JP3223398B2 - Knowledge information processing apparatus and learning method thereof - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

TECHNICAL FIELD The present invention relates to a knowledge information processing apparatus and a learning method thereof.

[0002]

2. Description of the Related Art As a learning method of a knowledge information processing apparatus, knowledge information is input and output data up to or at the present time and teacher output data (target value) desired for input data is used as knowledge information. Some devices sequentially modify knowledge data, processing parameters, and the like used in a processing device.
Further, there is a device in which the processing status at the current time in the knowledge information processing device is evaluated by another device using rule-based inference or the like, and the knowledge data of the knowledge information processing device is changed by this device.

In a system provided with such a knowledge information processing device as a controller, for example, in a multi-product production system that handles products of different types, when the type is switched, preset default knowledge data is added to new knowledge data. The information is set in the knowledge information processing apparatus as initial knowledge data on the product. Further, the knowledge data on the product before the product is switched is set as it is as initial knowledge data on a new product.

[0004] However, in such a conventional knowledge information processing apparatus, when the default knowledge data is set as the initial knowledge data, it takes time to learn until the production system reaches practical performance. Also, when the knowledge data before the type is switched is set as the initial knowledge data as it is, the optimal knowledge data is specialized for that type, so if the knowledge data of another type is used as it is, the default knowledge data is set. It takes more learning time to reach practical performance than in the case. In this case, the production system may not have practical performance. This is because the knowledge information processing procedure cannot select optimal knowledge data for a new situation.

Further, when learning is performed for the same varieties, the learning oscillates and does not converge, or the learning falls to a local optimum solution (local minimum) and cannot escape. The system performance may not be sufficient. This is because the knowledge information processing device considers only the latest learning result and does not consider the learning process up to that time.

[0006] As described above, the conventional knowledge information processing apparatus includes:
He does not know what kind of learning process has been performed in various past situations and what is the current situation. Many knowledge information processing devices that store past input / output data have been proposed, but all of these devices are "what did the object do in various past situations? ”, But does not consider the internal state of the knowledge information processing apparatus, that is, the processing parameters of the knowledge data.

[0007]

DISCLOSURE OF THE INVENTION An object of the present invention is to provide a knowledge information processing apparatus and a learning method thereof.

A knowledge information processing apparatus according to the present invention has a knowledge data base for storing knowledge data for controlling or processing an object, and the knowledge data base for input data.
Information processing means for performing information processing in accordance with base knowledge data and outputting output data; learning about processing parameters of the knowledge data based on input and output data of the information processing means; Learning history means for correcting the input and output data of the information processing means and processing parameters of the knowledge data as history data in the order in which they appear, and stored in the history memory.
Current characteristic analysis means for analyzing the historical data within the current characteristic analysis reference time determined retroactively from the current time to generate current characteristic data representing the characteristic of the target at the current time, stored in the historical memory A history characteristic analysis unit configured to analyze history data within a reference period of history characteristic analysis determined retroactively from an arbitrary time in the past to generate history characteristic data representing characteristics relating to the target at the past time; When the history characteristic data generated by the characteristic analysis means is within a predetermined range, the history characteristic memory for storing the history characteristic data, and the current characteristic data for all the history characteristic data stored in the history characteristic memory. History comparison means for calculating a history similarity determined based on the difference between the history comparison data and the history characteristic data, Situation recognizing means for selecting the largest one of the obtained history similarities as the maximum similarity, and history characteristic data giving the maximum similarity selected by the situation recognizing means among the processing parameters stored in the history memory. Is provided with a knowledge changing means for changing the processing parameter of the knowledge data with the processing parameter of the knowledge data on which is generated.

A learning method for a knowledge information processing apparatus according to the present invention provides a knowledge data base for storing knowledge data for controlling or processing an object and information processing for input data in accordance with the knowledge data of the knowledge data base. And a learning method for a knowledge information processing apparatus having information processing means for outputting and outputting output data. The learning method relates to processing parameters of the knowledge data based on input and output data of the information processing means. The processing parameters are corrected, and the input / output data of the information processing means and the processing parameters of the knowledge data are stored as history data in a history memory in the order in which they appear, and the current time stored in the history memory is stored. Analyze the historical data within the current characteristic analysis reference time determined retroactively from Generate the current characteristic data representing the characteristics of the target at the current time point, and analyze the history data stored in the history memory and within the historical characteristic analysis reference time determined retroactively from an arbitrary time in the past, Generating history characteristic data representing characteristics relating to the object at the past point in time, and when the history characteristic data is within a predetermined range, stores the history characteristic data in the history characteristic memory and stores the history characteristic data in the history characteristic memory. For all of the historical characteristic data, a historical similarity determined based on the difference between the current characteristic data and the historical characteristic data is calculated, the one with the greatest historical similarity is selected as the maximum similarity, and the history similarity is stored in the history memory. The processing parameter of the stored knowledge data, which is the basis for generating the history characteristic data giving the maximum similarity, It is to change the physical parameters.

According to the present invention, the processing parameters of input / output data and knowledge data are stored as history data in the history memory in the order in which they appear. The current characteristic data is generated by analyzing the history data within the current characteristic analysis reference time set retroactively from the current time. Historical characteristic data is generated by analyzing historical data within a historical characteristic analysis reference time determined retroactively from an arbitrary time in the past, and those having a predetermined range of the historical characteristic data are stored in a historical characteristic memory. . The history similarity between the current characteristic data and the current characteristic data is calculated for all the historical characteristic data stored in the history characteristic memory. The knowledge data is changed by the processing parameter of the knowledge data from which the history characteristic data having the maximum similarity is generated.

Therefore, a processing situation at a past time similar to the processing situation at the current time can be recognized based on the history similarity between the history characteristic data at an arbitrary time in the past and the current characteristic data at the current time. . With this, the processing parameter of the knowledge data at the current time can be changed using the processing parameter of the knowledge data that has been sufficiently learned at the past time that is the most similar to the current time. Can be shortened.

Further, even in the case where the learning of the processing parameter at the present time oscillates, the transition of the learning regarding the processing parameter of the knowledge data can be seen by the history characteristic data, and the learning oscillates. State,
It is possible to recognize, for example, a state of falling into a local optimum solution (local minimum). Therefore, the most suitable knowledge data can be obtained.

In one embodiment of the present invention, when there are a plurality of objects, knowledge data is prepared for each of the objects, and the knowledge data for the specified object is used for information processing. When history data relating to a plurality of objects is stored in the history memory, the maximum similarity is obtained for each object, and the maximum similarity that maximizes the history similarity of each object, and the history that gives this maximum similarity The processing parameter of the knowledge data is changed based on the processing parameter of the knowledge data on which the characteristic data is generated (for example, a weighted average value based on the maximum similarity of the processing parameters).

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS When the knowledge information processing apparatus according to the present invention controls a target, the knowledge data is a control rule or an identification model of the target, and when processing the target, recognition such as character recognition is performed. If so, the knowledge data is a standard pattern, a processing routine, and the like.

FIG. 1 is a diagram showing an overall configuration of a control system in which a knowledge information processing apparatus is applied to control of a winder.

The winding machine winds the product 2 wound on the unwinding reel 3 on the winding reel 5 with a constant tension. This winding machine targets a plurality of types of products. The take-up reel 3 includes a take-up motor 11
The take-up reel 5 is rotated by being driven by the take-up motor 15. The product 2 is tensioned by a step roller 3 on which a weight 7 is suspended via a spring 6. By controlling the position of the step roller 3 to be constant, a constant tension can be applied to the product 2.

For controlling the position of the step roller 4, the following data is detected.

The step roller position is the position of the step roller 4 and is detected by the step roller position sensor 13.

The unwinding peripheral speed is a speed at which the product 2 is unwound from the unwinding reel 3, and is a unwinding peripheral speed sensor 12
Is detected by

The take-up peripheral speed is the speed at which the product 2 is taken up from the take-up reel 5, and the take-up peripheral speed sensor 14
Is detected by

The unwinding motor rotation speed is the rotation speed of the unwinding motor 11, and is detected by a rotation speed sensor (not shown).

The winding motor rotation speed is the rotation speed of the winding motor 15 and is detected by a rotation speed sensor (not shown).

The pre-processing device 19 performs pre-processing on the data detected by the above-mentioned sensors.

The pre-processing device 19 determines that the step roller position deviation, the time during which the absolute value of the step roller position deviation continues within the stability judgment threshold value P0, and the absolute value of the step roller position deviation varies within the step roller position change reference time T0. The number of times exceeding the determination threshold value P1, the unwinding diameter, the winding diameter, and the difference between the unwinding peripheral speed and the winding peripheral speed (hereinafter, these are referred to as “input data”) are calculated.

FIG. 2 shows an example of time-series data of the position of the step roller.

The step roller position deviation is an error between the step roller position detected by the step roller position sensor 12 and the target position (set position) P of the step roller 4.

The time during which the absolute value of the level difference roller position deviation continues within the stability judgment threshold value P0 is the time when the detected level difference roller position is continuously within the range from the position (P-P0) to the position (P + P0). It is time to be.

For example, the time during which the absolute value of the step roller position deviation continues within the stability determination threshold value P0 is determined by Tc shown in FIG.
become.

The number of times that the absolute value of the step roller position deviation exceeds the variation determination threshold value P1 within the step roller position change reference time T0 is determined by the step roller position detected by the step roller position sensor 13 and the step roller position change reference time T0. Within (P-P1) or (P + P
This is the number of times greater than the position of 1).

For example, in the case shown in FIG. 2, the absolute value of the step roller position deviation is equal to the step roller position change reference time T0.
The number of times exceeding the variation determination threshold value P1 is “3”.

The unwinding diameter is such that the product 2 is the unwinding reel 3
Is calculated based on the detected unwinding motor rotation speed and unwinding peripheral speed.

The winding diameter is such that the product 2 is the winding reel 3
Is calculated based on the detected rotation speed of the winding motor and the peripheral speed of the winding.

The difference between the unwinding peripheral speed and the winding peripheral speed is the deviation of the winding peripheral speed from the detected unwinding peripheral speed,
That is, (unwinding peripheral speed)-(winding peripheral speed).

The preprocessed input data is provided from the preprocessing device 19 to the knowledge information processing device 20.

The knowledge information processing apparatus 20, which will be described later in detail, is provided with a preprocessing apparatus in accordance with input target information.
Information processing is performed on the input data provided from step 19, and a target unwinding peripheral speed value and a winding motor rotational speed command value (hereinafter, these are referred to as “output data”) are output. Here, the target information is information on the product 2 input by the operator, for example, the product 2 at the current time.
Varieties.

The unwinding peripheral speed target value is stored in the knowledge information processing device 20.
Are supplied to the servo device 18, and the take-up motor rotation speed command value is provided to the unwind motor driver 17.

The servo device 18 includes an unwinding peripheral speed target value provided from the knowledge information processing device 20 and an unwinding peripheral speed sensor.
An unwinding motor rotation speed command value is calculated based on the unwinding peripheral speed detected by (12). The unwinding motor rotation speed command value is given to the unwinding motor driver 16.

The unwinding motor driver 16 drives the unwinding motor 11 based on the unwinding motor rotation speed command value provided from the servo device 18.

The take-up motor driver 17 drives the take-up motor 15 based on the take-up motor rotation speed command value given from the knowledge information processing device 20.

As described above, the position of the step roller 4 is controlled so as to be at the set position P by operating the unwinding motor rotation speed and the winding motor rotation speed. Thereby, the tension applied to the product 2 is set to the set value.

The winding machine preferably sets the tension applied to the product 2 to a set value to increase the speed of winding the product 2 from the unwinding reel 3 to the winding reel 5.

However, the unwinding diameter and the rewinding diameter change every moment, the type of the product 2 is changed, and the dynamic characteristics of the unwinding motor 11 and the rewinding motor 15 change with the temperature. Parameters in a simple control system fluctuate. This makes it difficult to perform optimal control for setting the tension applied to the product 2 to the set value, that is, optimal control of the position of the step roller.

Therefore, the knowledge information processing apparatus 20 is capable of responding to the above-mentioned parameter fluctuation in the control system.
Optimal control of the position of the step roller is performed.

FIG. 3 is a functional block diagram showing a detailed configuration of the knowledge information processing apparatus 20.

This knowledge information processing apparatus 20 includes an information processing unit 2
1, knowledge database 22, knowledge learning unit 23, history memory 2
4, current characteristics analysis unit 25, history characteristics analysis unit 26, history characteristics memory 27, history comparison unit 28, situation recognition unit 29, and knowledge change unit
Consists of 30.

The knowledge information processing apparatus 20 can be entirely realized by a hardware architecture, or can be realized by a programmed computer system. Further, a part of the knowledge information processing apparatus 20 can be realized by hardware, and another part can be realized by software.

Input data given from the preprocessing device 19,
The input target information is provided to the information processing unit 21, the knowledge learning unit 23, and the history memory 24.

The information processing section 21 infers input data according to knowledge data (control rules) stored in a knowledge database 22 to be described later, and outputs output data (target unwinding peripheral speed, winding motor). (A rotation speed command value).

The output peripheral speed and the take-up motor rotational speed command value, which are output data, are supplied to the servo device 18 and the driver 17 as described above, as well as to the knowledge learning unit 23 and the history memory 24.

The knowledge database 22 stores knowledge data (control rules) for performing the above-described inference in the information processing section 21. Knowledge Data Database 22 is RA
M, and realized by a memory such as a disk memory.

As knowledge data, different knowledge data is prepared in advance for each type (type) of the target product 2, and knowledge data of the type specified by the input target information is selected. Further, all or part of the knowledge data is changed by a knowledge learning unit 23 and a knowledge changing unit 30 described later.

An example of the knowledge data (control rules) is shown below.

R1: If the position of the step roller remains normal, slightly increase the target unwinding peripheral speed and increase the take-up motor rotation speed command value by the ratio of the winding diameter. If the roller position is fluttering, slightly lower the target unwinding peripheral speed and lower the winding motor rotation speed command value by the ratio of the winding diameter. R3: If the unwinding diameter is small or the winding diameter is small (Because the position of the step roller tends to become unstable) Slow down the target unwinding peripheral speed slightly and reduce the take-up motor rotation speed command value by the ratio of the winding diameter. R4: If the step roller position is lowered Ba
Increase the take-up motor rotation speed command value. R5: If the step roller position is up,
Decrease the take-up motor rotation speed command value. R6: If the take-up peripheral speed is lower than the take-out peripheral speed
Then, increase the take-up motor rotation speed command value. R7: If the take-up peripheral speed is faster than the take-out peripheral speed
Then, reduce the take-up motor rotation speed command value. R8: If the unwinding diameter is small (because the unwinding diameter becomes smaller immediately), slightly lower the take-up motor rotation speed command value. R9: If winding If the diameter is small, slightly reduce the take-up motor rotation speed command value (since the take-up diameter becomes larger immediately).

An input variable in each control rule, for example, R
The description “No. 1 indicates that the state in which the step roller position is normal continues” is described as “the time during which the absolute value of the step roller position deviation continues within the stability determination threshold value P0 is long”. "The number of times the absolute value of the step roller position deviation exceeds the variation determination threshold value P1 within the step roller position change reference time T0" is described.

The knowledge learning unit 23 learns the knowledge data stored in the knowledge data base 22 based on the input data from the preprocessing device 19 and the output data from the information processing unit 21.

In this embodiment, a case in which the consequent part constant relating to the unwinding peripheral speed and the take-up motor rotation speed command value of the consequent part will be described below.

When the unwinding peripheral speed target value is input by the operator as target information, learning is performed by the steepest descent method using the unwinding peripheral speed target value as teacher data.

Learning is performed according to the meta-knowledge data on the winding motor rotation speed designation value. An example of the meta-knowledge data is shown below.

MR1: If the position of the step roller is fluttering on the average, the absolute value of the consequent constant of R6 and R7 is slightly reduced, and the absolute value of the consequent constant of R4 is slightly increased. Yes MR2: If the step roller position is fluttering on average, slightly decrease the absolute value of the consequent constant of R6 and R7, and slightly increase the absolute value of the consequent constant of R5 MR3 : If the step roller position is fluttering without offset, reduce the absolute value of the consequent part constant of R4, R5, R6, R7. MR4: If the step roller position falls on average, it falls down and arrives. If it is, increase the absolute value of the consequent part of R4. MR5: If the step roller position is on the average and rises and then falls down, increase the absolute value of the consequent part constant of R5. R6: If the position of the step roller continues to decrease, increase the absolute value of the consequent constant of R6. MR7: If the position of the step roller continues to increase, increase the absolute value of the consequent constant of R7. MR8: If winding If the unwinding diameter is small and the position of the step roller is lower on average, reduce the absolute value of the consequent constant of R8. MR9: If the unwinding diameter is small and the position of the step roller is lower on average, R9 is used. Decrease the absolute value of the consequent constant

As described above, the learning of the knowledge data is performed. FIG. 4 shows an example of the transition of the consequent part constant by learning.

The learning of the knowledge data is not limited to the learning of the consequent part constant, but may be the learning of the antecedent part constant and the membership function of the knowledge data.

The history memory 24 stores input data supplied from the preprocessing device 19, output data supplied from the information processing unit, the consequent constants of rules of the knowledge data stored in the knowledge database 22, and the status described later. The type similarity from the recognition unit 29 (hereinafter referred to as “history data”)
The knowledge information processing device 20 stores the information for each target product. The history memory 24 is realized by a memory such as a disk memory.

FIG. 5 shows a data structure of data stored in the history memory 24.

FIG. 6 shows an example of input / output data.

The current characteristic analysis section 25 analyzes the input and output data within the current characteristic analysis reference time T1 retroactively from the current time to generate current characteristic data.

The analysis of the current characteristics is as follows.

One of the current characteristic analysis is to calculate the amount of change within the current characteristic analysis reference time T1 for each of the input and output data of the knowledge data (control rule) and to calculate the amount of change for each of the input and output data. Characteristic data.

In the second analysis of the current characteristic, a correlation coefficient between input / output data and each consequent constant of the knowledge data is calculated, and the correlation coefficient is used as the current characteristic data.

In this case, the correlation coefficient is calculated by the following least square method.

The input / output data is a variable x and its value is x
n (n = 1 to N; N is the total number of data included in the current characteristic analysis reference time T1). The consequent constant of the knowledge data is defined as a variable y, and its value is defined as yn (n = 1 to N). Assuming that the correlation coefficients are a and b, the equation of the correlation line is represented by the following equation.

[0071]

(Equation 1)

Therefore, the correlation coefficients a and b are given by the following equation (2).
To (7). FIG.
An example showing the correlation between the output data and the consequent part constant is shown.

As described above, the analysis of the current characteristic is performed, and the generated current characteristic data is given to the history comparing unit 28.

The history characteristic analysis unit 26 calculates a certain past time Tp.
From the history characteristic analysis reference time T2
The output data is analyzed to generate history characteristic data.

The analysis of the history characteristic includes the following as in the case of the current characteristic analysis unit 25.

One analysis of the history characteristic is that a certain past time T
The input and output data within the history characteristic analysis reference time T2 are read from the history memory 24 retroactively from p, and the amount of change is calculated for each of the input and output data, and the amount of change is used as the history characteristic data.

In the analysis of the history characteristic, the correlation coefficient between the input / output data and each consequent constant is calculated by comparing the input / output data and the input / output data within the historical characteristic analysis reference time T2 with a certain past time Tp. For the data read from the history memory 24, the subject part constants are calculated using the above equations (1) to (7), and the correlation coefficient is used as the history characteristic data. However, equations (1) to
N in (7) is replaced with the total number M of data included in the history characteristic analysis reference time T2.

As described above, the generated history characteristic data is registered in the history characteristic memory 27. However, if all the history characteristic data are registered in the history characteristic memory 27, the capacity of the history characteristic memory 27 becomes large. Only the characteristic data of the history characteristic data is registered in the history memory 27. Therefore, data whose history characteristic data is equal to or larger than the threshold value d is registered in the history characteristic memory 27. The threshold value d is calculated by the following equation.

[0079]

(Equation 2)

Here, the threshold value d is normalized by Ayy because D in the equation (9) has the order of the square of the variable y.

In this way, the history characteristic data that has been subjected to the threshold processing using the threshold value d is registered in the history characteristic memory 27.

When the history characteristic analysis reference time T2 in the history characteristic analysis unit 26 is equal to the current characteristic analysis reference time T1 in the current characteristic analysis unit 25, the total number N of data used for the history characteristic analysis and the current characteristic analysis, respectively. Since M becomes equal, the analysis of the history characteristic is unnecessary. In this case, the history characteristic analysis unit 26 performs the above-described threshold processing (provided that M = N in equations (8) to (10)) on the current characteristic data calculated by the current characteristic analysis unit 25. Is registered in the history characteristic memory 27 as history characteristic data.

The history characteristic memory 27 stores the history characteristic data obtained as described above, and is realized by a memory such as a disk memory.

FIG. 8 is a flow chart showing a series of processing procedures from the history comparing section 28 to the knowledge changing section 30.

The history comparison section 28 compares the current characteristic data supplied from the current characteristic analysis section 25 with the history characteristic memory 27.
To calculate the history similarity with all the history characteristic data stored in.

The history similarity is calculated as follows.

All the history characteristic data Pijk (i represents the type of input / output data, j represents the product type, and k represents the time when the history characteristic data is generated) stored in the history characteristic memory 27 are read out. The difference Sijk of all the current characteristic data Ni and the read history characteristic Pijk given by the current characteristic analysis unit 25 is calculated by the following equation (step 41).

[0088]

(Equation 3)

When the difference Sijk is calculated for all the history characteristic data read from the history characteristic memory 27, the history similarity Mjk indicating how similar the history characteristic data at the past time point k is and the current present characteristic data is calculated. Is calculated by the following equation (step 42).

[0090]

(Equation 4)

The difference Sijk between the current characteristic data and the historical characteristic data calculated by the equation (11) may be counted below the threshold value set in advance, and the counted value may be calculated as the historical similarity. The difference Sijk may be used as an input to perform fuzzy inference according to a preset fuzzy rule to calculate the similarity.

The history similarity M calculated as described above
The jk is given from the history comparison unit 28 to the situation recognition unit 29.

The situation recognizing unit 29 calculates the kind similarity indicating how similar the current kind is to the past kind based on the history similarity Mjk given from the history comparing unit. Is what you do.

Based on the history similarity Mjk, the kind similarity Mj having the maximum history similarity among the varieties is calculated by the following equation (step 43).

[0095]

(Equation 5)

The kind similarity is given from the situation recognition unit 29 to the knowledge changing unit 30 and is also given to the history memory 24 and stored.

The knowledge changing section 30 changes the knowledge data in the knowledge database 22 based on the kind similarity provided by the situation recognition section 29.

The change of the knowledge data is performed as follows.

The situation changing unit 29 determines the degree of similarity between the other varieties targeted in the past and the current varieties based on the variety similarity Mj given from the situation recognizing unit 29. If the target products have the same kind similarity Mj, the learning is determined to be oscillating (step 44).

When the learning oscillates (YE in step 44)
S), the knowledge data of another type that has not been executed in the past and is stored in the history memory 24 is read (step 4).
5) Calculate the similarity between the read knowledge data and the past knowledge data of the current object (step 4).
6). If the similarity is smaller than the threshold, the current knowledge data is changed to the read knowledge data (step 47).

If the learning is not oscillating (step
(NO at 44), reads the consequent part constant Ujm (m represents the type of the consequent part constant) from the history memory 24, which is the knowledge data of each type, which is the basis for generating the history characteristic data given the type similarity. (Step 48) After that, the part constant and the kind similarity Mj
, The consequent constant Um of the current knowledge data is calculated by the following equation (step 49).

[0102]

(Equation 6)

The knowledge data in the knowledge database 22 is changed to the knowledge data calculated as described above (step 5).
0).

Since the recognition result of the situation recognition unit 29 is stored in the history memory 24, it is possible to determine whether the current recognition result is appropriate by comparing the current recognition result with the past recognition result. . If the recognition result is not appropriate, the current characteristic analysis unit 25, the history characteristic analysis unit 26, the history comparison unit
The characteristic analysis processing, the history comparison processing, and the situation recognition processing in each of the situation recognition unit 29 are changed so that the recognition result differs from the past. Further, the knowledge changing process in the knowledge changing unit 30 is changed so as to obtain different knowledge data from the past.

[Brief description of the drawings]

FIG. 1 is a diagram showing an overall configuration of a control system in which a knowledge information processing device is applied to control of a winder.

FIG. 2 shows an example of time-series data of a step roller position.

FIG. 3 is a functional block diagram illustrating an internal configuration of the knowledge information processing apparatus.

FIG. 4 shows an example of a transition of a consequent part constant by learning.

FIG. 5 shows a data structure of data stored in a history memory.

FIG. 6 shows an example of input and output data.

FIG. 7 shows an example showing a correlation between input and output data and consequent part constants.

FIG. 8 is a flowchart showing a series of processing procedures from a history comparing unit to a knowledge changing unit.

[Explanation of symbols]

 2 Product 3 Unwind reel 4 Step roller 5 Take-up reel 6 Spring 7 Weight 11 Unwind motor 12 Unwind peripheral speed sensor 13 Step roller position sensor 14 Take-up peripheral speed sensor 15 Rewind motor 16 Unwind motor / driver 17 Wind Motor / Driver 18 Servo unit 19 Preprocessing unit 20 Knowledge information processing unit 21 Information processing unit 22 Knowledge database 23 Knowledge learning unit 24 History memory 25 Current characteristics analysis unit 26 History characteristics analysis unit 27 History characteristics memory 28 History comparison unit 29 Situation Recognition Department 30 Knowledge Change Department

──────────────────────────────────────────────────の Continuing on the front page (72) Inventor Rosei Seiko, 10th Hanazono Todocho, Ukyo-ku, Kyoto-shi, Kyoto (72) Inventor Tsutomu Ishida 10th Hanazono-Todocho, Ukyo-ku, Kyoto, Kyoto Omron shares Inside the company (72) Inventor Yumi Tsutsumi 10 Okayama Todocho, Ukyo-ku, Kyoto City, Kyoto Prefecture (72) Inventor Nobuo Nagasaka 10 Ohanadodo-cho, Ukyo-ku, Kyoto City, Kyoto Omron Corporation (72) Invention Omori Mamoru Egi 10-chome, Hanazono-Todo-cho, Ukyo-ku, Kyoto, Japan (56) References JP-A-3-265902 (JP, A) JP-A-4-302305 (JP, A) JP-A-5 -108132 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G05B 11/00-13/04 G06F 9/44 G06F 17/30

Claims (2)

(57) [Claims] 1. A knowledge database for storing knowledge data for controlling or processing an object, an information processing means for performing information processing on input data according to the knowledge data of the knowledge data base, and outputting output data; Knowledge learning means for learning the processing parameters of the knowledge data based on the input and output data of the information processing means and correcting the processing parameters of the knowledge data; input and output data of the information processing means and the knowledge data A history memory that stores the processing parameters as history data in the order in which they appear, by analyzing the history data stored in the history memory and within the current characteristic analysis reference time determined retroactively from the current time, The current characteristic that generates the current characteristic data representing the characteristic of the object at the current point in time Analyzing means, stored in the history memory, analyzes history data within a history characteristic analysis reference time determined retroactively from an arbitrary time in the past, and obtains history characteristic data representing characteristics of the target at the past time. When the history characteristic data generated by the history characteristic analysis unit is within a predetermined range, a history characteristic memory for storing the history characteristic data, and all the history characteristics stored in the history characteristic memory are stored. History comparing means for calculating a history similarity determined based on a difference between the current characteristic data and the history characteristic data; and a maximum similarity calculated by the history comparing means as a maximum similarity. Situation recognition means to choose out as
And a processing parameter of the knowledge data, which is a processing parameter of the knowledge data from which the history characteristic data giving the maximum similarity selected by the situation recognition means is selected from the processing parameters stored in the history memory. A knowledge changing means for changing a parameter; 2. A knowledge data base for storing knowledge data for controlling or processing an object, and information processing means for performing information processing on input data according to the knowledge data of the knowledge data base and outputting output data. A learning method for a knowledge information processing apparatus, the method comprising: learning about processing parameters of the knowledge data based on input / output data of the information processing means; correcting the processing parameters of the knowledge data; Input and output data and the processing parameters of the knowledge data are stored in a history memory in the order in which they appear as history data. Analyze the historical data over time to show the characteristics of the object at the current point in time. The current characteristic data is generated, and the history data stored in the history memory and within the historical characteristic analysis reference time determined retroactively from an arbitrary time in the past is analyzed, and the characteristic relating to the target at the past time is analyzed. Generating history characteristic data to represent, when the history characteristic data is within a predetermined range, storing the history characteristic data in a history characteristic memory; and for all the history characteristic data stored in the history characteristic memory, A history similarity determined based on the difference between the current characteristic data and the history characteristic data is calculated, the one with the largest history similarity is selected as the maximum similarity, and the maximum similarity stored in the history memory is calculated. Changing the processing parameters of the knowledge data with the processing parameters of the knowledge data from which the given history characteristic data is generated; Learning method of location.