JP2686023B2 - Pattern sorter - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern classifying device for classifying a pattern of an input signal by using a so-called neural network and a device for supporting the creation of the network. This pattern classifying device can be used, for example, for classifying a temperature distribution pattern of a blast furnace in an iron mill.

[0002]

2. Description of the Related Art In the field of automatically performing pattern classification processing and the like, recently, a technique of a neural network is often used. Studies on this type of neural network have already been made for use in various fields, and development and commercialization of products using the neural network are also being carried out in various fields. Also, tools that support the creation of the neural network are commercially available from various companies. As a conventional technique of this kind of neural network, for example, Japanese Patent Laid-Open No. 2-1
Those disclosed in Japanese Patent No. 57981 and Japanese Patent Laid-Open No. 3-33968 are known.

[0003]

[Problems to be Solved by the Invention] A neural network
In general, the performance of the network greatly changes depending on the network structure (the number of layers and the number of units in each layer), the number of learnings, the learning algorithm, the type of data used for learning, etc. The conditions will vary depending on the field to which it is applied.
It is very difficult to create a ku.

[0004] For example, in a steel mill, it is necessary to identify the state of formation of deposits on the wall of the furnace with respect to the circumferential temperature distribution pattern detected in the horizontal plane of the blast furnace. Balance is important, and for example, it is necessary to classify into eight types of patterns shown in FIG. 5, and control the furnace according to which pattern classification the current furnace condition belongs to.

The signal of this temperature distribution pattern was subjected to a classification process by simulation using a general neural network, but a low accuracy rate (the same classification result as an expert can be obtained). Rate). There are several possible causes for this result as follows.

That is, patterns of shapes to be classified into the same type.
However, if there is a rotation of the position in the horizontal plane between a plurality of patterns, a large change occurs in those patterns, and this change causes a decrease in the classification ability. Also,
In the pattern classification of FIG. 5, since the characteristics of the pattern are emphasized rather than the input data (temperature) itself, the temperature data
Pattern by just inputting the data group into the network as it is.
It is difficult for the characteristics of the computer to be reflected in the results. Furthermore, if some of the assigned classification patterns have a large difference from the others, an attempt to increase the accuracy rate for some of the classification patterns will result in the remaining classification patterns. The correct answer rate will be lower. For example, patterns 1 and 2 in FIG.
Since the difference is large compared to other patterns, if you try to classify all patterns using a single neural network, increase the learning amount for patterns 1 and 2 Instead of improving the accuracy rate for
The correct answer rate for patterns 3 to 8 decreases, and if the learning amount for patterns 3 to 8 increases, the correct answer rate for patterns 3 and 8 decreases, but the correct answer rate for patterns 1 and 2 decreases.

Therefore, an object of the present invention is to provide a pattern classifying device having a high classifying ability.

[0008]

[MEANS FOR SOLVING THE PROBLEMS]
, In pattern classification apparatus for outputting a signal indicating a classification of a combination pattern of a plurality of input signals is applied an input signal to the input terminal of the network comprises a network having an input terminal and a plurality of output terminals of the multiple, Attributes similar
Input multiple input signals on the radiation at the same angle from the origin.
Place the points according to the signal size and connect each point
A polygon on the surface is processed , the geometrical shape of the polygon is rotated , and the input signal is processed based on a predetermined evaluation.
The rotation processing means for changing the output and the plurality of input signals.
Feature values and geometric shapes obtained by using the number of samples as sample values
Statistics calculation processor that outputs the statistical features calculated from the
Stage, each of the rotation processing means and the statistic calculation processing means
Input the output and classify signals of the combination pattern of the input signals
Network means for outputting

Preferably, the rotation processing means is
Rotate the sequence of the entire input signal that has a geometric shape to
The correlation coefficient with the predetermined reference signal for each turn classification is the highest.
The input signal is output by selecting the rotation position that is the largest.

Further, the statistic calculation processing means is
Number of input signals as sample values
The shape of the mountain obtained from the three adjacent points from the geometrical shape
Output the statistical feature amount.

In addition, neural networks are
The first set of learning data with a high appearance ratio of input signal patterns
The first set of nets for which the weighting coefficient learning is performed
Work means and appearance ratio of the specific input signal pattern
Learning of the weighting coefficient for the second set of learning data
The second set of network measures performed and unknown inputs
A signal is applied to the first and second sets of network means
The first set of network means forwards the input signal when
When classified as a specific pattern, the first set of networks
The output of the network means and the first set of network means
When the input signal is classified as other than the specific pattern
Select the output of the second set of network means and select
An output selection means for outputting a class signal is provided.

[0012]

[Function] By rotating the arrangement order of all the plurality of input signals (see FIG.
Since the position-compensated signal is applied to the input terminal of the network, even if the pattern of the input signal causes rotation such as the temperature distribution signal of the blast furnace, the position in the direction of rotation is Reference pattern (that is, learning pattern)
Is applied to the network after being aligned so as to coincide with the position in the rotation direction. Therefore, the rotational position of the input signal pattern does not affect the network. As a result, the pattern classification ability of the network is greatly improved.

The first set of network means has a high classification ability for a specific input signal pattern but a low classification ability for other patterns, and the second set of network means.
The clock means has a low classification ability for the specific input signal pattern, but has a high classification ability for other patterns. When the neural network is trained by the usual method, that is, back propagation, the updating of the weighting factor by learning is applied to almost all units (neurons). Therefore, for example, when the learning for another pattern C / D classification is executed in the state where the learning for the pattern A / B classification has already been completed, the weighting coefficient for the already learned A / B classification. Is also updated,
A / B classification ability (correct answer rate) decreases. The tendency of the classification ability to decrease is such that a plurality of patterns having a large difference in shape (for example, an active pattern and a small active / inactive pattern in FIG. 5) have a small difference in shape. This becomes remarkable when the pattern classification (for example, the classification of patterns 3 to 8 in FIG. 5) is attempted at the same time. In the present invention, the learning data
Since the network means of the first set and the network means of the second set, which have different classification capabilities by changing the data, are provided, for example, classification of a plurality of patterns having a large difference in shape is performed in the first set. By using the network means and classifying the plurality of patterns having a small shape difference by the second set of network means, a plurality of patterns having a large shape difference and a plurality of patterns having a small shape difference can be obtained. High classification ability is obtained for any of the above.

The network is a pattern in which an input signal and a statistical amount of the signal (for example, an average value of a plurality of signals) are combined.
Type in to classify it. For example, in the classification of the horizontal plane temperature distribution pattern of the blast furnace shown in FIG. 5, the characteristics of the pattern of the entire temperature data are emphasized rather than the value of the data itself. Even if only data is input to the network, high classification ability cannot be expected. However, in the present invention, the statistical amount is also input as a feature amount to the network together with the input signal to classify both combination patterns, so that high classifying ability can be expected.

[0015]

FIG. 1 shows the configuration of a pattern classification device according to an embodiment. The pattern classification device of this embodiment has a temperature signal group (which indicates a temperature distribution in a horizontal plane at a predetermined position, which is detected by a large number of temperature sensors (thermocouples) arranged on the wall surface of the blast furnace as shown in FIG. The temperature of 12 points) is input to classify the steve temperature circumferential balance pattern as shown in FIG. 5, and the work station and process having the configuration shown in FIG. It is configured as a software process that runs on a computer. Of course, it is also possible to replace the apparatus of FIG. 1 with only hardware, or replace it with a combination of hardware and software.

Referring to FIG. 1, this pattern classifying device comprises a rotation processing section 1, a statistical calculation processing section 2, neural networks 3 and 4, and a data selector section 5. The 12 temperature signals obtained in the blast furnace are the rotation processing unit 1
And is input to the two sets of neural networks 3 and 4 through. In addition, the temperature signals of 12 points are calculated by the statistical calculation processing unit 2
Is also entered. The input layers of the neural networks 3 and 4 are composed of 16 units, and the output of the rotation processing unit is applied to 12 of them, and four types of statistical signals a and b output by the statistical calculation processing unit 2 are output. , C and d are applied to the remaining input units of the neural network.

The rotation processing unit 1 rotates the alignment order of the temperature signals of 12 points and aligns them. That is, in this embodiment, since it is necessary to classify the pattern shape with emphasis on it, the pattern obtained by rotating the reference typical pattern is the same as the original typical pattern. -There is a need to classify (see Figure 5). However, if only a specific pattern (typical pattern) is learned by the neural network, there is a high possibility that the rotated pattern will be misclassified. Therefore, in this example, in the rotation processing unit 1, the temperature signals (D1 to D12) of 12 points are rotated and aligned as shown in FIG. 7, and then applied to each neural network. .

As the concrete contents of the processing of the rotation processing section 1, first, the input signal is rotated so that the position of the maximum value of the data coincides with the typical pattern and the input signal pattern. Alignment was performed. As a result, the classification ability was improved compared to the case without alignment, but it was necessary to further improve the performance. Therefore, in this embodiment, the further improved rotation process shown in FIG. 6 is performed.

The rotation process will be described with reference to FIG. In step S1, 12 temperature signals are input, and in the next step S2, the value of the input temperature signal is normalized. In a succeeding step S3, reference patterns 1 to 8 which are created in advance.
Read the data (1-8 of the typical pattern in FIG. 5).
In step S4, the correlation coefficient between the normalized input temperature signal and each reference pattern is calculated. Save the calculation results. In step S5, the reference pattern to be compared is exchanged with another one. Then, the process of step S4 is repeated until the calculation of the correlation coefficient with all the reference patterns is completed. In step S7, the positions of the 12 input temperature signals are rotated by one step (one point) as shown in FIG.
Then, the process proceeds to step S4 again to calculate the correlation coefficient between the input temperature signal after rotation and each reference pattern. When the calculation of the correlation coefficient is completed for all the rotational positions, the process proceeds from step S8 to S9. In step S9, the largest one is searched among the correlation coefficients calculated in step S4, and the rotational position where the maximum correlation coefficient is obtained is detected. In the next step S10, the 12 input temperature signals are aligned so that the detected rotational position is reached.

Next, the processing contents of the statistical calculation processing unit 2 will be described. In this embodiment, the statistical calculation processing section 2 carries out the processing shown in FIG. That is, the temperature signals x (i) of 12 points are input in step S21, the average value a of x (i) is calculated in step S22, and the average value a of x (i) in step S23 is smaller than the average value a. The average value b is calculated, and in step S24 the angle of each crest of the input temperature signal pattern (see FIGS. 4 and 9) is detected. In the next step S25, the number c of points where the crest angle is 110 degrees or less. Is detected, and in the subsequent step S26, x larger than the average value + 1 sigma is detected.
The number d of (i) is counted, and in the final step S27, the four detected statistical values a, b, c and d are output.

The mountain angle detected in step S24 will be described. In this embodiment, the shape of the temperature pattern is represented by the chart shown in FIG. In FIG. 4, the center of the circle is 0 ° C., and the temperature scale is increased linearly from the center, and the position of the inner circle is 50 °.
The outer circle position is 125 ° C., and the position of each sensor is shown in the circumferential direction position shifted by 30 °.
The angle of each i-th peak of the temperature pattern on this chart is α [i + 1] + β [i] shown in FIG. 9, and is calculated by processing as shown in FIG. That is, when the distance between the point indicating the temperature value detected by the i-th sensor and the i + 1-th sensor and the center of the chart is L [i] and L [i + 1], respectively, the condition of each point is Accordingly, α [i] and β [i] are calculated as follows.

When L [i] = L [i + 1]: α [i] = β [i] = (π-π / 6) / 2 [rad] L [i]> L [i + 1] When: la = L [i + 1] × Cos (π / 6) lb = L [i] −la t = L [i + 1] × Sin (π / 6) α [i] = arctan (t / lb) [rad] β [i] = π-π / 6−α [i] [rad] When L [i] <L [i + 1]: la = L [i] × Cos (π / 6) lb = L [i + 1] −la t = L [i] × Sin (π / 6) β [i] = arctan (t / lb) [rad] α [i] = π−π / 6−β [ i] [rad] In this way, α [i] and β [i] at each point are obtained, and the result is used to obtain the angle α [i + 1] + β [i] of each mountain.

In the pattern classifying apparatus shown in FIG. 1, two sets of neural networks 3 and 4 are used, but these two have a similar structure except the learning contents. As shown in FIG. 2, the structure of the neural networks 3 and 4 is a three-layer structure of an input layer, an intermediate layer and an output layer. In the network 3, the input layer has 16 units and the output layer has The network 4 has 16 units in the input layer and 6 units in the output layer. The temperature signals of 12 points aligned by the rotation process are applied to 12 units out of 16 units of the input layer, and the statistical values a, b, c and d are applied to the remaining 4 units, respectively. 1 of output layer of network 3
The second unit is assigned to patterns 1 and 2 and 2
The 3rd, 3rd, 4th, 5th, 6th and 7th units are, respectively, pattern 3, pattern 4, pattern 5, pattern 6, pattern 7 and pattern 8 Assigned to.

As shown in FIG. 5, typical patterns of this embodiment are divided into eight types, but pattern 1 and pattern 2 are highly confused with each other, and both are also Since they are active and do not require classification in practice, as a result of integrating them into the same classification, the device of FIG. 1 was configured to classify 7 types of patterns.

The structure of each unit (cell) constituting the neural networks 3 and 4 is basically the same as that of a conventionally known technique, for example, the output of a unit having 1 to n input terminals. For the value, the input value and the weighting coefficient (that is, the coupling coefficient) for the i-th input terminal are respectively set to D (i)
And w (i) are output values of a function (for example, a sigmoid function) with respect to the sum of i = 1 to n of D (i) × w (i).

In this embodiment, as the learning data for learning one neural network 3, the data classified into patterns 1 and 2 shown in FIG. Twenty types of patterns classified into types and other types were prepared, and learning was performed a predetermined number of times for a total of 30 types of patterns (group B). A known back propagation was used as a learning algorithm.
As learning data for learning the other neural network 4, the pattern 3 shown in FIG.
24 in total for items belonging to 4, 5, 6, 7 or 8
Prepare different types of patterns (group A), each pattern
Training was conducted a predetermined number of times. Back-propagation was used as a learning algorithm.

Accordingly, one of the neural networks 3
Has a high classification ability between patterns 1 and 2 and the other patterns, and the other neural network 4 has a pattern.
The ability to classify those that do not belong to patterns 1 and 2 into any of patterns 3 to 8 is high. Therefore, as shown in FIG. 1, the switching control of the data selector 5 is performed by utilizing the classification outputs of the patterns 1 and 2 of the neural network 3, and the neural network 3 is unknown. Pattern 1, Pattern 1,
When classified into 2, the pattern of the neural network 3
When the classification output of patterns 3 to 8 is selected by the data selector, and the neural network 3 classifies an unknown pattern into a pattern other than patterns 1 and 2, the neural network 3 is selected.
The classification outputs of patterns 3 to 8 of group 4 are selected by the data selector, and the classification outputs of patterns 3 to 8 selected and the pattern of neural network 3 are selected. 1,2
The classification output of (1) is selectively output as the classification result of the pattern classification device. To be precise, the neural network 3 is dedicated to detecting the patterns 1 and 2.

The classification ability of this pattern classifier was compared with that of a skilled operator, and if the same classification result was regarded as the correct answer, a correct answer rate of about 85% was obtained. In addition, it is very delicate to judge which pattern should be classified for incorrect data, and even if this pattern classifier is used for actual operation, there is no problem. Was found not to occur.

An embodiment of the network creating device which is useful for efficiently developing the pattern classifying device as described above will be described below.

The network creating apparatus of this embodiment is constructed on the workstation shown in FIG. Referring to FIG. 10, this work station includes a central processing unit CPU, a display unit CRT, a keyboard KB, a mouse MS, a main storage unit MM, an auxiliary storage unit HD, a printer PR and a communication unit CU. Configured, communication unit CU
Is connected to the local area network. On the auxiliary storage device HD, basic software (that is, operating software)
Application system) and application software for a network creation device. After turning on the workstation and starting the operating system, log in to the system and then use the keyboard.
The system of the network creation device is activated by performing a predetermined input operation with the mouse KB or the mouse MS. When this system is activated, the processing shown in FIG. 18 is started.

In the first step S31 of FIG. 18, the main menu is displayed on the display device CRT. As a result, a menu screen as shown in FIG. 11, for example, is displayed. On this menu screen, as an operable key, an end key K1a, a model registration key K1b, a structure setting key K1c, a learning mode key K1d, a cancel key K1e, a model deletion key K1f, and an actual machine transfer Key K1g and evaluation mode
Doki-K1h is provided, and model NO area A
1a, model list area A1b, layer number list area A1c
Is provided.

By operating the model registration key K1b on the screen with the mouse or the keyboard, step S43 of FIG. 18 is executed. At this time, the model number, the model name, and the memo can be input to each column of the model list area A1b. The entered information is registered in the memory and displayed on the model list.

By designating the model NO area A1a on the screen, step S44 of FIG. 18 is executed. At this time, one model can be selected and designated from the models registered in advance (the models displayed in the model list). The number of the selected model is registered in the memory and displayed in the model NO area A1a.

By operating the structure setting key K1c on the screen with the mouse or the keyboard, step S45 of FIG. 18 is executed. At this time, the layer number list area A
The number of cells in each layer can be input for 1c.
The input cell number of each layer is registered in the memory as structural information of each layer for the model selected at that time, and the updated information is displayed in the layer number list area A1c.

When the cancel key K1e is operated, step S47 in FIG. 18 is executed. However, if the structure setting is completed in step S45 before that, one neural network is created in the memory in step S47. Created. In particular,
The source file is compiled based on the information of the C language source file prepared in advance, the set number of layers and the parameter of the number of cells in each layer, and the network data structure. And an object program that executes calculations for the network are created. This network is composed of two sets, a learning model and an evaluation model.

The created neural network is provided with a desired pattern classification function by performing learning. By operating the learning mode key K1d on the screen of FIG. 11, step 48 of FIG. 18 is executed, which allows the processing to proceed to the learning mode shown in FIG.

In the first step S51 of FIG. 19, the learning menu is displayed on the display device CRT. by this,
For example, a menu screen as shown in FIG. 12 is displayed.
On this menu screen, the operable keys are the end key K2a, the learning NO registration key K2b, and the data creation key.
K2c, cell allocation key K2d, cancel key K2e, data deletion key K2f, learning mode setting key K2g, learning key
K2h and a data analysis key K2i are provided, and further, a model NO area A2a, a learning NO area A2b, a real machine data file area A2c, a learning mode area A2e, an end condition area A2f, and a learning end value area A2g. , A learning data list area A2h and a learning status area A2i are provided.

By operating the learning NO registration key K2b on the screen with the mouse or the keyboard, step S63 of FIG. 19 is executed. At this time, the learning data number and data name can be entered. The input information is registered in the memory and displayed on the learning data list.

Further, by pointing the learning NO area A2b with the mouse or the keyboard, step S in FIG.
64 is executed. At this time, one of the registered learning data (data displayed in the list) can be selected. The number of the selected learning data is registered in the memory and displayed in the learning NO area A2b.

After selecting the learning data, by instructing the data preparation key K2c, step S in FIG.
66 is executed. In this step, the contents of the learning data are created and registered by the input operation of the operator. Actually, since a screen as shown in FIG. 15 is displayed, a line to be input is designated on this screen, and a pattern name, learning data and teacher data are input in each line. The example shown in FIG. 15 shows a case where the input layer of the network is composed of 6 units (excluding the unit of the statistical value) and the output layer is composed of 4 units. For example, the data on the first line
For example, when 50.0, 54.0, 50.0, 44.5, 35.0 and 16.4 are applied to the 6 units of the input layer, the teacher data of the 1st, 2nd, 3rd and 4th units of the output layer are It becomes 1.0, 0.0, 0.0 and 0.0 respectively.

On the display screen of FIG. 12, the cell allocation key
When K2d is designated by the mouse or the keyboard, step S67 of FIG. 19 is executed, and the process shown in FIG. In the first step S81 of FIG. 20, a cell allocation screen as shown in FIG. 13, for example, is displayed. FIG.
The above example shows the case where the input layer of the network is composed of seven units (cells). Information about the data assigned to each unit is specified on this screen. Items that can be specified are NO, input data name,
There are return value NO, normalization method, normalization constant and normalization function name. NO indicates the cell number to which data is assigned.

On the display screen of FIG. 13, the input key K3b
Is designated by the mouse or the keyboard, step 93 of FIG. 20 is executed and the process proceeds to FIG. In this case, by executing step S101, the displayed content of the screen opens a choice window as shown in FIG. 14, and a new key K4b,
K4c, K4e and K4f are displayed. By designating keys K4b, K4c, K4e and K4f on the screen shown in FIG. 14, original data, direct designation, statistical value, and user designation can be selected as the type of input data.

The selection of original data means that the data created on the screen shown in FIG. 15 is normalized and input to each cell, and the selection directly specified is created on the screen shown in FIG. It means that the obtained data is directly input to each cell without normalization, and the selection of the statistical value is performed by the data created on the screen shown in FIG.
It means that the statistical value of the data is input to each cell, and the selection by the user selects the data created on the screen shown in FIG.
This means inputting the result of processing by a function (subroutine) created by the user in advance to each cell. Also,
When the statistic value is selected, the following selection items are further displayed, and it is possible to select which of the basic statistic values is to be assigned. In this embodiment, as the basic statistical values, the average value, standard deviation, sum of squared deviations, variance, standard error, maximum value,
The minimum, range, median, and maximum position are provided. One of these is selected. FIG. 13 and FIG.
In the example shown in FIG. 4, the original data is assigned to the units 1 to 6 in the input layer of the network, and the average of the statistical values is assigned to the unit 7.

The return value NO indicates the cell to which the return value from the user specifying function is assigned. The normalization method specifies a normalization process when input data is normalized and applied to cells. In this example, the data is divided by a constant value and divided by a width (for each column). It is possible to select from three types: a method of normalizing by maximum and minimum values) and a method of normalizing by a user specification function. The normalization constant indicates a constant value when the method of dividing data by a constant value is selected as the normalization method. The normalization function name indicates the name of the function used when the user specification function is selected as the normalization method.

When the learning mode setting key K2g is designated on the screen of FIG. 12, step S68 of FIG. 19 is executed. At this time, the learning mode, the end condition, and the learning end value are set. As for the learning mode, either new learning or update learning of an existing learning file (weighting coefficient file) can be selected, and as for the end condition, the number of times of learning, error, and automatic are used as conditions for ending the learning. When the number of times of learning is selected as a condition, the number of times of learning is specified as a learning end value, and when an error is selected as a condition, an error between the network output and the teacher data is set as a learning end value (convergence value ) Is specified. The set conditions are registered, and the areas A2e, A2f and A2 on the screen are registered.
It is displayed in g.

After preparing for learning by operating as described above, when the learning key K2h is instructed on the screen shown in FIG. 12, the process proceeds to step S69 in FIG. 19 and the learning process is executed. In this learning process, a backpropagation algorithm is used. That is, the specified learning data is normalized as necessary, statistical processing is performed as necessary, input to the specified input layer unit, network calculation is executed, and the result is calculated. -The value obtained in each unit of the output layer of the output data is compared with the teacher data (see FIG. 15) accompanying the specified learning data, and the error between the two is approximated to 0. From the to the input layer, the value of the weighting coefficient (coupling coefficient) of each unit is modified little by little. This operation is repeated many times. The learning situation is displayed in a region A2i of the screen shown in FIG. 12 as a learning curve showing a change in the correspondence between the learning number and the error.

When the data analysis key K2i is designated on the display screen shown in FIG. 12, step S71 of FIG. 19 is executed. In this process, in order to analyze the characteristics of the designated learning data (see FIG. 15), the learning data is displayed as a graph, the learning data is printed, and the statistical value of the learning data is calculated and printed. It can be processed according to the operator's instruction. When displaying learning data as a graph, it is possible to selectively display two types of graphs, a line graph and a radar chart, and move the display position of each element on the graph (in the case of a line graph Can be moved in the axial direction and rotated in the case of radar chart).

After the learning is completed, the screen shown in FIG. 11 is returned to. When the evaluation mode key K1h is designated by the mouse or the keyboard, step S49 of FIG. 18 is executed and the evaluation mode shown in FIG. -Go to processing. In this case, first step S1
At 21, the evaluation menu as shown in FIG. 16 is displayed. On the screen of FIG. 16, the end key K6a, the evaluation NO registration key K
6b, data creation key K6c, cell allocation key K6d, pretreatment generation key K6e, cancel key K6f, data deletion key
K6g, input mode key K6h, display specification key K6i,
Assignment copy key K6j, evaluation key K6k, display cancel key
K6l, model NO area A6a, learning NO area A6b,
Evaluation NO area A6c, input mode area A6d, display mode
A display area A6e, a graph mode area A6f, a scale area A6g, and a display data list area A6h are provided.

In this evaluation mode, an unknown data pattern can be input to the created neural network and the classification result output by the network can be examined. When carrying out this evaluation, first the evaluation NO registration key K6b
After operating, enter the evaluation number and name. The entered information is registered and displayed on the evaluation data list area A6h. Next, the learning NO area A6b is moved to the mouse or keyboard.
22 is executed, the weighting factor data is selected here. That is, the network to be evaluated is created in advance in the learning mode.
One is selected from one or a plurality of weight coefficient data files. It is not possible to evaluate a network that has not completed learning.

Next, when the evaluation NO area A6c is designated by the mouse or the keyboard, step S145 in FIG. 22 is executed. Therefore, the evaluation data (area A registered in advance) (area A) is executed.
Select one of the data displayed in 6h). Then, when the data creation key K6c is instructed, step S146 in FIG. 22 is executed. The contents of the evaluation data selected here are input to create the main body of the evaluation data. in this case,
Although not shown, for example, a screen similar to the learning data creation screen shown in FIG. 15 is displayed, and similarly to that case, an unknown data pattern to be evaluated is input.
In the learning mode, it is necessary to input the learning data and the teacher data, but in the evaluation mode, the data to be evaluated.
Data (data applied to the input layer of the neural network)
Enter only).

Then, when the cell allocation key K6d is designated by the mouse or the keyboard, step S147 of FIG. 22 is executed. Here, the evaluation data is assigned to each unit of the network. In this cell allocation, the same processing as in the cell allocation in the learning mode is performed.

Next, when the mouse or the keyboard is used to instruct the pretreatment generation key K6e, step S148 of FIG. 22 is executed. In this process, the data set in step S147
A pretreatment module is created based on the data allocation.
This pre-processing module corresponds to, for example, the rotation processing unit 1 in FIG. 1, and is used after pre-processing the input data to be evaluated (alignment in the rotation direction in the example of FIG. 1). Apply to Lar network.

When the input mode key K6h is designated, as shown in FIG.
Step S150 of 2 is executed. Here, it is possible to select whether the data to be evaluated is read from a previously prepared evaluation data file and input, or directly input from the keyboard. In addition, when the display specification key K6i is designated, the display specification on the screen displaying the evaluation result can be changed.

Next, when the evaluation key K6k is designated, as shown in FIG.
The second step S151 is executed. When a file is selected as the input mode, when the file number is input, the data of the file is read as the evaluation target data, applied to the network, and the evaluation is started. That is,
The network is calculated for the evaluation data, the output value of each unit in the output layer of the network is displayed as the evaluation value, and the classification pattern of the unit with the highest evaluation result is displayed. -Is displayed as the evaluation result. Also, the evaluation target data and the learning data are displayed in a graph. As a result, a screen as shown in FIG. 17, for example, is displayed. In the example of FIG. 17, the evaluation target data is displayed as a line graph, but if the display specifications are changed, the evaluation target data can be displayed by the radar chart as shown in FIG. 4, for example. You can also

When the allocation copy key K6j is designated on the display screen shown in FIG. 16, the selected learning NO.
The cell allocation is automatically performed with reference to the registered information so that the data matches the cell allocation set when the learning is performed in the learning mode. That is, the information specified in each mode for each network is
As a data file associated with that network,
Since the data is stored in the main memory MM or the auxiliary memory HD, the cell allocation information specified in the learning mode,
That is, the original data associated with each input layer unit
Data, direct designation, statistical values (including categories such as average value and standard deviation) and user-designated category information, and information for normalization are read out, and the information is also used in the evaluation mode. When performing this automatic cell allocation, it is not necessary to input again for cell allocation in the evaluation mode, and the same conditions as in the learning mode are automatically reproduced and the network An evaluation can be carried out. As a result, it is possible to easily check which kind of statistical value is effective as the input information of the network for pattern classification.

[0056]

As described above, according to the pattern classifying apparatus of the present invention , the alignment order of the entire plurality of input signals is rotated and the alignment-compensated signal is applied to the input terminal of the network. For example, even if the pattern of the input signal causes rotation like the temperature distribution signal of the blast furnace, the position in the direction of rotation coincides with the position in the direction of rotation of the reference pattern (that is, the pattern to be learned). Are aligned and applied to the network. Therefore, the rotational position of the input signal pattern does not affect the network. As a result, the pattern classification ability of the network is greatly improved.

The first set of network means has a high classification ability for a specific input signal pattern, but has a low classification ability for other patterns, and the second set of network means.
The clock means has a low classification ability for the specific input signal pattern, but has a high classification ability for other patterns. When the neural network is trained by a normal method, that is, back propagation, the weighting factor update by learning is applied to almost all units, so that, for example, the learning of pattern A / B is already sufficient. If the learning for another pattern C / D classification is executed in the state that has already been completed, the weighting coefficient for the already learned A / B classification is also updated, and the ability of A / B classification decreases. Such a tendency of the classification ability to decrease is classified into a plurality of patterns having a large difference in shape (for example, the active pattern of FIG. 5).
And small active / inactive patterns) and a plurality of patterns with a small shape difference (for example, patterns 3 to 8 in FIG. 5).
It becomes remarkable when trying to carry out at the same time. In the present invention, since the first set of network means and the second set of network means having different classification capabilities by changing the learning data are provided, for example, a plurality of patterns having a large difference in shape. Classification is performed by the first set of network means, and classification of a plurality of patterns having a small shape difference is performed by the second set of network means, whereby a plurality of patterns and shapes having a large shape difference are formed. High classification ability can be obtained for any of a plurality of patterns with a small difference.

[0058] Ne Ttowa - click the pattern combines the statistics of the input signal and the signal - so by entering the down classify it can be expected high classification ability.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a pattern classification device according to an embodiment.

FIG. 2 is a block diagram showing the configuration of one network.

FIG. 3 is a perspective view showing a positional relationship between a blast furnace and a temperature sensor.

FIG. 4 is a radar chart showing one input signal pattern.
It is.

FIG. 5 is a radar chart showing a classification of typical signal patterns.

FIG. 6 is a flowchart showing the contents of rotation processing.

FIG. 7 is a map showing a sequence of signals before and after rotation processing.

FIG. 8 is a flowchart showing the contents of statistical calculation processing.

FIG. 9 is a radar chart showing one input signal pattern.
It is.

FIG. 10 is a block diagram showing a hardware configuration of the device.

FIG. 11 is a front view showing an example of a display screen of the network creation device.

FIG. 12 is a front view showing an example of a display screen of the network creation device.

FIG. 13 is a front view showing an example of a display screen of the network creation device.

FIG. 14 is a front view showing an example of a display screen of the network creation device.

FIG. 15 is a front view showing an example of a display screen of the network creation device.

FIG. 16 is a front view showing an example of a display screen of the network creation device.

FIG. 17 is a front view showing an example of a display screen of the network creation device.

FIG. 18 is a flowchart showing a part of the processing of the network creation device.

FIG. 19 is a flowchart showing a part of the processing of the network creation device.

FIG. 20 is a flowchart showing a part of the processing of the network creation device.

FIG. 21 is a flowchart showing a part of the processing of the network creation device.

FIG. 22 is a flowchart showing a part of the processing of the network creation device.

[Explanation of symbols]

 1: rotation processing unit 2: statistical calculation processing unit 3, 4: neural network 5: data selector

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Kiyoaki Nishio 5-3 Tokai-cho, Tokai-shi Nippon Steel Co., Ltd. Inside the Nagoya Works (72) Inventor Hiroyuki Miyazaki 20-1 Shintomi, Futtsu-shi (56) References JP-A-4-112277 (JP, A) JP-A-2-165288 (JP, A) JP-A 1-251288 (JP, A) JP-A-hei 4-195267 (JP, A) JP 2-81279 (JP, A) JP 2-287860 (JP, A) JP 2-101586 (JP, A) JP 2-310780 (JP, A) Yoshihisa Otsuka "Application of neural network to blast furnace data recognition" System / Control / Information VOL. 35, NO. 1, PP. 41-46, 1991

Claims (4)

(57) [Claims] 1. A pattern classifying apparatus including a network having a plurality of input terminals and a plurality of output terminals, applying a plurality of input signals to the input terminals of the network and outputting a signal indicating a classification of a combination pattern of the input signals. in, attribute
Of multiple similar input signals on the radiation at the same angle from the origin
And place each point according to the size of the input signal.
As a polygon on the connected plane, the geometrical shape of the polygon
Rotation processing is performed on the contrary, and input based on a predetermined evaluation
A rotation processing unit that rearranges and outputs signals, and
The statistical feature value obtained by using the input signal of
A statistic meter that outputs statistical features calculated from geometric shapes
Arithmetic processing means, the rotation processing means, and the statistic calculation processing
Input each output of the means, and
A pattern classification device comprising network means for outputting a classification signal . 2. The rotation processing means transforms the geometric shape into
Rotate the order of the entire input signal to make each pattern classification
Rotation that maximizes the correlation coefficient with the predetermined reference signal
A contract characterized by selecting a position and outputting an input signal
The pattern classification device according to claim 1 . 3. The statistic calculation processing means is configured to process the plurality of inputs.
The statistical feature and the geometry obtained by using the force signal as a sample value
On the shape of the mountain obtained from three adjacent points from the geometrical shape
2. The method according to claim 1, wherein the characteristic feature quantity is output.
Pattern classifier. 4. A plurality of input terminals and a plurality of output terminals are provided.
A network, including multiple input terminals
Apply input signals and classify combination patterns of input signals
In a pattern classification device that outputs a signal indicating
Of multiple similar input signals on the radiation at the same angle from the origin
And place each point according to the size of the input signal.
Let the polygon on the connected plane be the geometrical shape of the polygon.
Rotate the order of the entire input signal to make each pattern classification
Rotation that maximizes the correlation coefficient with the predetermined reference signal
Rotation processing means for selecting a position and outputting an input signal;
A statistic calculation processing means for outputting the statistical characteristic amount of the angle of the mountain determined from three points adjacent the serial plurality of input signals from the geometry as statistical characteristic amount obtained as a sample value, specific The first set of network means in which the weighting coefficient is learned with respect to the first set of learning data in which the appearance ratio of the input signal pattern is increased, and the second set in which the appearance ratio of the specific input signal pattern is decreased When the unknown input signal is applied to the first set and the second set of network means, the first set of network means performs weighting coefficient learning on the learning data. When the input signal is classified as the specific pattern, the output of the first set of network means is selected, and when the first set of network means classifies the input signal as other than the specific pattern, the second set of nets is selected. Selects the output of the over-click means outputs the classification signal selected, pattern classification apparatus comprising an output selection means.