Method of configuring a neural network and a diagnosis/recognition system using the same

• US 5,598,509 A
• Filed: 03/18/1994
• Issued: 01/28/1997
• Est. Priority Date: 08/28/1992
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Takahashi, Isao, Endo, Fumihiro, Yamagiwa, Tokio
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): KATBAB, ABDOLLAH
• Application Number:
  US08/214,769
• Time in Patent Office: 1,047 Days
• Field of Search: 395/22, 395/23, 382/14, 382/157, 382/159
• US Class Current: 706/20
• CPC Class Codes: G06F 18/00 Pattern recognition
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: Y10S 128/925 Neural network

Abstract

In a recognition/diagnosis method, a relation between total sums of inputs of respective cases and teacher data is listed in the order of magnitude of the total sums of inputs. Based on the value of the teacher data for the input having the maximum total sum and the number of times of change of teacher data in the table are considered, a configuration of a neural network (the number of hidden layers and the number of neurons thereof) is determined. Coupling weights are analytically calculated based on the table.

Claims

1. A method for analytically determining coupling weights using a neural network having an input layer, at least a hidden layer and an output layer, said method comprising the steps of:

classifying learning data input to said input layer based on features thereof into classifications of learning data;

inputting learning data for each of said classifications of learning data;

calculating a total sum for each of said classifications of learning data;

correlating teacher data corresponding to the calculated total sum for each of said classifications of learning data in the order of magnitude of the calculated total sum for each output of said output layer; and

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums of said classifications of learning data before and after any change of the teacher data, and the teacher data; and

performing recognition or diagnosis by using said neural network having the determined coupling weights.

4. A method according to claim 1 wherein two hidden layers are provided and the coupling weights are selected such that when the teacher data corresponding to the classification of learning data having the maximum total sum is 0, the second hidden layer functions as an AND circuit and said output layer functions as an OR circuit.

5. A method according to claim 1, wherein two hidden layers are provided and the coupling weights are selected such that when the teacher data corresponding to the classification of learning data having the maximum total sum is 1, the second hidden layer functions as an OR circuit and said output layer functions as an AND circuit.

6. A method according to claim 1, wherein said teacher data are used by reciprocals thereof for each classification of learning data and a reversal operation in the magnitude of the input data is executed.

7. A method according to claim 1 wherein a back propagation iterative learning method is used to calculate the coupling weights which are principal parts of the neural network.

2. A method for analytically determining coupling weights using a neural network having an input layer, at least a hidden layer and an output layer, said method comprising the steps of:

inputting input data for each value for each classification of learning data based on an upper limit and a lower limit thereof to said input layer;

calculating a total sum for each of said classifications of learning data;

correlating teacher data corresponding to the calculated total sum for each of said classifications of learning data in the order of magnitude of the calculated total sum for each output of said output layer; and

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums of said classifications of learning data before and after any change of the teacher data, and the teacher data; and

performing recognition or diagnosis by using said neural network having the determined coupling weights.

3. A method according to claim 2, wherein the input data input to said input layer are products of the upper limit and the lower limit multiplied by a same coupling weight.

8. A method for analytically determining coupling weights using a neural network having an input layer, at least a hidden layer and an output layer, said method comprising the steps of:

selecting coupling weights such that an input sum determined for a classification of learning data other than a particular classification of learning data input to said input layer does not fall between two input sums determined based on an upper limit and a lower limit of input data of said particular classification of learning data, and determining the input sum;

inputting said input sum as input data;

calculating an input sum for each classification of learning data;

calculating a total sum for each classification of learning data;

correlating teacher data corresponding to the input sum for each classification of learning data in the order of magnitude of the calculated total sum for each output of said output layer; and

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum total sum, total sums of classifications of learning data before and after any change of the teacher data, and the teacher data; and

performing recognition or diagnosis by using said neural network having the determined coupling weights.

9. A method according to claim 8, wherein the teacher data of said particular classification of learning data and the teacher data of a classification of learning data other than said particular classification of learning data are different from each other.

10. A recognition or diagnosis method using a neural network having an input layer, at least a hidden layer and an output layer, said method comprising the steps of:

inputting learning data to said input layer for each classification of learning data;

calculating a total sum for each classification of learning data;

correlating teacher data to the calculated total sum for each classification of learning data in the order of magnitude of the calculated total sum for each output of said output layer;

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums of said classifications of learning data before and after the change of the teacher data, and the teacher data; and

indicating the deviation from a specified range of scatter of the learning data when input diagnosis data is out of the specified range; and

performing recognition or diagnosis by using said neural network having the determined coupling weights.

11. A recognition or diagnosis method using a neural network having an input layer, at least a hidden layer and an output layer said method comprising the steps of:

inputting learning data to said input layer for each classification of learning data;

calculating a total sum for each classification of learning data;

correlating teacher data corresponding to the calculated total sum for each classification of learning data in the order of magnitude of the calculated total sum for each output of said output layer;

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums before and after the change of the teacher data and the teacher data; and

transferring diagnosis data and a corresponding correct solution as a portion of the learning data after the diagnosis when the correct solution to diagnosis data is known; and

performing recognition or diagnosis by using said neural network having the determined coupling weights.

12. A recognition or diagnosis method using a neural network having an input layer, at least a hidden layer and an output layer, said method comprising the steps of:

inputting learning data to said input layer for each classification of learning data;

calculating a total sum for each classification of learning data;

correlating teacher data corresponding to the calculated total sum for each classification of learning data in the order of magnitude of the calculated total sum for each output of said output layer;

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums before and after the change of the teacher data, and the teacher data; and

calculating an output for diagnosis data when the diagnosis data is input;

inferring at least one probable learning data from the output corresponding to the diagnosis data;

comparing the diagnosis data with the learning data; and

indicating the result of said comparing step.

13. A recognition or diagnosis method according to claim 12, wherein said inferred probable learning data includes a scatter, and the diagnosis data is compared with the learning data having the scatter and the result thereof is indicated.

14. A recognition or diagnosis method according to claim 12 wherein a first coupling weight to be multiplied to the input data of said input layer is selected to be different from other coupling weights.

15. A recognition or diagnosis method, comprising the steps of:

classifying learning data input to an input layer based on characteristics thereof;

inputting learning data for each classification of learning data;

calculating a total sum for each classification of learning data;

correlating teacher data corresponding to the calculated total sum for each classification of learning data in order of magnitude of the calculated total sum for each output of said input layer;

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums before and after a change of the teacher data, and the teach data;

comparing diagnosis data directly with learning data;

determining whether the diagnosis data and the learning data of a particular classification of learning data totally match, partially match or totally mismatch; and

indicating the result if said determining step.

16. A diagnosis evolution method, comprising the steps of:

comparing diagnosis data directly with learning data having a range of scatter;

determining whether the diagnosis data is totally within said range, partially within said range or totally out of said range; and

indicating the result of said determining step.

17. A recognition or diagnosis method, comprising the steps of:

classifying learning data input to an input layer based on characteristics thereof;

inputting learning data for each classification of learning data;

calculating a total sum for each classification of learning data;

correlating teacher data corresponding to the calculated total sum for each classification of learning data in order of magnitude of the calculated total sum for each output of said input layer;

determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums before and after a change of the teacher data, and the teach data;

preparing simultaneous equations by using calculation formulas of inputs to a neuron based on an upper limit and a lower limit of input data having a range of scatter;

solving the simultaneous equations to determine coupling weights; and

performing recognition or diagnosis by using said neural network having the determined coupling weights.

18. An apparatus for diagnosing an electrical circuit to be used under a plurality of conditions by using a neural network comprising;

means for utilizing the plurality of conditions as learning data to determine coupling weights which form said neural network;

means for classifying learning data input to an input layer based on characteristics thereof;

means for inputting learning data for each classification of learning data;

means for calculating a total sum for each classification of learning data;

means for correlating teacher data corresponding to the calculated total sum for each classification of learning data in order of magnitude of the calculated total sum for each output of said input layer; and

means for determining coupling weights based on the teacher data corresponding to the classification of learning data having a maximum calculated total sum, total sums before and after a change of the teacher data, and the teach data; and

means for performing recognition or diagnosis by using said neural network having the determined coupling weights.