JPH04199302A - Neural network learning method in fuzzy control - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to the control of plants and large equipment or equipment, and more specifically to information processing using a neural network that constitutes a neural network. .

[Prior Art] Today, various types of information processing based on fuzzy reasoning called expert systems have come to be performed.

Using this fuzzy inference, for example, in order to perform process control, the state of the equipment to be controlled, that is, the deviation of the current control amount from the target value, the amount of change in that deviation, and the degree of conformance are set as fuzzy sets as antecedent members. Define the ship function,
In addition, the amount of manipulation is determined as a consequent membership function, the observed value (a foot value) is associated with the antecedent membership function to find the great, and fuzzy inference is performed to calculate the consequent centroid. However, the control is performed so that the operating state of the equipment to be controlled approaches the ideal state.

In order to perform control and judgment based on this fuzzy inference even more reliably, it is necessary to improve the responsiveness of control by the inference device (
For example, JP-A-63-113.733).

Removal of noise included in measured values (for example, Japanese Patent Application Laid-Open No. 1-11
．． 9.803) is being attempted.

On the other hand, in place of the conventional von Neumann computer, neural computers that form a neural network using the human brain as a model have come to be developed and used.

In this neural computer, layers are formed by a large number of elements called processing elements (hereinafter referred to as PE elements) as shown in FIG. 14, and an input layer and an output layer are formed. This method uses a neural network in which a hidden layer is formed between the input layer and the output layer. In some cases, the PE element has multiple layers, and each PE element receives each input from other PE elements multiplied by a weighting coefficient, and each PE element becomes excited when the sum of the inputs exceeds a predetermined threshold value k. It is configured to output a signal rlJ.

Note that a PE element that is in an excited state and outputs "1" is called a firing element, and a state in which it outputs a signal "1" is also called a firing state. In addition, the input to each PE element is
There are cases of excitatory signals that make the E element excited, and cases of inhibitory signals that make it rest (output "0").

This neural network generates a desired unique output by giving a weighting coefficient to the signal input to each PE element. Obtaining the weighting coefficient of each connection is called learning, and comparison is performed. In order to perform learning with a small amount of data, various methods such as the hack propagation method and the steepest descent method are employed (for example, Japanese Patent Application Laid-Open No. 1-282.678).

[Problem to be solved by the invention] Control using an expert system that employs fuzzy reasoning facilitates the creation of control programs for large-scale equipment such as plants, and facilitates control and judgment when the number of detection elements is extremely large. Although it may be possible to make it usable, there may be slight control errors or judgment errors due to the inherent characteristics of the equipment to be controlled or the characteristics of the detection device, etc., and it is necessary to correct such control errors. It was still difficult to do.

[Means for Solving the Problems] The present invention uses a neural computer to perform fuzzy inference calculations to control various devices and make information judgments, and each time the control or judgment is repeated, the fuzzy inference judgment criteria are determined. The neural network is trained by changing the coupling coefficients of the PE elements so as to change the fuzzy set.

[Operation] Since the present invention performs control and judgment based on fuzzy inference, advanced control and judgment are enabled, and the learning machine part of the neural network is used to change the fuzzy set that is the basis of inference. By repeating the process and judgment, it is possible to easily realize more accurate control and judgment suitable for the device in question.

[Embodiment] An embodiment of the present invention uses a neural network that has a multi-A hidden layer between an output layer and an input layer and has no coupling between PE elements constituting each layer. It is a neural network that allows inference to be made, and in this neural network,
The PE elements of the specific layer forming the hidden layer are arranged to form a phase plane in which the horizontal axis is the deviation from the target value and the vertical axis is the differential value of the deviation, as shown in Figure 2. Allocate to.

In other words, this phase plane is 1. For example, when controlling the temperature of a kiln in a cement factory, the center of the phase plane is the origin, the horizontal axis corresponds to the deviation between the measured value and the target value, and the vertical axis represents the amount of change of the deviation. By making the differential values correspond and further forming a block for each of a plurality of PE elements in this phase plane, the block is used as a discrimination region and a correction unit of a membership function to be described later.

In this neural network, the relationship between input and output is based on fuzzy calculations such as the minimax synthesis method, and as shown in Figure 3, if the measured value Y is much larger than the target value, Membership function that is larger than the target value (deviation or positive maximum: LM)
, Membership function that is slightly larger than the target value (deviation or positive: LS), Membership function that is appropriate (deviation is O:MM), Membership function that is slightly smaller than the target value ( Deviation or negative.

SS), a membership function that is smaller than the target value (large when the deviation is negative: SM), and a membership function that is much smaller than the target value (maximum when the deviation is negative:
Similarly, for the differential value ΔY of the deviation, positive is maximum (LB), positive is large (LM), positive is small (LS), 0 (MM), negative is small ( SS), large negative (
SM), negative maximum (SB), and a fuzzy set membership function that determines the amount of fuel operation as shown in Figure 4. The amount of change is determined as a membership function, such as ZE, which maintains the current supply amount, PS, which slightly increases it, PP, which maximizes the amount of increase, NS, which slightly decreases it, and NB, which greatly decreases it.

Furthermore, the deviation or differential value of the above-mentioned measured value is used as the antecedent part, and the fuel operation amount is used as the consequent part, so that if the measured temperature, which is the antecedent part, is higher than the target temperature, the supplied fuel is reduced. , a consequent fuzzy set is assigned that increases the decrease amount, that is, the manipulated variable, as the measured temperature is higher, and increases the amount of fuel supplied when the measured temperature is lower than the target temperature, and increases the amount of increase as the measured temperature decreases. In addition, if the measured temperature is smaller than the target temperature but the measured temperature is rising, the manipulated variable is set to 0, and even if the measured temperature is slightly larger than the target temperature, the measured temperature is falling. In this case, the manipulated variable is set to 0, and even if the measured temperature is equal to the target temperature, when the temperature rises, the supply amount is decreased, and when the temperature falls, the supply amount is increased, If the measured temperature is high and rising, maximize the operation amount to reduce the supply amount.
When the measured temperature is low and falling, the operation amount to increase the supply amount is set to be maximum, and the antecedent membership function and the consequent membership function are related, for example, as shown in Fig. 5. Set chiful.

Then, according to the table, for example, if the conditional statement is IF Y
The coupling coefficient of each PE element is determined so as to output the conclusion of the fuzzy operation such as =SS, ΔY=LM THEN D=NS, etc.

As shown in FIG. 6, the content of this fuzzy operation is to determine the degree of fitness of the membership function based on the value of the measured value y, and to determine whether the degree of fitness α11 for the membership function SM is 0.
7, the fitness α12 for the membership function SS is 0.
．． 3, and the goodness of fit of the membership function is similarly determined by the value of the differential value Δy, and the goodness of fit for LS α21
is 0.9 and the fitness for LM is α22 or 0.1 (see Figure 7). Based on the table above, IF Y2SM, ΔY=LM THEN D=ZE IF Y2SS, ΔY=LM THEN D='N5 IF Y=SM, ΔY=LS THEN D=PS IF Y= SS, ΔY=LSTHEN
Based on each conditional statement of D=ZE, first the fitness of Y=SM α11-0.
7 and the fitness α22=0.1 of ΔY=LM, and the membership function ZE is calculated using the value of fitness α22.
is copied from the master table, the copied membership function ZE is α-cut as shown in the 8th UAA, and the fitness of Y=SS α12=O,:l and ΔY are similarly calculated.
= LM's fitness α22 = 0.1 and the membership function NS copied from the master table is α-cut as shown in FIG. α-cut the membership function PS copied from the master table using the fitness value obtained by mini-synthesizing the fitness degree α11=0.7 and the fitness degree α21=0.9 of ΔY=LS (see Figure 8C), y α-cut the membership function ZE copied from the master table using the mini-composite fitness value of = ss fitness α]2 and ΔY = LS fitness α21 (see diagram 8)
do. After that, each of the α-cut consequents is max-synthesized as shown in Figure 9, the center of gravity ω of this composite set is calculated to obtain the manipulated variable, and this manipulated variable is output. .

In this embodiment, as described above, the neural network is used to output the manipulated variable based on fuzzy inference, and the history of the temperature change of the kiln is known from the firing state of the PE element in the above-mentioned phase plane, and the consequent member The ship function is corrected, and this correction is performed in a region where the measured value is larger than the target value and rises in the phase plane shown in Figure 2, so the deviation is positive and the differential value of the deviation is also positive. Then, we will concentrate the consequent membership function copied from the master table according to each discriminant region, and if the deviation is negative, and if the differential value of the deviation is also negative, then we will expand the consequent membership function copied from the master table. In regions where the differential value is negative and the differential value is positive, and in regions where the differential value is negative and the deviation is positive, the brightness and darkness of the copied consequent membership function is enhanced.

This concentration is achieved by setting up a table that records the number of firings of each PE element that forms a phase plane based on the deviation of the measured temperature and the differential value of the deviation, and within a period of n cycles within a discrimination area obtained by dividing the phase plane. The number of firings of a specific PE element in a
Then, when other PE elements have the maximum number of firings in the discrimination region, the coefficient α is determined as α = b / a, and the curve of the output fuzzy set is made steep by the CON operation, and the other , the expansion is done by the coefficient β−a/
b, a DIL operation is performed using the coefficient β to make the gradient gentler, and for contrast enhancement, an INT operation is performed using the coefficient α to widen the width of the high grade and narrow the width of the low grade.

Note that if the specific PE element has the maximum number of firings in the relevant discrimination region, it is α-β-1, and the output fuzzy set is not changed.

Moreover, the above CON operation is an operation that lowers the grade of the fuzzy set A and makes the curve steeper, as shown in FIG. 10, and CON (A) -A2. As shown in FIG. 11, the DIL operation is an operation that increases the grade of the fuzzy set A and makes the curve gentler, and D I L (A)
=A05. As shown in Figure 12, INT@ calculation further increases the high grades of fuzzy set A and further lowers the low grades, and when 0≦1LA(y)≦0.5, I NT (A) −2A2,0.5≦River A (y
)≦1, the calculation INT(A)=12 (mA) 2 is performed.

In this way, the area where the deviation and the differential value of the deviation is positive is suppressed by the CON operation which is concentration, and the area where the deviation and the differential value of the deviation is positive, that is, the measured temperature is higher than the target temperature and the temperature is In the case of an increase, correction is made to suppress it, and the deviation and the differential value of the deviation or the negative area are enlarged DI.
Excitation is performed by the L calculation, and it can be corrected to increase if the deviation and the differential value of the deviation are negative, that is, if the measured temperature is lower than the target temperature and the temperature is dropping.

Note that the discrimination area is not limited to the case where a large number of discrimination areas are provided so as to divide the entire surface of the phase plane as shown in FIG. In some cases, a coefficient α or a coefficient β is calculated as a discrimination area.

Furthermore, calculations based on fuzzy inference are not limited to α-cuts, and may include multiplication of grade α (degree of fitness) values, and calculations of numerical values from composite sets are not limited to the centroid method. The area method, entropy method, etc. can also be used.

In this way, this embodiment calculates the number of firings of a specific PE element within a predetermined period, and if the other PE element in the discrimination area including the specific PE element has the maximum number of firings, the maximum number of firings in the discrimination area. Modify each output fuzzy set copied from the master table for each discrimination area based on the ratio with the number of firings,
The device to be controlled by repeating the control performs learning to change the coupling coefficient with the excitatory element group and the inhibitory element group so as to perform calculations based on the modified fuzzy set and output an output signal. control according to the characteristics of
In addition, the control amount can be modified to take into account the characteristics of the control device and the detection device.

[Effects of the Invention] The present invention allows a neural network to perform control based on fuzzy operations, and also causes the neural network to learn to perform control by modifying fuzzy sets based on a history of repeated control. Therefore, not only the characteristics of the controlled device but also the characteristics of the detection device are taken into account, learning is done to eliminate control errors and judgment errors, and proper control and f4r! It can lead to I+.

[Brief explanation of the drawing]

FIG. 1 is a flowchart showing an embodiment of the present invention, FIG. 2 is a diagram showing a phase plane, FIG. 3 is a diagram showing an example of an antecedent membership function, and FIG. 4 is a diagram showing an example of a consequent membership function. FIG. 5 is a diagram showing an example of a conditional expression between an antecedent membership function and a consequent membership function, and FIGS. 6 to 9 are diagrams showing fuzzy FIGS. 10 to 12 are diagrams showing examples of correction calculations, FIG. 13 is a diagram showing another embodiment, and 14th VA is a schematic diagram of a PE element,
Figure 515 is a simplified diagram of a neural network. Patent applicant Ube Industries Co., Ltd. 1: Representative patent attorney Hitoshi Kitamura...! Figure 2 71Y -ΔY Figure 5 Measured value (Y) SB SM SS MM LS LM L
Bpp Positive maximum PB, positive large PM, positive medium ps Positive small ZE/Zero NS: Negative small NM: Negative medium NB: Negative large NN: Negative maximum Fang 8 ■ A Engraving of drawings (contents) (no change)? ↑Figure 2? 13 m ΔY -8Y O T4 ■ 151 m Procedural amendment (method) % formula % Engineering Display of the case Patent application No. 2-331273 2, Title of invention Learning method of neural network in fuzzy control 3 Person making the correction Relationship to the case Applicant: 1-12-32 Nishikicho, Ube City, Yamaguchi Prefecture (020) Ube Industries Co., Ltd. 4, Agent: Kinoshita Building Heisei, 13-3 Hisamatsucho, Nihonbashi, Chuo-ku, Tokyo
February 25, 1991 (shipped on March 12, 1991) 6. Subject of amendment (1) Drawing 7 attached to the specification, contents of amendment (1) Figures 1 and 12 of the drawing attached to the application The figure is corrected as shown in the attached sheet (no change in content).

Claims (1)

[Claims] Each layer consists of a plurality of PE elements (processing elements), includes a plurality of hidden layers formed between an input layer and an output layer, and has mutual coupling between PE elements of adjacent layers, and When performing control using a neural network that has no connections in each layer, a phase plane that has an origin at the center and whose axes are the deviation of the measured value from the target value and the differential value of this deviation is defined as By assuming a hidden layer and giving an initial value to the weighting coefficient of each connection, the output layer outputs a result by fuzzy inference according to the membership function based on the deviation and the differential value of the deviation, and further performs measurement. A learning method for a neural network in fuzzy control, characterized in that the weighting coefficients are changed based on the trajectory of values in the phase plane to centralize, expand, and strengthen the brightness of a fuzzy set.