unit (15) comprising a plurality of rule units (15) that calculates and outputs a grade value according to an assigned operation using as input a multiplication value with an internal state value.
an output unit that calculates and outputs the control operation amount or its corresponding value from the output value of the corresponding rule unit (15) of the second rule layer and the consequent membership function of the fuzzy control rule. (15); and a desired control operation amount or its corresponding value from the output layer when the control state quantity is given to the calculation layer as the internal state value. A hierarchical network configured fuzzy controller, characterized in that the controller is configured to set a value that will be outputted. (2) In the hierarchical network configuration fuzzy controller according to claim (1), between the first rule layer and the second rule layer, the output value of the corresponding rule unit of the previous rule layer and the rule unit are connected. There is one or more rule layers composed of one or more rule units that calculate and output a grade value according to the assigned operation, taking as input the multiplication value with the internal state value assigned to the internal connection between A hierarchical network configured fuzzy controller, characterized in that the controller is configured to have multiple layers.