Neuro-fuzzy fusion data processing system

Claim 2

A neuro-fuzzy integrated data processing system according to Claim 1, wherein said rule part is pre-wired between adjacent layers in said rule part, between said antecedent membership function realizing part and said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part where the existence of connection and a part or the whole of a weight value are set according to a fuzzy rule.

Claim 3

A neuro-fuzzy integrated data processing system according to Claim 1, wherein said rule part is completely connected between adjacent layers in said rule part, between said antecedent membership function realizing part and said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part where the existence of connection and a part or the whole of a weight value are set according to a fuzzy rule.

Claim 4

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein said antecedent membership function realizing part like said rule part, comprises, a hierarchical i.e. layered network.

Claim 5

A neuro-fuzzy integrated data processing system according to Claim 4, wherein said antecedent membership function realizing part, includes a part of a unit for outputting a grade value of the applicability of said antecedent membership function is capable of outputting a grade value of a membership function indicating "an input signal is SMALL" with a sign of a weight value of the input connection to said unit and a sign of a threshold assumed as negative.

Claim 6

A neuro-fuzzy integrated data processing system according to Claim 4, wherein said antecedent membership function realizing part includes a part of a unit for outputting a grade value of the applicability of said antecedent membership function is capable of outputting a grade value of a membership function indicating "an input signal is LARGE" with a sign of a weight value of input connection to said unit and a sign of a threshold assumed as positive.

Claim 7

A neuro-fuzzy integrated data processing system according to Claim 4, wherein said antecedent membership function realizing part includes a unit capable of outputting a grade value of a membership function indicating "an input signal is NORMAL", said unit comprising: two units for receiving said input signal and outputting a Sigmoid function of said input signal corresponding to a weight value and a threshold of each input signal; and a unit for obtaining the difference in the output of said two units and for outputting said grade value of said antecedent membership function.

Claim 8

A neuro-fuzzy integrated data system according to Claims 1, 2, or 3, wherein said rule part includes a part of a unit capable of calculating in a logical operation, in response to two input values, a logical product, a logical sum, an acuity product, an acuity sum, an average, a permanent TRUE, and an permanent FALSE, or of calculating in a logical operation in response to either of two input values.

Claim 9

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein said rule part includes a part of a unit for setting S as positive and t as between 0 and 1 in response to a plurality of input values X₁, X₂, ---, X n and for determining each of weight values W₁, W₂, ---, W n of connection and a threshold ϑ in response to a plurality of said input values as follows: where [X i ]   = +1 X i : when positive = -1 X i : when negative, and thereby performing addition of a plurality of said input values as follows:

Claim 10

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein said rule part includes a part of a unit for setting S as positive and t as between 0 and 1 in response to a plurality of input values X₁, X₂, ---, X n and for determining each of weight values W₁, W₂, ---, W n of connection and a threshold ϑ in response to a plurality of said input values as follows: where [X i ]   = +1 X i : when positive = -1 X i : when negative, and thereby performing multiplication of a plurality of said input values as follows:

Claim 11

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein said consequent membership function realizing/non-fuzzy-processing part comprises: one or a plurality of said consequent membership function realizing parts for enlarging/reducing said consequent membership function corresponding to each of one or a plurality of system output signals according to said enlargement/reduction rate of a consequent membership function provided by said rule part; and one or a plurality of non/fuzzy-process calculation realizing parts for calculating, using output values provided by one or a plurality of said consequent membership function realizing parts, a representative value as said corresponding system output signal.

Claim 12

A neuro-fuzzy integrated data processing system according to Claim 11, wherein one or a plurality of said consequent membership function realizing parts comprises a hierarchical neural network having one or a plurality of layers where connection between adjacent layers in said hierarchical neural network, and a weight value of said connection is set corresponding to a consequent membership function.

Claim 13

A neuro-fuzzy integrated data processing system according to Claim 12, wherein each of said consequent membership function realizing part comprises a plurality of units each corresponding to a plurality of abscissa values of a curve of said consequent membership function; and a grade value of said consequent membership function at each of said abscissa values is set, in said rule part, as a weight value of each connection between each unit for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions related to said corresponding system output signals and a plurality of said units.

Claim 14

A neuro-fuzzy integrated data processing system according to Claim 12, wherein each of said consequent membership function realizing parts comprises a plurality of units each corresponding to a plurality of abscissa values of a curve of said consequent membership function; and in said rule part, a value of 1 is set for the connection to a unit corresponding to an abscissa value of a consequent membership function curve specified by the consequent part of each of said fuzzy rules, and a value of 0 is set for other connection as a weight value of each connection between each unit for outputting a grade value of a fuzzy rule of one or a plurality of consequent membership functions related to said corresponding system output signals and a plurality of said units.

Claim 15

A neuro-fuzzy integrated data processing system according to Claim 14, wherein some of a plurality of units provided in said consequent membership function realizing part outputs a logical sum of a plurality of input values applied to said units.

Claim 16

A neuro-fuzzy integrated data processing system according to Claim 14, wherein some of a plurality of units provided in said consequent membership function realizing part output an algebraic sum of a plurality of input values applied to said units.

Claim 17

A neuro-fuzzy integrated data processing system according to Claim 11, wherein each of said non-fuzzy-process calculation realizing parts comprises: a center-of-gravity determining element output part comprising two units for receiving a plurality of signals as a result of the enlargement/reduction of a consequent membership function provided by said consequent membership function realizing part; and a center-of-gravity calculating part for outputting a center-of-gravity value as said system output signal, that is, a representative value, using two center-of-gravity determining elements provided by said center-of-gravity determining element output part.

Claim 18

A neuro-fuzzy-integrated data processing system according to Claim 17, wherein the first unit of said two units in said center-of-gravity determining element output part outputs ④ the sum of products obtained by multiplying the difference, that is, a weight value (a weight value = ② - ①), between the minimum value ① in the abscissa values of a consequent membership function curve corresponding to a plurality of output values as a result of the enlargement/reduction of said consequent membership function and values ② corresponding to each of a plurality of said output values by output values in a plurality of said output values; and the second unit of said two units outputs a sum of products obtained by multiplying the difference, that is, a weight value (a weight value = ② - ③), between the maximum value ③ in the abscissa values corresponding to a plurality of output values and values ② corresponding to each of a plurality of output values by output values in a plurality of said output values, thus obtaining said center-of-gravity values as follows:,

Claim 19

A neuro-fuzzy integrated data processing system according to Claim 18, wherein said first and second units output a result by multiplying each of said sums by a constant.

Claim 20

A neuro-fuzzy integrated data processing system according to Claims 17, 18, or 19, wherein a center-of-gravity determining element output unit as said center-of-gravity determining element output part comprises: an endpoint coordinate and oblique storing means for storing coordinates of two endpoints, that is, the maximum and minimum values, in a plurality of abscissa values of said consequent membership function curve and for storing the oblique of a line passing through said center of gravity; a teaching signal calculating means for obtaining and outputting teaching signals for two center-of-gravity determining elements used to obtain center-of-gravity values using a true center-of-gravity value, said endpoint values stored in said endpoint coordinate and oblique storing means, and an oblique of a line passing through the center-of-gravity; and a teaching signal determining unit for determining said teaching signals according to an expression of a line passing through a center-of-gravity having a predetermined oblique.

Claim 21

A neuro-fuzzy integrated data processing system according to Claims 17, 18, or 19, wherein a center-of-gravity determining element output unit as said center-of-gravity determining element output part comprises: an endpoint coordinate storing means for storing coordinates of two endpoints, that is, the maximum and minimum abscissa values of said consequent membership function curve; a teaching signal calculating means for obtaining and outputting teaching signals for two center-of-gravity determining elements used to obtain center-of-gravity values using true center-of-gravity values, two center-of-gravity determining element values, and endpoint coordinates stored in said endpoint coordinate storing means; and a teaching signal determining unit for determining said teaching signals according to an expression of a line passing through true center-of-gravity values having the same oblique as the line determined by output values of said two center-of-gravity determining elements.

Claim 22

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein said consequent membership function realizing/non-fuzzy-processing part comprises a hierarchical neural network having a plurality of layers.

Claim 23

A neuro-fuzzy integrated data processing system according to Claim 22, wherein said consequent membership function realizing/non-fuzzy-processing part comprises: a consequent membership function realizing part for enlarging/reducing said consequent membership function where the connections of adjacent layers are set corresponding to said consequent membership functions; and a center-of-gravity calculation realizing part as a hierarchical neural network having one or a plurality of intermediate layers for calculating a center-of-gravity value as said output signal according to the output of said consequent membership function realizing part.

Claim 24

A neuro-fuzzy integrated data processing system according to Claims 23, wherein an input normalizing unit is provided between said consequent membership function realizing part and said center-of-gravity calculation realizing part for mapping the output of said consequent membership function realizing part within the predetermined range of an input unit in said center-of-gravity calculation realizing part using an appropriate function; and an output restoring unit is provided between the consequent step of said center-of-gravity calculation realizing part for copying a center-of-gravity coordinate as the output of said center-of-gravity calculation realizing part within an appropriate range of coordinates.

Claim 25

A neuro-fuzzy integrated data processing system according to Claim 24, wherein said input normalizing unit uses the linear function for mapping according to the following expression: function value = 0.6 x (output of consequent membership function realizing part) / 10 + 0.2; and said output restoring unit uses other linear functions for mapping according to the following expression: function value = 4 x (output of center-of-gravity calculation realizing part - 0.2) / 0.6 + 1

Claim 26

A neuro-fuzzy integrated data processing system according to Claims 23, 24, or 25, wherein random numbers are used as values corresponding to the abscissa of said consequent membership function curve; teaching data are generated by mapping said random numbers by a linear function within the input range of a hierarchical neural network as said center-of-gravity calculation realizing part, and are supplied as learning data; and a center-of-gravity learning unit is provided for learning said hierarchical neural network.

Claim 27

A neuro-fuzzy integrated data processing system according to Claim 26, wherein said center-of-gravity learning unit comprises: a random number generating part for generating random numbers as values corresponding to the abscissa of said consequent membership function curve and a constant storing part for storing constants including a plurality of abscissa values of said consequent membership function curve and teaching data values; and a teaching data generating part generates, according to constants stored in said constant storing part and random numbers generated by said random number generating part, teaching data by mapping said random numbers by a linear function within the input range of a hierarchical neural network as said center-of-gravity calculation realizing part.

Claim 28

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein said consequent membership function realizing/non-fuzzy-processing part forms a division network comprising one output unit and one or more than one intermediate layers; a weight value of the connection between a unit for outputting a grade value of a fuzzy rule as an enlargement/reduction rate of one or a plurality of consequent membership functions outputted by said rule part and a first unit of said two input units is set as a value of the abscissa of a consequent membership function curve specified in the consequent part of each fuzzy rule; a weight value of the connection between a unit for outputting a grade value of said fuzzy rule and a second unit of said two input units is set to 1; and said division network outputs a result of a sum of input to said first unit divided by a sum of input to the second unit.

Claim 29

A neuro-fuzzy integrated data processing system according to Claim 28, wherein a weight value of the connection between the input of two input units in said division network can be learned in the back propagation method during the learning process of said consequent membership function realizing/non-fuzzy-processing part.

Claim 30

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the upper limit of the sensitivity determined and with a weight value W and a threshold ϑ defined as follows: W = 4 / (b - a), ϑ = 2 (a + b) / (b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = l / (1 + exp (-WX + ϑ)

Claim 31

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein approximation of a membership function obtained by a neuro-process can be realized with the upper limit of the sensitivity determined and with a weight value W and a threshold ϑ defined as follows: W = 2 / (b - a), ϑ = (a + b) / (b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = 0.5 + 0.5 tanh (WX - ϑ)

Claim 32

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the integration of an absolute value of an error minimized and with a weight value W and a threshold ϑ defined as follows: W = 5.3605/(b - a), ϑ = 2.6802(a + b)/(b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = l / (1 + exp (-WX + ϑ)

Claim 33

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the integration of an absolute value of an error minimized and with a weight value W and a threshold ϑ defined as follows: W = 2.6802/(b - a), ϑ = 1.3401(a + b)/(b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = 0.5 + 0.5 tanh (WX - ϑ)

Claim 34

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the integration of two square of an error minimized and with a weight value W and a threshold ϑ defined as follows: W = 5.3012/(b - a), ϑ = 2.6506(a + b)/(b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = l / (1 + exp (-WX + ϑ)

Claim 35

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the integration of two square of an error minimized and with a weight value W and a threshold ϑ defined as follows: W = 2.6506/(b - a), ϑ = 1.3253(a + b)/(b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = 0.5 + 0.5 tanh (WX - ϑ)

Claim 36

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the maximum error minimized and with a weight value W and a threshold ϑ defined as follows: W = 5.648/(b - a), ϑ = 2.824(a + b)/(b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = 1 / (1 + exp (-WX + ϑ))

Claim 37

A neuro-fuzzy integrated data processing system according to Claims 1, 2, or 3, wherein the approximation of a membership function obtained by a neuro-process can be realized with the maximum error minimized and with a weight value W and a threshold ϑ defined as follows: W = 2.824/(b - a), ϑ = 1.412(a + b)/(b - a) where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a, Y = (X - a) / (b - a) when a < X < b, Y = 1 when b ≦ X and the characteristic of a neuron is defined as follows: Y = 0.5 + 0.5 tanh (WX - ϑ)

Claim 38

A neuro-fuzzy integrated data processing system according to Claims 1, 2, and 3 wherein a weight value W and a threshold ϑ of one of two units in the second layer is defined as follows: W = 4/C, ϑ = 2 (2a - 2b - c)/C and a weight value W and a threshold ϑ of the other unit are defined as follows: W = -4/C, ϑ = -2 (2a + 2b + c)/C where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a - b - c, Y = (X - (a-b-c))/C when a-b-c < X ≦ a-b, Y = 1 when a-b < X ≦ a+b, Y = -(X - (a+b+c))/C when a+b < X a+b+c, Y = 0 when a+b+c < X, and a three layer hierarchical neural network for realizing the approximation of said membership function with the upper limit of the sensitivity determined as ±1/C, said network comprising: a first layer of one unit for outputting an input value as is; a second layer of two units having a non-linear characteristic; and a third layer of one neuron for outputting a result obtained by subtracting 1 from a sum of the output of two neurons in the second layer, wherein the characteristic of a unit in the second layer is: Y = 1 / (1 + exp (-WX + ϑ))

Claim 39

A neuro-fuzzy integrated data processing system according to Claims 1, 2, and 3 wherein a weight value W and a threshold ϑ of one of two units in the second layer is defined as follows: W = 2/C, ϑ = (2a - 2b - c)/C and a weight value W and a threshold ϑ of the other unit are defined as follows: W = -2/C, ϑ = - (2a + 2b + c)/C where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a - b - c, Y = (X - (a-b-c))/C when a-b-c < X ≦ a-b, Y = 1 when a-b < X ≦ a+b, Y = -(X - (a+b+c))/C when a+b < X ≦ a+b+c, Y = 0 when a+b+c < X, and a three layer hierarchical neural network for approximating said membership function with the upper limit of the sensitivity determined as ±1/C, said network comprising: a first layer of one unit for outputting an input value as is; a second layer of two units having a non-linear characteristic; and a third layer of one neuron for outputting a result obtained by subtracting 1 from a sum of the output of two neurons in the second layer, wherein the characteristic of a unit in the second layer is: Y = 0.5 + 0.5 tanh (WX - ϑ))

Claim 40

A neuro-fuzzy integrated data processing system according to Claims 1, 2, and 3 wherein a weight value W and a threshold ϑ of one of two units in the second layer is defined as follows: W = 5.468/C, ϑ = 2.824(2a - 2b - c)/C and a weight value W and a threshold ϑ of the other unit are defined as follows: W = -5.468/C, ϑ = -2.824 (2a + 2b + c)/C where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a - b - c, Y = (X - (a-b-c))/C when a-b-c < X ≦ a-b, Y = 1 when a-b < X ≦ a+b, Y = -(X - (a+b+c))/C when a+b < X ≦ a+b+c, Y = 0 when a+b+c < X, and a three layer hierarchical neural network for realizing the approximation of said membership function with the maximum error minimized, said network comprising: a first layer of one unit for outputting an input value as is; a second layer of two units having a non-linear characteristic; and a third layer of one neuron for outputting a result obtained by subtracting 1 from a sum of the output of two neurons in the second layer, wherein the characteristic of a unit in the second layer is: Y = 1 / (1 + exp (-WX + ϑ))

Claim 41

A neuro-fuzzy integrated data processing system according to Claims 1, 2, and 3 wherein a weight value W and a threshold ϑ of one of two units in the second layer is defined as follows: W = 2.824/C, ϑ = 1.412(2a - 2b - c)/C and a weight value W and a threshold ϑ of the other unit are defined as follows: W = -2.824/C, ϑ = 1.412 (2a + 2b + c)/C where said antecedent membership function is defined as having the relation between an input X and an output Y as follows: Y = 0 when X ≦ a - b - c, Y = (X - (a-b-c))/C when a-b-c < X ≦ a-b, Y = 1 when a-b < X ≦ a+b, Y = -(X - (a+b+c))/C when a+b < X ≦ a+b+c, Y = 0 when a+b+c < X, and a three layer hierarchical neural network for realizing the approximation of said membership function with the maximum error minimized, said network comprising: a first layer of one unit for outputting an input value as is; a second layer of two units having a non-linear characteristic; and a third layer of one neuron for outputting a result obtained by subtracting 1 from a sum of the output of two neurons in the second layer, wherein the characteristic of a unit in the second layer is: Y = 0.5 + 0.5 tanh (WX - ϑ))

Claim 42

A network structure conversion system in a hierarchical i.e. layered network in a neuro-fuzzy integrated data processing system for calculating and outputting according to a fuzzy control rule the amount of control operation corresponding to the inputted control state value, said system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizing part for forming a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein at least said rule part comprises a hierarchical neural network; and the network structure can be simplified by detecting and eliminating insignificant connection to the whole data process between the specified adjacent layers in said hierarchical neural network.

Claim 43

A network structure conversion system in a hierarchical network according to Claim 42, wherein said hierarchical neural network is a pure neuro at first where all adjacent layers are completely connected; and said pure neuro is converted to a completely-connected-rule-part neuro where complete connection is made only between said antecedent membership function realizing part and said rule part, between adjacent parts in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part; while insignificant connection to the whole data process is deleted between all the other parts in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part.

Claim 44

A hierarchical network structure conversion system according to Claim 42, wherein said hierarchical neural network is a pure neuro at first where all adjacent layers are completely connected; and said pure neuro is converted to a pre-wired-rule part neuro where complete connection is made between said antecedent membership function realizing part and said rule part, between adjacent parts in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part; while insignificant connection to the whole data process is deleted between all the other parts in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part.

Claim 45

A hierarchical network structure conversion system according to Claim 42, wherein said hierarchical neural network comprises a neural network at least in said rule part at first, and is a completely-connected-rule-part neuro where complete connection is made only between said antecedent membership function realizing part and said rule part, between adjacent parts in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part; and said completely-connected-rule-part neuro is converted to a pre-wired-rule-part neuro with the insignificant connection to the whole process deleted between said antecedent membership function realizing part and said rule part, between adjacent parts in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part.

Claim 46

A fuzzy model extracting system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizer for comprising a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said neuro-fuzzy integrated data processing system for calculating and outputting the control operation value corresponding to the inputted control state value according to a fuzzy control rule has a configuration in which: at least said rule part has a hierarchical neural network; said hierarchical neural network is a pre-wired-rule-part neuro; and a fuzzy rule can be extracted from a pre-wired-rule-part by analyzing capabilities of each unit of said rule part forming said pre-wired-rule-part neuro and then associating them with the capabilities of a specific logical element.

Claim 47

A fuzzy model extracting system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizer for comprising a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said neuro-fuzzy integrated data processing system for calculating and outputting the control operation value corresponding to the inputted control state value according to a fuzzy control rule has a configuration in which: said input part, antecedent membership function realizing part, rule part, and a part of at least said consequent membership function realizing/non-fuzzy-processing part have a hierarchical neural network; said hierarchical neural network is a completely-connected-rule-part neuro that is completely connected only between said antecedent membership function realizing part and said rule part, between layers of said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-realizing part, while insignificant connection is deleted between all the other parts in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy/processing part; and a membership function is extracted by analyzing said completely-connected-rule-part neuro.

Claim 48

A fuzzy model extracting system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizing part for forming a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said neuro-fuzzy integrated data processing system for calculating and outputting the control operation value corresponding to the inputted control state value according to a fuzzy control rule has a configuration in which: said input part, antecedent membership function realizing part, rule part, and a part of at least said consequent membership function realizing/non-fuzzy-processing part have a hierarchical neural network; said hierarchical neural network is a pure neuro at first where all adjacent layers are completely connected; and said pure neuro is converted to a completely-connected-rule-part neuro where complete connection is made only between said antecedent membership function realizing part and said rule part, between adjacent parts in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part; while insignificant connection to the whole data process is deleted between all the other parts in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part; and a membership function is extracted by analyzing said completely-connected-rule-part neuro.

Claim 49

A fuzzy model extracting system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizer for comprising a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said neuro-fuzzy integrated data processing system for calculating and outputting the control operation value corresponding to the inputted control state value according to a fuzzy control rule has a configuration in which: said input part, antecedent membership function realizing part, rule part, and a part of at least said consequent membership function realizing/non-fuzzy-processing part have a hierarchical neural network; said hierarchical neural network is a pure neuro at first where all adjacent layers are completely connected; and said pure neuro is converted to a pre-wired-rule-part neuro where insignificant connection to the whole data process is deleted between layers of said antecedent membership function realizing part and said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part; and a membership function is extracted by analyzing said pre-wired-rule-part neuro.

Claim 50

A fuzzy model extracting method comprising:    an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part, preceded by said antecedent membership function realizing part for forming a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said neuro-fuzzy integrated data processing system for calculating and outputting the control operation value corresponding to the inputted control state value according to a fuzzy control rule has a configuration in which: said input part, antecedent membership function realizing part, rule part, and a part of at least said consequent membership function realizing/non-fuzzy-processing part have a hierarchical neural network; said hierarchical neural network is a pure neuro at first where all adjacent layers are completely connected; and said pure neuro is converted to a pre-wired-rule-part neuro where insignificant connection to the whole data process is deleted between layers of said antecedent membership function realizing part and said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part; and a fuzzy rule is extracted by analyzing said pre-wired-rule-part neuro.

Claim 51

A fuzzy-model extracting system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizing part for forming a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said neuro-fuzzy integrated data processing system for calculating and outputting the control operation value corresponding to the inputted control state value according to a fuzzy control rule has a configuration in which: at least said rule part has a hierarchical neural network; said hierarchical neural network comprises a completely-connected-rule-part neuro where complete connection is made only between said antecedent membership function realizing part and said rule part, between adjacent layers in said rule part, and between said rule part and said consequent membership function realizing/non=fuzzy-processing part; and said completely-connected-rule-part neuro is converted to a pre-wired-rule-part neuro with the insignificant connection to the whole process deleted between said antecedent membership function realizing part and said rule part, between adjacent layers in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part, and a fuzzy rule is extracted by analyzing said pre-wired-rule-part neuro.

Claim 52

A learning system for a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following three-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said rule part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said antecedent membership function realizing part according to the learning data; and the third step to learn a total weight value of said antecedent membership function realizing part, said rule part, and said consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 53

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following three-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said rule part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said rule part according to the learning data; and the third step to learn a total weight value of said antecedent membership function realizing part, said rule part, and said consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 54

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following four-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said rule part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said rule part according to the learning data; the third step to learn a weight value of said antecedent membership function realizing part according to the learning data; and the fourth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 55

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following four=step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said rule part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said antecedent membership function realizing part according to the learning data; the third step to learn a weight value of said rule part according to the learning data; and the fourth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 56

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following three-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part and rule part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said antecedent membership function realizing part and rule part according to the learning data; and the third step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 57

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following four-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said rule part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said antecedent membership function realizing part according to the learning data; the third step to learn a weight value of said consequent membership function realizing/non-fuzzy-processing part according to the learning data; and the fourth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 58

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following four-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said rule part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said rule part according to the learning data; the third step to learn a weight value of said consequent membership function realizing/non-fuzzy-processing part according to the learning data; and the fourth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 59

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following five-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said rule part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said rule part according to the learning data; the third step to learn a weight value of said antecedent membership function realizing part according to the learning data; the fourth step to learn a weight value of the consequent membership function realizing/non-fuzzy-processing part according to the learning data; and the fifth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 60

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely connected between adjacent layers of all units, but is partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following five-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said rule part and consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said antecedent membership function realizing part according to the learning data; the third step to learn a weight value of said rule part according to the learning data; the fourth step to learn a weight value of the consequent membership function realizing/non-fuzzy-processing part according to the learning data: and the fifth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 61

A learning method of a pre-wired-rule-part neuro in a hierarchical network comprising said antecedent membership function realizing part, one or a plurality of said rule parts, and said consequent membership function realizing/non-fuzzy-processing part, wherein said rule part is not completely-connected between each of adjacent layers of all units, but partially connected internally, according to the control rules, between said antecedent membership function realizing part in the preceding step and said consequent membership function realizing/non-fuzzy-processing part in the following step, or between adjacent layers in said rule part; and outputs one or a plurality of control operation values (Y) corresponding to inputted control state values (X₁, X₂, --- X n ); and the following four-step process is performed: the first step to initialize a weight value according to the knowledge pre-stored in said antecedent membership function realizing part and rule part or according to random numbers and to initialize a weight value according to the knowledge pre-stored in said consequent membership function realizing/non-fuzzy-processing part; the second step to learn a weight value of said antecedent membership function realizing part and rule part according to the learning data; the third step to learn a weight value of said consequent membership function realizing/non-fuzzy-processing part according to the learning data; and the fourth step to learn a total weight value of said antecedent membership function realizing part, rule part, and consequent membership function realizing/non-fuzzy-processing part according to the learning data.

Claim 62

A learning method of a pre-wired-rule-part neuro according to Claims 52 to 61, wherein an initial weight value in said first step for first learning a weight value is initialized such that a weight value recognizable according to prestored knowledge is initialized by said prestored knowledge, and others are initialized according to random numbers.

Claim 63

A conversion system from a fuzzy model to a pre-wired-rule-part neuro, wherein a weight value of the connection and a threshold are set between layers in said antecedent membership function realizing part, between said antecedent membership function realizing part band said rule part, between layers in said rule part, between said rule part and said consequent membership function realizing/non-fuzzy-processing part, and between layers in said consequent membership function realizing/non-fuzzy processing part according to a antecedent membership function, a fuzzy rule and a consequent membership function described in the fuzzy presumption form of the input/output signal relation of target data to be processed; and said fuzzy model is converted to a pre-wired-rule-part neuro.

Claim 64

A conversion system from a fuzzy set to a pre-wired-rule-part neuro according to Claim 63, wherein said pre-wired-rule-part neuro is tuned by having said pre-wired-rule-part neuro learn input/output data of said fuzzy model.

Claim 65

A conversion method from a fuzzy model to a completely-connected-rule-part neuro, wherein a weight value of the connection and a threshold are set between layers in said antecedent membership function realizing part, and in said consequent membership function realizing/non-fuzzy processing part according to a antecedent membership function and a consequent membership function described in the fuzzy presumption form of the input/output signal relation of target data to be processed; and said fuzzy model is converted to a completely-connected-rule-part neuro where adjacent layers are completely connected between said antecedent membership function realizing part and said rule part, in said rule part, and between said rule part and consequent membership function realizing/non-fuzzy-processing part.

Claim 66

A conversion system from a fuzzy model to a completely-connected-rule-part neuro according to Claim 65, wherein said completely-connected-rule-part neuro is tuned by having said completely-connected-rule-part neuro learn input/output data of said fuzzy model.

Claim 67

A conversion system from a fuzzy model to a pure neuro comprising: a fuzzy model described in the fuzzy presumption form of the input/output signal relation of target data to be processed; and a pure neuro which is a hierarchical neural network where all adjacent layers are completely connected, wherein data processing capabilities of said fuzzy model are mapped in said pure neuro by having said pure neuro learn input/output data of said fuzzy model.

Claim 68

A conversion system from a pure neuro to a pre-wired-rule-part neuro comprising: a pure neuro which is a hierarchical neural network where all adjacent layers are completely connected; and a pre-wired-rule-part neuro where insignificant connection to the whole data process is deleted between layers of said antecedent membership function realizing part and said rule part, between layers in said rule part, between layers of said rule part and said consequent membership function realizing/non-fuzzy-processing part, wherein data processing capabilities of said pure neuro are mapped in said pre-wired-rule-part neuro by having said pre-wired-rule-part neuro learn input/output data of said pure neuro.

Claim 69

A conversion system from a pure neuro to a completely-connected-rule-part neuro comprising: a pure neuro which is a hierarchical neural network where all adjacent layers are completely connected; and a completely-connected-rule-part neuro where adjacent layers are completely connected only between said antecedent membership function realizing part and said rule part, in said rule part, and between said rule part and said consequent membership function realizing/non-fuzzy-processing part, while insignificant connection to the whole data process is deleted between other adjacent layers in said antecedent membership function realizing part and in said consequent membership function realizing/non-fuzzy-processing part, wherein data processing capabilities of said pure neuro are mapped in said completely-connected-rule-part neuro by having said completely-connected-rule-part neuro learn input/output data of said pure neuro.

Claim 70

A neuro-fuzzy integrated data processing system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting one or a plurality of grade values of a antecedent membership function; a rule part preceded by said antecedent membership function realizing part for forming a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting an enlargement/reduction rate of one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said enlargement/reduction rate of a consequent membership function, calculating a non-fuzzy process after enlarging/reducing a consequent membership function according to said enlargement/reduction rate, and outputting a system output signal, wherein said rule part comprises neurons.

Claim 71

A neuro-fuzzy integrated data processing system comprising: an input part for receiving one or a plurality of input signals; a antecedent membership function realizing part preceded by said input part for outputting, in response to one or a plurality of said input signals, one or a plurality of grade values indicating the applicability of a antecedent membership function; a rule part preceded by said antecedent membership function realizing part for forming a network having one or a plurality of layers, receiving a grade value of said membership function, and for outputting transformation information for transforming one or a plurality of consequent membership functions corresponding to one or a plurality of system output signals as a grade value of a fuzzy rule; and a consequent membership function realizing/non-fuzzy-processing part preceded by said rule part for receiving said transformation information of a consequent membership function, calculating a non-fuzzy process after transforming a consequent membership function according to said transformation information, and outputting a system output signal.