JP2763370B2 - Hierarchical network configuration fuzzy controller - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a hierarchical network configuration fuzzy controller that functions as an adaptive fuzzy controller.

Fuzzy control is becoming popular as a new control processing method. In this fuzzy control, a control algorithm including ambiguity such as human judgment is expressed in an if-then form, and the control algorithm is executed in accordance with fuzzy inference. This is to calculate and control the control target. Since the fuzzy controller that executes the fuzzy control handles indefinite ones, it is necessary to adopt a configuration that can adaptively cope with adjustment and change of the control algorithm.

[Conventional technology]

The fuzzy control rule is of the form if x ₁ is big and x ₂ is small then y ₁ is big (the IF part is called the antecedent part and describes the conditions for the control state quantity such as temperature data. , THEN part is called a consequent part, and is a part that describes the condition of the control operation amount such as the operation end). The fuzzy controller describes the control logic of the control object. By managing such multiple fuzzy control rules prepared as, and quantifying the meaning of ambiguous linguistic expressions such as "large" and "small" described in each fuzzy control rule as a membership function The configuration to manage is adopted.

When the fuzzy controller is given a control state quantity such as temperature data and water level data from the control object, first, a membership function of the control state quantity being managed (becomes a membership function of the antecedent part). Calculates the membership function value (truth value) of the control state quantity given from. Next, according to the antecedent operation such as selecting the minimum value,
A process for determining an applied value to the consequent part in each fuzzy control rule is executed. That is, if the truth value of “x ₁ is big” is “0.8” and the truth value of “x ₂ is small” is “0.5”, the minimum value According to the antecedent operation to select the value,
Processing is performed so as to determine “0.5” as an applied value to the consequent part of the fuzzy control rule.

Subsequently, the fuzzy controller, in accordance with the consequent part operation such as selecting the maximum value, calculates the fuzzy control rules given for the same membership function of the same control operation amount (which becomes the consequent part membership function). From the applied value to the consequent part, a process of determining an applied value to the membership function is executed. That is, the applied value "0.5" for "y ₁ IS big" of the consequent part in the above-described fuzzy control rules, on the other hand, with a different fuzzy control rules for "y ₁ IS big""0.6" In the case where is set as the applied value, the "0.6"
Than it is processed to determine the applied value of the for the "big" control operation amounts y ₁ a.

Subsequently, the fuzzy controller reduces the membership function of the control operation amount in accordance with the determined applied value, and also obtains the fuzzy center of the function sum of the membership function for the same control operation amount. A process of calculating a control operation amount which is an inference value is executed, and a process of outputting the control operation amount to an operation terminal or the like is executed.

The conventional fuzzy controller employs a configuration in which the execution of such fuzzy control rules is implemented by executing the program. That is, according to the software means of the computer system that performs the sequential processing,
First, the truth value of the control state quantity is calculated, then, the application value for the consequent part is determined according to the antecedent operation, and then the membership function of the control operation amount according to the consequent operation. , The control operation amount is calculated and output by the center of gravity calculation or the like.

[Problems to be solved by the invention]

However, the fuzzy control rules are generated in accordance with the knowledge of the target process of the operator who is familiar with the control of the control target. From now on, the fact is that it is not possible to generate fuzzy control rules that can achieve the desired control from the beginning, and we will tune it by trial and error while evaluating the generated fuzzy control rules by simulation and on-site tests. Therefore, it is necessary to take a procedure of completing a fuzzy control rule suitable for the control target.

However, in the conventional method of constructing a fuzzy controller using software means, the fuzzy controller simply executes the control logic in accordance with the description procedure of the generated fuzzy control rule and has no adaptability. is not. Therefore, the conventional fuzzy controller does not exert a function of reducing the load of tuning the fuzzy control rules for the operator.

The present invention has been made in view of such circumstances,
The provision of a hierarchical network configuration fuzzy controller that constitutes a new fuzzy controller that adopts a configuration that can adaptively respond to adjustment or change of the fuzzy control rule while adopting a configuration for realizing the fuzzy control rule. It is the purpose.

[Means for solving the problem]

 FIG. 1 is a diagram illustrating the principle of the present invention.

In the figure, reference numeral 10 denotes a hierarchical network configuration fuzzy controller according to the present invention, which adopts a configuration for realizing a fuzzy control rule and has a configuration capable of adaptively coping with adjustment and change of the fuzzy control rule. 11 is a fuzzy controller having a hierarchical network configuration.
A fuzzy control rule according to a data conversion function defined by a hierarchical network structure and an internal state value assigned to an internal connection of the hierarchical network structure, and 16 is an internal state value management unit. The hierarchical network unit 11 manages internal state values required at the time of execution of the fuzzy control rule, and 20 is a learning signal presenting device, which stores a control state amount data group for learning in the hierarchical network unit 11. A control operation amount data group for learning, which is presented to the input layer and forms a pair with the control state amount data for learning (when a control state amount data for learning is provided, a desired control state is realized. , Which is a control operation amount for learning), is presented to the learning processing device 30 described below. Between the control operation amount data group from the hierarchical network unit 11 output in response to the presentation of the learning control state amount data group and the learning control operation amount data group presented from the learning signal presentation device 20 Is calculated, and the internal state value to be learned is sequentially updated based on the error value, thereby learning the internal state value in which the error value is within an allowable range.

The hierarchical network unit 11 includes an input layer constituted by a plurality of input units 12 provided in a format corresponding to the input control state quantity, and a layer subsequent to the input layer, which is a member of the antecedent of the fuzzy control rule. A first calculation layer constituted by a plurality of processing units 13 provided in a format corresponding to the ship function, and one or more rule units 14 positioned as a subsequent layer of the first calculation layer Calculate the function sum of an intermediate layer composed of one or a plurality of rule layers with the rule layer as a layer unit and a consequent part membership function whose size is changed, positioned as a subsequent layer of the intermediate layer. A second calculation layer composed of a plurality of processing units 15a and a pair of two calculation position derivation values that are located as subsequent layers of the second calculation layer are associated with the control operation amount. Set Configured with a hierarchical network structure that constituted the output layer, the output unit 15b to be.

The input unit 12 of the input layer receives the control state quantity of the control target as an input, and the corresponding processing unit of the first calculation layer.
The input control state quantity is distributed to 13.

The processing unit 13 of the first calculation layer receives a multiplied value of a control state amount notified from the input unit 12 and an internal state value allocated to an internal connection between the input unit 12 and a prescribed operation. To calculate the truth value of the membership function of the antecedent part of the fuzzy control rule,
All rule units 14 in the first rule layer of the middle layer
Output to

The rule unit 14 of the foremost rule layer of the intermediate layer receives, as an input, a multiplied value of a plurality of truth values notified from the processing unit 13 and an internal state value assigned to an internal connection between the processing units 13. By executing the assigned operation, the grade value is calculated and processed so as to be output to all the rule units 14 of the next rule layer, and the rule units 14 of each rule layer are compared with the rule units 14 of the preceding rule layer. A grade value is calculated by executing an assigned operation by inputting a multiplied value of the grade value notified from the internal unit and an internal state value assigned to an internal connection between the rule units 14 and calculating a grade value, thereby executing a subsequent rule layer. Is processed so as to be output to all rule units 14 of the intermediate layer, and the rule unit 14 of the last rule layer of the middle layer is transmitted from the rule unit 14 of the previous rule layer A grade value is calculated by inputting a multiplied value of a grade value to be assigned and an internal state value assigned to an internal connection between the rule units 14 and executing an assigned operation to obtain a grade value. Output to the corresponding processing unit 15a of the layer.

The processing unit 15a of the second calculation layer includes a grade value notified from the rule unit 14 of the last rule layer of the intermediate layer, and a consequent membership function assigned to the internal connection between the rule unit 14 and the grade value. The multiplication value with the division value (the discrete function value cut out from the consequent part membership function) is input, and the function sum of the consequent part membership function whose size is changed by the grade value is calculated (each processing The unit 15a calculates a partial value of the function sum, and the sum is a function sum), and outputs the corresponding output unit of the output layer.
Output to 15b.

The output unit 15b of the output layer includes a function sum partial value of the consequent part membership function notified from the processing unit 15a,
A derived value used for deriving a control operation amount is obtained by taking as input a multiplied value by a constant division value (a numerical value obtained by partitioning between predetermined constant values) allocated to the internal connection with the processing unit 15a. calculate.

[Action]

In the hierarchical network unit 11 of the present invention, when a control state quantity is given via the input unit 12, the processing unit 13 of the first calculation layer performs a function conversion function assigned to the control state quantity. Then, a truth value of the membership function for the control state quantity, which is an operation result value, is calculated, and the calculated truth value is distributed and output to the rule unit 14 of the foremost rule layer of the intermediate layer. When the truth value of the membership function for the control state quantity is received from the processing unit 13 in this manner, the rule unit 14 of the intermediate layer executes the operation processing such as the function conversion function assigned to the intermediate unit, thereby obtaining the operation result value and By calculating the grade value and outputting the calculated grade value to the subsequent rule unit 14, the processing of the second calculation layer in which the grade value, which is the operation result value as the intermediate layer, is internally combined is performed. Output to the unit 15a.

Upon receiving this grade value, the processing unit 15a
In accordance with the consequent part membership function division value assigned to the inner connection with the rule unit 14, the function sum of the consequent part membership function whose size is changed by the grade value is calculated, and the calculated function is calculated. The sum partial value is output to the output unit 15b of the output layer that is internally coupled. Then, upon receiving the partial value of the function sum of the consequent part membership function, the output unit 15b outputs the derived value used for deriving the control operation amount according to the constant division value allocated to the internal connection with the processing unit 15a. Is calculated.

As described above, 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, the hierarchical network configuration fuzzy controller 10
If you want to construct as the desired fuzzy controller,
The operator firstly asks the processing unit 13
A function for calculating the truth value of the membership function for the generated control state quantity is assigned, and the membership function for the generated control operation quantity is registered in an internal connection between the rule unit 14 and the processing unit 15a. And
The internal connection between the processing unit 13 of the first calculation layer and the rule unit 14 of the foremost rule layer of the intermediate layer and the internal connection between the rule units 14 of the adjacent rule layer of the intermediate layer are determined by the internal state value. Initial values are set and stored in the internal state value management unit 16.

Subsequently, the operator activates the learning signal presentation device 20 and the learning processing device 30. When activated in this way,
The learning signal presenting device 20 presents the learning control state amount data to the input layer of the hierarchical network unit 11 and presents the learning control operation amount data to the learning processing device 30. In response to the presentation of the learning signal, the learning processing device 30 outputs the control operation amount data from the hierarchical network unit 11 output in response to the presentation of the learning control state amount data, to the learning control amount to be presented. The internal state value of the internal connection between the processing unit 13 of the first calculation layer and the rule unit 14 of the foremost rule layer of the intermediate layer and the adjacent state of the intermediate layer so as to substantially match the operation amount data A process of learning the internal state value of the internal connection between the rule units 14 of the rule layer and setting the learned value of the obtained internal state value in the internal state value management unit 17 is executed.

In the learning process of the internal state value, the internal state value of the internal connection set between the input unit 12 and the processing unit 13 and the rule unit 14 and the processing unit of the last rule layer of the intermediate layer are set. It is possible to set the internal state value of the internal connection set between the target and the target 15a as the learning target.

Assuming that the internal state value of the internal connection set between the input unit 12 and the processing unit 13 is to be learned, the membership function of the control state quantity assumed in advance is further improved by a new internal state value obtained by this learning processing. Tuning can be performed to match the control target.

Further, when learning the internal state value of the internal connection set between the rule unit 14 and the processing unit 15a, the membership function of the control operation amount assumed in advance is calculated by the new internal state value obtained by this learning processing. It is possible to tune to one that is more suitable for the control target.

When the learning value of the internal state value is set, the hierarchical network configuration fuzzy controller 10 functions as a fuzzy controller that executes a 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 the fuzzy control rules.

As is well known as a neurocomputer, a data processing device that follows a data conversion function of a hierarchical network structure, when an internal state value assigned to an internal connection of the hierarchical network structure is set, an unexpected input signal is generated. This realizes "soft" data processing of outputting an appropriate output signal even if it is input. According to the present invention, the hierarchical network structure of such a hierarchical network configuration data processing device is structured so as to be adapted to fuzzy control rules and provided as a hierarchical network unit 11. Then, the internal state value of the internal connection of the rule part of the fuzzy control rule can be learned separately from the membership function for the previously generated control state amount and control operation amount.

By using such a hierarchical network configuration data processing device, the learning processing device 30 can mechanically determine the association of the fuzzy control rules. Can be greatly reduced. And
Instead of using the hierarchical network configuration data processing device as it is, using the hierarchical network unit 11 structured in a format that conforms to the fuzzy control rules, the membership function is tuned separately with the membership function. And the consistency with the conventional fuzzy controller can be maintained.

〔Example〕

 Hereinafter, the present invention will be described in detail with reference to examples.

FIG. 2 shows an embodiment for constructing the hierarchical network configuration fuzzy controller 10 of the present invention. In the figure, the same components as those described in FIG. 1 are indicated by the same symbols. 1'-h is an input unit corresponding to the input unit 12 of FIG. 1, 1-i is a processing unit 13 of FIG. 1 which is a basic unit constituting a rule unit 14,
1-j is a basic unit constituting the output unit 15 of FIG. 1, and 16a is a weight value management unit corresponding to the internal state value management unit 16 of FIG. Is managed (corresponding to the internal state value in FIG. 1).

Reference numeral 21 denotes a learning signal storage unit, and a hierarchical network unit 11
And a learning signal presenting unit 22 for storing a learning signal (constituted as a basic unit of a pair of control state amount data and control operation amount data) used for learning the weight value of the learning signal. The learning signal is read from the storage unit 21, the control state amount data therein is presented to the hierarchical network unit 11, and the other control operation amount data forming a pair is described below with a weight value changing unit 30a. Learning convergence judgment unit 2
What is presented in 3 is a learning convergence determination unit 23, which 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, and It is for determining whether or not an error in the data processing function of the unit 11 is within an allowable range, and notifying the learning signal presentation unit 22 of the determination result.

30a is a weight value changing unit corresponding to the learning processing device 30 in FIG. 1, and receives learning control operation amount data from the learning signal presenting unit 22 and network output data from the hierarchical network unit 11, 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 the convergence value.

As described 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 rules. Hierarchical network unit 11 characteristic of the present invention
Prior to the description, a hierarchical network structure (hereinafter, referred to as a known hierarchical network structure) provided in a generally known hierarchical network configuration data processing device will be described, and the weight value changing unit 30a will be described. The back propagation method, which is a learning algorithm to be implemented, which is a learning algorithm for determining the value of the weight value of the internal connection of the known hierarchical network structure, will be described in detail.

In a known hierarchical network structure, a hierarchical network is basically configured from a kind of node called a basic unit and an internal connection having a weight value. FIG. 3 shows the basic configuration of the basic unit 1. This basic unit 1
A multi-processing unit 2 for multiplying a plurality of inputs by respective weights of the internal connection, an accumulating unit 3 for adding all multiplication results thereof, And a threshold processing unit 4 for performing a non-linear threshold processing on the output of the first and second outputs.

Assuming that the h-th layer is the first layer and the i-th layer is the second layer, the accumulation processing unit 3 executes the operation of the following expression (1), and the threshold processing unit 4 executes the operation of the following expression (2).

y _pi = 1 / (1 + exp (−x _pi + θ _i )) (2) where h: unit number of h layer i: unit number of i layer p: pattern number of input signal θ _i : i number of i layer Unit threshold value W _ih : Weight value of internal connection between _{hi and} i layers x _pi : Sum of products of inputs from each unit of h layer to unit i of y layer y _ph : h layer of input signal of p-th pattern Output from the h-th unit y _pi : Output from the i-th unit in the i-th layer with respect to the input signal of the p-th pattern As shown in FIG. 4, an input unit 1 'for directly distributing and outputting as an input layer is hierarchically connected as shown in FIG. (Output pattern) parallel data processing It will exert its logical functions.

In the bag propagation method, learning is performed by adaptively and automatically adjusting the weight value W _ih and the threshold value θ _i of the hierarchical network by error feedback. (1)
As is apparent from the equation (2), the adjustment of the weight value W _ih and the threshold value θ _i needs to be performed simultaneously, but this operation is a difficult operation that interferes with each other. Therefore, the applicant has
As previously disclosed in "Japanese Patent Application No. 62-333484 (filed on December 28, 1987," Network Configuration Data Processing Apparatus ")," 1 "is always output to the h layer on the input side. In addition, it has been proposed to provide a unit for assigning the threshold value θ _i as a weight value to the output, thereby incorporating the threshold value θ _i into the weight value W _ih and treating the threshold value θ _i as a weight value. By doing so, the above (1) and (2)
ceremony, y _pi = 1 / (1 + exp (−x _pi )) (4)

Next, a description will be given of a weight value learning processing method by the back propagation method according to the description format of the expressions (3) and (4). Here, the description will be made on the assumption that the hierarchical network unit 11 has a three-layer hierarchical network of h layer-i layer-j layer shown in FIG.

By analogy with equations (3) and (4), the following equations (5) and (6)
An expression is obtained. That is, y _pj = 1 / (1 + exp (−x _pj )) (6) where j: the unit number of the j-th layer W _ji : the weight of the internal connection between the i-j layers x _pj : the j-th layer from each unit of the i-th layer Y _pj : output from the j-th unit in the j-th layer with respect to the p-th pattern input signal. The weight value changing unit 30a outputs when the input pattern for learning is presented. Output pattern y _pj from output layer
And a teacher pattern d _pj which is a signal to be taken by the output pattern y _pj (j layer j for the p-th pattern input signal)
Upon receiving the teacher signal) and to the second unit, first difference value between the first output pattern y _pj and the teacher pattern d _pj [dy
_pj− y _pj ], and then α _pj = y _pj (1−y _pj ) (d _pj −y _pj ) (7) Here, the update amount ΔW _ji (t) of the weight between the i-th layer and the j-th layer is calculated according to ε: learning constant ζ: momentum t: learning number. Here, for adding those pertaining to the update amount of the weight values determined in the previous update cycle of "ζΔW _ji (t-1)" is order to speed up the learning.

Subsequently, the weight value changing unit 30a firstly uses the calculated α _pj to , Then , The update amount ΔW _ih (t) of the weight value between the h layer and the i layer
Is calculated.

Subsequently, the weight value changing section 30a, the weight value _{W ji (t) = W ji} (t-1) + ΔW ji (t) W ih (t) = W for the next update cycle according to the update amount calculated as described above By repeating the method of determining _ih (t−1) + ΔW _ih (t), an output pattern y _pj from the output layer that is output when an input pattern for learning is presented, and its output weight value W _ji where the teacher pattern d _pj is the signal to be taken by the pattern y _pj is to be consistent, and treated to learn the W _ih.

Then, the hierarchical network unit 11 includes the g-layer-h-layer-i-layer-
When a four-layer hierarchical network called a j-layer is provided, the weight value changing unit 30a firstly , Then The update amount ΔW _hg (t) of the weight value between the g layer and the h layer is calculated according to the following formula. It is processed so that it is determined while using it.

Here, when the basic unit 1 does not include the threshold processing unit 4, the above equation (7) becomes α _pj = (d _pj −y _pj ) (7), and the above equation (9) becomes And the above equation (11) becomes Becomes

Next, the hierarchical network unit 11 characteristic of the present invention will be described.

FIG. 5 shows an embodiment of the hierarchical network unit 11. In the figure, 1 'is the input unit shown in FIG. 4, 1 is the basic unit shown in FIG. 4, 1a is the weight value "1" and "-" for the output values of the two basic units 1 located at the preceding stage. Subtractors 1a and 1b each configured by a basic unit 1 having a multiplied value of 1 "as an input and having no threshold processing unit 4 multiply output values of a plurality of basic units 1 located at the preceding stage by a specified weight value. Value adder composed of a basic unit 1 having no threshold value processing unit 4 as input, and 1c is an input value obtained by multiplying the output value of the truth adder 1b located at the preceding stage by a specified weight value. And a center-of-gravity derivation value calculator constituted by a basic unit 11 not having the threshold value processing unit 4.

Reference numeral 40 denotes a prerequisite part truth value calculation layer constituting the hierarchical network unit 11, which receives an input control state quantity and calculates a truth value of a membership function for the control state quantity, and 41 denotes a hierarchical network. By performing a function conversion process assigned to the truth value of the membership function for the control state amount calculated by the antecedent part truth value calculation layer 40 in the grade value calculation layer constituting the unit 11 A grade value calculation unit 42 is a consequent part truth value processing layer constituting the hierarchical network unit 11 and controls a control operation amount reduced / expanded based on the grade value calculated by the grade value calculation layer 41. 43 is a centroid derived value calculation layer constituting the hierarchical network unit 11, which is used when calculating a centroid value which is a fuzzy inference value. The center of gravity derived value and calculates from the function sum calculated by the consequent truth treatment layer 42.

The grade value calculation layer 41 corresponds to the intermediate layer described with reference to FIG. 1, and includes one or a plurality of rule layers in which a rule layer configured by one or a plurality of basic units 1 is a layer unit. Will be composed. In the embodiment of FIG. 5, the antecedent part truth value calculation layer 40 constitutes the input layer and the calculation layer of FIG. 1, and the consequent part truth processing layer 42 and the centroid derivation value calculation layer 43 are included. This constitutes the output layer shown in FIG.

As shown in FIG. 5, the antecedent part truth value calculation layer 40 has the membership of the input control state quantity such as the truth value of X1 obtained by the membership function SA for the control state quantity X1. Performs the function of calculating the truth value of a function. 6 and 7 show an embodiment for realizing the function of the antecedent truth value calculation layer 40. FIG.

The embodiment shown in FIG. 6 has an output value y output from the basic unit 1. However, as shown in FIG. 6 (a), when ω <0, θ <0, the function shape becomes similar to the membership function “temperature is high” shown in FIG. 8, and ω> 0 , θ> 0, focusing on the fact that the function shape is similar to the membership function “low temperature” shown in FIG. 8, and as shown in FIG.
By appropriately setting the weight value ω and the threshold value θ to be assigned to the input of, the processing of calculating the truth value of the membership function having the function shape of “low temperature” or “high temperature” in FIG. 8 is realized. Disclose something.

In the embodiment shown in FIG. 7, the difference y between the output values output from the two basic units 1 However, as shown in FIG. 7 (a), focusing on the fact that the function shape is similar to the membership function "temperature is normal" in FIG. 8, and as shown in FIG. 7 (b), It comprises two basic units 1 and the above-described subtractor 1a for processing to calculate the difference between the output values of the two basic units 1, and the weights assigned to the inputs of the two basic units 1 FIG. 8 discloses that a value of ω ₁ , ω ₂ and a threshold θ ₁ , θ ₂ are appropriately set so as to realize a process of calculating a truth value of a membership function having a relational shape of “normal temperature” in FIG. I have.

As shown in FIG. 5, the consequent part truth value processing layer 42 outputs the grade value associated with the membership function SA, MM, or LA for the control operation amount of Y from the grade value calculation layer 41. According to this grade value, the function of reducing the membership function corresponding to the same control operation amount (may be expanded) and calculating the function sum of the reduced membership function is executed. FIG. 9 shows an embodiment for realizing the function of the consequent part truth value processing layer. Although the embodiment of FIG. 5 discloses that there is only one type of control operation amount, when there are a plurality of types such as a valve opening degree and a heating amount of a heater, each control operation amount is reduced. A function for calculating the function sum of the membership functions is provided.

The embodiment shown in FIG. 9 is, as shown in FIG.
The membership function for the same control operation amount is finely divided at regular intervals to identify the truth value y _i of each division,
As shown in FIG. 9 (b), the inputs of the above-described truth value adders 1b prepared for the number of the specified truth values y _i (the basic unit 1 of the last rule layer of the grade value calculation layer 41). (Which is given an output from the data processor) is assigned a weight value “y _i ” to realize the function of the consequent part truth value processing layer 42.

As shown in FIG. 5, the center-of-gravity derivation value calculation layer 43 outputs the function sum of the membership function for the reduced control operation amount from the consequent truth value processing layer 42. The function of calculating _two center-of-gravity derivation values Y ₁ and Y ₂ required for obtaining the center of gravity is executed. FIG. 10 shows an embodiment for realizing the function of the center-of-gravity derived value calculation layer 44.

In general, a fuzzy controller obtains a barycenter of a figure of a function sum of reduced membership functions for the same control operation amount according to the following equation. A method of calculating a control operation amount which is a fuzzy inference value is employed. Therefore, in the embodiment shown in FIG. 10, the above-described two center-of-gravity derived value calculators 1c are prepared, and the weight value of the internal connection with the truth value adder 1b of the consequent part truth value processing layer 42 is provided. As for one of the center-of-gravity derived value calculators 1c, they increase at equal intervals, for example, from 0 to 1 in order of increasing from the minimum value of the control operation amount associated with the truth value adder 1b as a starting point. A weight value is assigned to the other center-of-gravity derived value calculator 1c. For example, from 0 to -1 in the order of decreasing from the maximum value of the control operation amount associated with the truth value adder 1b. A technique is disclosed that realizes a process of calculating _two center-of-gravity derivation values Y ₁ and Y ₂ by assigning weight values that decrease at equal intervals between them. Note that the final barycenter value is configured to be obtained using the output values of the two barycenter derived value calculators 1c in accordance with arithmetic means provided separately from the hierarchical network unit 11.

That is, with this configuration, one center of gravity derived value calculator
1c, for example, if a description will be given of a case truth adder 1b is 6 _{units, Y 1 = 0 · C 1} +0.2 · C 2 +0.4 · C 3 +0.6 · C 4 +0.8 · C ₅ +1 · C ₆ However, C _i outputs the output value of the truth adder 1b, contrast, other centroid derived value calculator. _{1c, Y 2 = -1 · C 1} -0.8 · Since C ₂ −0.6 ・ C ₃ −0.4 ・ C ₄ −0.2 ・ C ₅ −0.C ₆ is output, the output values Y ₁ , Y
_{Using 2} , Calculate Thus, the center of gravity value described in the above equation (13) is obtained.

In the hierarchical network unit 11 thus configured, the basic unit 1 and the subtractor of the truth value calculation layer 40 of the antecedent part
A structure is adopted in which internal connection is made between 1a and the basic unit 1 of the rule layer at the first stage of the grade value calculation layer 41 and between the basic units 1 of the adjacent rule layers of the grade value calculation layer 41. . Note that between the input unit 1 ′ and the basic unit 1 in the antecedent part truth value calculation layer 40, the membership function whose truth value is to be calculated has a different control function such as an unpleasant index defined by temperature and humidity. When it is related to the state quantity, a configuration is adopted in which a plurality of states are internally connected.

In the present invention, the weight values that have not been determined among the weight values of the internal connection of the hierarchical network unit 11 configured as described above are calculated according to the weight value changing unit 30a using actual control data obtained from the control target. It will be processed to 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 the internal connection of each input of the basic unit 1 of the antecedent part truth value calculation layer 40. A weight value and a threshold value are set, and a control operation amount described in a fuzzy control rule is defined as a function of a membership function for an internal connection of each input of the truth value adder 1b of the consequent truth value processing layer 42. When a weight value for expressing a pattern is set, an actual control data group obtained from a control object is presented as a learning signal to learn a weight value assigned to each input of each basic unit 1 of the grade value calculation layer 41. Learning signal storage of device 20
21 and then instructs the learning signal presenting apparatus 20 to present the registered control data group to the hierarchical network unit 11 and the weight value changing unit 30a, so that the weight value of the hierarchical network unit 11 is Instruct the execution of learning.

When receiving the control state quantity data for learning from the learning signal presentation device 20, the input unit 1 'of the hierarchical network unit 11
Distributes the control state quantity data to the corresponding basic unit 1 of the antecedent part truth value calculation layer 40. Upon receiving this control state quantity data, the basic unit 1 and the subtractor 1a of the antecedent part truth value calculation layer 40 calculate the truth value of the membership function for the control state quantity according to the above-described configuration, and calculate the grade value. The calculated truth value is output to all the basic units 1 in the foremost rule layer of the layer 41. Upon receipt of this truth value, the basic unit 1 of the foremost rule layer of the grade value calculation layer 41 calculates the grade value according to the above-described equation (2), and applies it to all the basic units 1 of the next rule layer. And outputs the calculated grade value. Less than,
When receiving the grade value from the previous stage, the grade value calculation layer
The basic unit 1 of each rule layer 41 calculates the grade value according to the above equation (2), and outputs the calculated grade value to all the basic units 1 of the subsequent rule layer. Processing is performed to output the grade value to the consequent part truth value processing layer.

Then, upon receiving the grade value from the grade value calculation layer 41, the truth value adder 1b of the consequent part truth value processing layer 42 calculates the function sum of the membership function for the control operation amount reduced according to the grade value. Then, the calculated function sum is output to the center-of-gravity derivation value calculator 1c which is internally connected to the next-stage center-of-gravity derivation value calculation layer 43. The calculator 1c calculates and outputs the above-described two center-of-gravity derivation values Y ₁ and Y ₂ , and receives this output, and a calculating means (not shown) calculates control operation amount data as a final output, and Change unit 30a and learning convergence determination unit 2
Output to 3.

When the control operation amount data is received from the hierarchical network unit 11 in this manner, the weight value changing unit 30a calculates α _pj according to the above equation (7), and calculates the center of gravity derived value calculator 1c according to the equation (8). The update amount of the weight value of the internal connection with the value adder 1b is calculated. Since the weight value of the inner connection is to be fixedly set, the process of changing the weight value according to the calculated update amount is not performed. At the start of the learning process, the two above-mentioned two signals used as teacher signals from the control operation amount data for learning are used.
It is required to calculate Y ₁ and Y _2. This process is performed, for example, on a straight line having a specific slope passing through the learning control manipulated variable data value and a left and right end point of the consequent part membership function. This is performed by a method such as determination according to the intersection with.

Subsequently, using the calculated α _pj , the weight value changing unit 30a calculates β _pi according to the above equation (9), and calculates the truth value adder 1b and the last of the grade value calculation layer 41 according to the equation (10). The update amount of the weight value of the internal connection with the basic unit 1 of the rule layer of the stage is calculated. This weight value is to be learned when tuning the membership function for the control operation amount, but needless to be learned when an approximately appropriate membership function is obtained. Therefore, it is not necessary to perform the process of updating the weight value according to the calculated update amount.

Subsequently, using the calculated β _pi , the weight value changing unit 30a calculates γ _ph according to the above equation (11), and calculates the basic value of the last rule layer of the grade value calculation layer 41 according to the equation (12). The update amount of the weight value of the internal connection between the unit 1 and the basic unit 1 of the preceding rule layer is calculated. Since the weight value of the inner connection is to be learned, the weight value changing unit 30a calculates a weight value for the next update cycle according to the calculated update amount and stores the weight value in the weight value management unit 16a. I do.

Subsequently, the weight value changing unit 30a performs a similar process,
The update amount of the weight value of the internal connection between the basic units 1 of the rule layer adjacent to the grade value calculation layer 41 to be learned is calculated, and the weight value for the next update cycle is calculated according to the calculated update amount. The basic unit 1 of the foremost rule layer of the grade value calculation layer 41 to be learned, the basic unit 1 of the antecedent part truth value calculation layer 40, and the subtractor 1a are stored in the weight value management unit 16a. , And calculates a weight value for the next update cycle according to the calculated update amount, and calculates a weight value management unit 16a.
To be stored.

Then, the weight changing unit 30a sets the weight of the internal connection in the antecedent part truth value calculation layer 40 to be a learning target when a roughly appropriate membership function of the control state quantity is obtained. Since it is not necessary, the updating process of the weight value can be terminated here.

If the learning convergence determining unit 23 of the learning signal presenting device 20 does not determine the learning convergence of the weight value, the learning signal presenting unit 22 presents the same learning control data group again. The value changing unit 30a repeats the same procedure as the above-described processing, thereby obtaining the weight value of the internal connection between the basic units 1 of the adjacent rule layers of the grade value calculation layer 41 and the maximum value of the grade value calculation layer 41. The update processing of the weight value of the internal connection between the basic unit 1 of the antecedent truth value calculation layer 40 of the basic unit 1 of the preceding rule layer and the subtractor 1a is continued. Note that the initial values at the start of the learning of the weight values of these internal connections are set, for example, in random numbers.

By this learning process, the weight of the internal connection between the basic units 1 of the adjacent rule layers of the grade value calculation layer 41 that defines the association of the learning control data group, and the first rule layer of the grade value calculation layer 41 And the weight value of the internal connection between the basic unit 1 of the antecedent part and the basic unit 1 of the antecedent part truth value calculation layer 40 and the subtractor 1a. This weight value is controlled by the hierarchical network unit 11. This indicates that the weight value allows the object to be accurately controlled according to the fuzzy control rules. From now on, by finally registering the learned weight value in the weight value management unit 16a, the hierarchical network unit 11 can function as a desired fuzzy controller.

In the embodiment of FIG. 5, the grade value calculation layer 41
In which each rule layer executes a function operation process (rule operation process) for obtaining a grade value in accordance with one basic unit 1, uses a partial hierarchical network in the hierarchical network unit 11. It is also possible to configure so as to obtain the grade value according to the method. That is, a basic unit group for realizing a rule operation is secured in the hierarchical network unit 11, and a fixed unit learned to realize a specified rule operation with respect to the weight value of the internal connection of the basic unit group is learned. The grade value may be obtained by setting a proper value. In this way, a more complicated calculation function can be realized as a rule calculation.

In the embodiment shown in FIG. 5, the arithmetic processing of the basic unit 1 performs the conversion processing of the sigmoid function represented by the expression (2). However, the conversion processing of the linear function and the conversion processing of the step function are disclosed. It is also possible to realize by processing.

As a method of configuring the hierarchical network unit 11 using hardware components, there is a method disclosed in Japanese Patent Application No. 63-216865 filed by the present applicant.
No. (filed on Aug. 31, 1988, "Network Configuration Data Processing Apparatus").

That is, as shown in FIG. 12, the basic unit 1
Multiplication type for multiplying the output from the preceding layer input through the input switch unit 7 and the weight value held by the weight value holding unit 8
A D / A converter 2a, an analog adder 3a that adds the output value of the multiplying D / A converter 2a and the previous accumulated value to calculate a new accumulated value, and an addition result of the analog adder 3a A sample-and-hold circuit 3b for holding, a non-linear function generating circuit 4a for non-linearly converting data held in the sample-and-hold circuit 3b when the accumulation processing is completed, and an analog of the non-linear function generating circuit 4a for outputting to a subsequent layer. This is realized by including an output holding unit 5 that holds a signal value, an output switch unit 6 that outputs data held by the output holding unit 5, and a control circuit 9 that controls these processing units. Here, regarding the subtractor 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 output to the output holding unit without passing through the nonlinear function generating circuit 4a. 5 will be adopted.

The hierarchical network unit 11 is realized by a configuration in which the basic units 1 having this configuration are electrically connected by one common analog bus 70 as shown in FIG. In the figure, reference numeral 71 denotes a weight output circuit for giving a weight value to the weight holding unit 8 of the basic unit 1, 72 denotes an initial signal output circuit corresponding to the input unit 1 ', and 73 denotes a control signal for data transfer. A synchronous control signal line 74 for transmitting a signal to the weight output circuit 71, the initial signal output circuit 72, and the control circuit 9;
Is a main control circuit for transmitting a synchronization control signal.

In the hierarchical network unit 11 having this configuration, the main control circuit 74 sequentially selects the basic units 1 in the preceding layer in a time-series manner, and synchronizes with the selection processing to output the output holding units 5 of the selected basic units 1. Is processed so as to be output to the multiplication type D / A converter 2a of the basic unit 1 in the subsequent stage via the analog bus 70 in accordance with a time-division transmission format. Upon receiving this input, the multiplying D / A converter 2a of the subsequent-stage basic unit 1 sequentially selects corresponding weight values, performs multiplication processing of the input values and the weight values, and performs analog-adder 3a and sample-and-hold operations. The accumulation processing unit 3 constituted by the circuit 3b sequentially accumulates the multiplied values. Subsequently, when all the accumulating processes for the basic unit 1 of the preceding layer are completed, the main control circuit 74 activates the nonlinear function generating circuit 4a of the basic unit 1 of the subsequent layer to calculate the final output, and The output holding unit 5 performs processing to hold the final output of this conversion processing result. And the main control circuit
Reference numeral 74 designates a process in which an output pattern corresponding to the input pattern is output by repeating the same processing for the next subsequent layer using the subsequent layer as a new preceding layer.

The hierarchical network unit 11 can be configured by software means without depending on such a configuration by hardware components. With this configuration, the truth value of the membership function for the control state quantity is accurately managed by a table means or the like, and the managed membership function is searched to execute the antecedent part truth value calculation layer 40. Since the process of calculating the truth value of the membership function with respect to the control state quantity, which is a function to perform, can be realized, it is not limited to the membership function having the function shape as described with reference to FIG. Then, as the rule calculation process of the grade value calculation layer 41, a more complicated rule calculation process can be adopted without being restricted by the calculation process of the basic unit 1.

Although the illustrated embodiment has been described, the present invention is not limited to this. For example, the applicant has previously filed “Japanese Patent Application No. 63-227825 (filed on September 12, 1988,“ Learning processing method of network configuration data processing device ”)”.
The invention has been disclosed in which the back-propagation method is improved so that the learning process of the weight value can be realized in a shorter time. However, the present invention provides such an improved back-propagation method and back-propagation method. A learning method of another weight value other than the modulus can also be used. The method of realizing the arithmetic processing of each unit constituting the hierarchical network unit 11 does not constitute the essence of the present invention, and may adopt any configuration method.

〔The invention's effect〕

As described above, according to the present invention, since a fuzzy controller is constructed according to a hierarchical network structure, a fuzzy controller that executes a desired control logic according to mechanical processing can be constructed. become. Thus, the load on the operator for tuning the fuzzy control rules can be greatly reduced.
Then, instead of using the hierarchical network structure of the known hierarchical network configuration data processing apparatus as it is, it is structured to be structured in a format compatible with the fuzzy control rules. The consistency with the container can be maintained.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram illustrating the principle of the present invention, FIG. 2 is an embodiment of the present invention, FIG. 3 is a diagram illustrating the basic structure of a basic unit, and FIG. Fig. 5 is an embodiment of the hierarchical network unit, Figs. 6 and 7 are embodiments of the truth value calculation layer of the antecedent part, Fig. 8 is an explanatory diagram of the membership function, and Fig. 9 is the consequent. FIG. 10 shows an embodiment of a centroid derived value calculation layer, FIG. 11 shows an embodiment of a basic unit, and FIG. 12 shows an embodiment of a hierarchical network unit. In the figure, 1 is a basic unit, 1 'is an input unit, 2 is a multiplication processing unit, 3 is an accumulation processing unit, 4 is a threshold processing unit, 10 is a hierarchical network configuration fuzzy controller, 11 is a hierarchical network unit, and 12 is a hierarchical network unit. Input unit, 13 is a processing unit, 14 is a rule unit, 15a is a processing unit, 15b is an output unit,
16 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, 23 is a learning convergence determination unit, 30 is a learning processing device, and 40 is a truth value calculation of the antecedent unit The layer, 41 is a grade value calculation layer, 42 is a consequent part truth value processing layer, 43 is a center of gravity derived value calculation layer, and 44 is an average value calculation layer.

──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Ryusuke Masuoka 1015 Uedanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Inside Fujitsu Limited (72) Inventor Yuri Owada 1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Fujitsu Limited ( 72) Inventor Kazuo Asakawa 1015 Uedanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Within Fujitsu Limited (72) Inventor Shigenori Matsuoka 1 Fujimachi, Hino-shi, Tokyo Fuji Facom Control Co., Ltd. (72) Inventor Hiroyuki Okada Tokyo No. 1, Fujimachi, Hino-shi, Tokyo (56) References JP-A-3-182903 (JP, A) JP-A-2-292602 (JP, A) JP-A-2-189635 (JP, A) JP-A-3-260807 (JP, A) WO 89/11684 (WO, A1) Akira Nishiguchi, "Neural Network Automatic adjustment of fuzzy membership function, "Proc. Of the 15th System Symposium, Proc. Of the Society of Instrument and Control Engineers, October 19, 1989, p. 261-266 (58) Field surveyed (Int. Cl. ⁶ , DB name) G05B 13/00-13/02 G06F 9/44 554 G06F 15/18 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A hierarchical network configuration fuzzy controller that calculates and outputs a control operation amount corresponding to an input control state amount by executing a fuzzy control rule according to a hierarchical network unit (11). A hierarchical network unit (11) receives a multiplied value of a control state quantity and an internal state value assigned to an internal connection for inputting the control state quantity, and calculates a truth value of a membership function of the antecedent part. A first calculation layer composed of a processing unit, and a multiplication of output values of all units in a preceding layer located at a stage subsequent to the first calculation layer and an internal state value allocated to an internal connection therebetween A middle layer composed of one or a plurality of rule layers composed of one or a plurality of rule units for calculating a grade value by using a value as an input, and a final rule layer The output value of the corresponding rule unit and the multiplication value of the discrete function value allocated to the inner connection therebetween and cut out from the consequent membership function are input, and the size is changed by the output value. A second calculation layer composed of a plurality of processing units for calculating a function sum of the consequent part membership function, an output value of a corresponding processing unit of the second calculation layer, and an internal connection therebetween And the multiplication value with the numerical value obtained by dividing between the specified constant values as input,
An output layer comprising an output unit, which is a set of two units for calculating a derived value of the control operation amount, a fuzzy controller configured as a hierarchical network.