JPH08110326A - Pattern generating method for neural network system - Google Patents

Abstract

PURPOSE: To generate a pattern with less data quantity by inputting and holding a processed value in a plurality of data processing areas for generating a teacher pattern and an input pattern, obtaining a representative value for the plurality of data in the area, and outputting the processed value of the areas in a preset order. CONSTITUTION: The shape of an object to be recognized is defined, stored in a memory, its shape interior is divided into a plurality of data processing areas N, a name of an aggregate for constituting the area N is set, and stored in the memory. The shape and the area N are displayed on a display, confirmed and then stored in the memory. The positions X(i), Y(i), Z(i) of a starting point and the sizes x(i), y(i), z(i) of the area are set for all the data processing areas (i) (=l-N), and stored in the memory. Then, a representative value calculating method in the area (i) is set, S=1 is set when the representative value is the maximum value, S=2 is set when the value is the mean value, S=3 is set when the value is the minimum value, stored in the memory, the order of outputting the representative values is set for the area (i), and stored in the memory.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern generation method in a neural network system, and more particularly to a pattern generation method in a neural network system applicable to any dimension.

[0002]

2. Description of the Related Art By irradiating a weld with ultrasonic waves and processing reflected ultrasonic information reflected from the weld by a neural network, the type of welding defects existing in the weld can be judged subjectively by an inspector. As a technique for objectively detecting without relying on judgment, “Japanese Patent Laid-Open No. 5-333001”
(First conventional technology). Also, as a technique for recognizing a two-dimensional handwritten character by a neural network, there is "Personal Neurocomputer" (second conventional technique) by NEC Information Technology, Inc.

Next, the first conventional technique, "JP-A-5-333001" will be described with reference to FIGS. FIG. 7 is a diagram showing a principle configuration according to a first conventional technique, in which 50 is a first member to be welded and 51 is a welding target.
Is a second member, 52 is a welded portion, and shows an example in which the first member 50 and the second member 51 are connected according to a three-stage welding layer formed by three welding processes. Reference numerals 53 and 53 are backing plates, which are used to prevent burn-through when forming the welded portion 52. Reference numeral 60 denotes an ultrasonic probe, which irradiates the cross section (YZ plane) of the welded portion 12 perpendicular to the welding line (X direction) with ultrasonic waves, and in response to the irradiation of this ultrasonic wave. The ultrasonic waves from the defective portion of the welded portion 52 that is reflected are received. When the ultrasonic probe 60 is scanned in a direction (Y direction) orthogonal to the welding line (X direction) to cover the entire cross-sectional area of the welded portion 52, the ultrasonic probe 60 moves in the welding line direction (X direction) by a predetermined distance. Has been
By scanning in the direction orthogonal to the welding line (Y direction) to cover the entire cross-sectional area of the welded portion 52 at the movement position,
Ultrasonic waves will be applied to the entire welded portion 52.

An ultrasonic flaw detector 61 controls the movement of the ultrasonic probe 60, controls the transmission of the ultrasonic waves, and receives the ultrasonic waves including the attenuation correction of the ultrasonic power depending on the propagation distance. Control and execute. This ultrasonic flaw detector 61 divides the welded portion 52 into a plurality of divisions in terms of coordinates, and controls the ultrasonic flaw detection device 61 in accordance with the divisions so that, for example, the reflection is reflected from each division. The maximum value of the coming distance-corrected ultrasonic power is detected as the representative reflected ultrasonic power of the section. Generally, the ultrasonic flaw detector 61 has a welded portion 52 when viewed in a coordinate space as shown in FIG.
With respect to a cross section (YZ plane) perpendicular to the welding line (X direction), in the Y direction, a portion (A1 to A5) adjacent to the first member 50 side and a second member 51 side are adjacent to each other. The part (B1 to B5) and the part sandwiched between these two parts (D
1 to D5, but not limited to one), while the Z direction is divided into subscripts 1 to 5 according to the number of weld layers. In addition, the welding line direction (X direction) is sectioned by the cross-section unit of the welded portion 52 for ultrasonic irradiation.

Reference numeral 62 denotes a weld defect inspection apparatus for detecting the defect type of the defect of the weld 52, which receives one or a plurality of inputs and an internal state value to be multiplied by these inputs, and sums the products. The basic unit for obtaining the final output by converting the sum of products value by the prescribed function is used as a basic unit, and the network data processing means 63 constituted by the network connection of the basic unit and the network data processing means 63 And an internal state value management means 64 for managing the internal state value possessed. In the case where the network data processing unit 63 matches the section configuration of the welded portion 52 by the ultrasonic flaw detector 61 and sets the unit portion of the welded cross section of the ultrasonic irradiation perpendicular to the weld line (X direction) as the defect detection target region. In the case where the representative reflected ultrasonic wave power of the partition included in the detection target area is input and all the welded parts included in the predetermined welding length are to be the defect detection target areas, they are included in the detection target area. When the representative reflected ultrasonic power of the partition is input and the specific welded portion included in the predetermined welding length is set as the defect detection target area, the representative reflected ultrasonic power of the partition included in the detection target area is set. input. Then, the input of the representative reflected ultrasonic power is matched with the defect type to be detected, and the type information of the defect type is output.

A teacher signal management device 65 manages a teacher signal required when the network data processing means 63 is constructed as the weld defect inspection device 62. This managed teacher signal includes various types of defects that may exist in the welded portion 52, and the representative reflected ultrasonic power that each section of the defect detection target area of the welded portion 52 indicates when the defect exists. Of the paired data, the representative reflected ultrasonic wave power among them becomes the input presentation side teacher signal, and the defect type becomes the output presentation side teacher signal. Reference numeral 66 denotes a learning processing device, which inputs a teacher signal on the input presentation side managed by the teacher signal management device 65 to the network data processing means 63 in accordance with the back propagation method or the like. The internal state value of the network data processing means 63 is learned so that the output presenting side teacher signal is output. Reference numeral 67 is a data preprocessor, and the ultrasonic flaw detector 61 has the same input configuration as the input data given to the network data processing means 63 by the learning processing device 66.
The representative reflected ultrasonic wave powers indicated by the sections are aligned and input to the network data processing means 63. Further, FIG. 8 is a diagram showing a system configuration of a welding defect inspection system, 70 is an XY table, 7
1 is a weld defect inspection device, and 72 is a weld defect inspection device 7.
It is the console of 1. FIG. 10 is a diagram showing a network configuration data processing function. Reference numeral 76 is an input unit to which a teacher signal and an unknown input signal are input, and 77 is neural network data processing of the teacher signal and the unknown input signal, and outputs the processed result. A basic unit 78 is a network configuration data processing means including an input unit 76 and a basic unit 77.

Next, the operation of the first conventional technique will be described with reference to FIGS. The learning processing device 66, when inputting the representative reflected ultrasonic wave power of the input presenting side teacher signal managed by the teacher signal managing device 65 to the network data processing means 63, the output presenting side forming a pair from the network data processing means 63. The internal state value of the network data processing means 63 is learned so that the defect type information of the teacher signal will be output. After that, the network data processing means 63, via the data pre-processor 67, each partition of the welded portion 52 (see FIGS. 9 and 10).
When the representative reflected ultrasonic power indicated by A1 to A5, D1 to D5, B1 to B5 shown in FIG. 10 is inputted, defect type information displayed by the inputted representative reflected ultrasonic power (shown in FIG. 10). , The first member side residual route, the second member side residual route, the first member side defective fusion, the second member side defective fusion,
Residual route on both sides, residual route gap failure, solidification cracking,
By operating so as to output slag inclusion, pinholes, blowholes, etc., it functions as the weld defect inspection device 62.

When the welded portion 52 is formed by welding the first member 50 and the second member 51 according to the number of times of welding a plurality of times, a boundary portion between the first member 50 and the welded portion 52 is formed. Then, a defect is likely to occur at the boundary between the second member 51 and the welded portion 52 and the adjacent boundary between the overlapping welding layers. From this, with respect to the welding cross section (YZ plane) perpendicular to the welding line (X direction), the welding portion 52 is adjacent to the first member 50 in the Y direction (A1 shown in FIGS. 9 and 10). .About.A5), a portion adjacent to the second member 51 (B1 to B5 shown in FIGS. 9 and 10), and a portion sandwiched between these two portions (D1 to D5 shown in FIGS. 9 and 10).
, And not necessarily one of the D columns), while the Z direction is divided according to the number of welding layers, and the network data processing means 63 is provided for each of the divided portions. Detects the types of defects existing in the welded portion 52. In this way, it becomes possible to objectively and accurately identify the type of defect existing in the welded portion 52 without resorting to the subjective judgment of the inspector.

Next, a second conventional technique, "a technique for recognizing a two-dimensional handwritten character by a neural network" will be described with reference to FIG. Figure 11
11A to 11C are diagrams showing a second conventional technique, and FIG. 11A is a diagram showing creation of a teacher signal and learning of a network. Prepare squares and set the characteristics of each category to the size of ■ (as data [1-0]
Corresponding to) indicates visually the teacher signal. In FIG. 11A, a plurality of teacher signals expressing the characteristics of the recognition target (in this figure, hiragana “o”) are prepared for each category (identification type “a” to “o”), and network learning is performed. I do.

Next, pattern recognition for an unknown input signal is performed in the following order using the learned network that learned the teacher signal. (1) Input of unknown input signal As shown in FIG. 11B, a pattern to be recognized is input to the square displayed on the computer screen by an input device such as a mouse. In the figure, 22 squares × 18 horizontals (396 in total) are prepared and stored as input patterns [0; white, 1; black].
indicate. (2) Execution of pattern recognition The following processing is performed on the signal input in (a). FIG. 11 (B)
2 x 2 squares as one unit, and [0, 0.25, 0.5, 1.0;
[Display in size]] is assigned to each unit. (B) Which category ("a" to "" is output by the learning network after being converted into the signal shown in FIG.
It is determined whether it is close to "O" identification type).

[0011]

However, in the first conventional technique shown in FIGS. 7 to 10, the types of welding defects to be detected in order to create the teaching signal, that is, the residual route, the fusion, It is necessary to prepare multiple welding test pieces each having defects, residual root gap defects, solidification cracking, slag entrainment, pinholes, blowholes, and a large number of test pieces are required, so much cost and There has been a problem that a large amount of measurement time is spent to acquire the data of each test piece. In addition, FIG.
In the second conventional technique shown in FIG. 11, the recognition target is limited to two dimensions, and as shown in FIGS. 11A to 11C, the size of the squares of the section is fixed, and the recognition of characters is performed. However, in order to correctly express the characteristics of welding defects such as blowholes scattered in a specific range of the welded portion, as in the first conventional technique, a specific range is required. It is necessary to set a large section in the, but when a fixed small section is applied, there is a problem that the feature quantity is extracted in a random place and the feature of the welding defect cannot be expressed correctly. Further, the method of converting from FIG. 11 (B) to FIG. 11 (C) is determined (averaged) by the number of input patterns of [1] in the unit, but like the first conventional technique. To
There is a problem that the maximum value or the minimum value is adopted in some cases, which is not always effective for other recognition targets.

[0012]

A pattern recognition method in a neural network system according to the present invention is made in order to solve the above-mentioned problems in the prior art, and one or more inputs and this input are multiplied. A network consisting of a network connection of the basic unit with the basic unit as a basic unit, which receives the internal state value to be obtained and obtains the product sum, and which converts the product sum value by the specified function to obtain the final output. From the step of using the configuration data processing means to learn the network by a preset teacher pattern for each identification type, and the next step of performing pattern recognition for an unknown input pattern using the learned network. Pattern generation method in a neural network system A function of setting a plurality of data processing areas in order to generate a turn and inputting and holding a processing value in each data processing area, and an operation of obtaining a representative value for a plurality of data existing in the processing area. It has a function of performing processing, and outputs the processing value of each of the data processing areas in a preset order.

The plurality of data processing areas can be set in any position by an empirical unique technique and in any size with respect to any dimension. By using the function of inputting and holding the processing value and inputting an empirical value to each of the data processing areas, it is possible to obtain a teacher pattern for each identification type, which exists in the plurality of data processing areas. It is also possible to obtain an input pattern and a teacher pattern from experimental data by using a function of performing a calculation process for obtaining a representative value for a plurality of data to be obtained.
It is possible to select from multiple calculation methods (maximum value, average value, minimum value, etc.), and to perform the calculation processing to generate a pattern with a small amount of data for experimental data and to perform network learning. Next, the processing speed at the stage of pattern recognition for an unknown input pattern may be improved.

[0014]

The pattern generating method in the neural network system according to the present invention, which is configured as described above, operates as follows. In the pattern generation method in the neural network system, in the plurality of data processing areas set to generate the teacher pattern and the input pattern, the processed values are input and held, and
Arithmetic processing for obtaining a representative value is performed on a plurality of data existing in the data processing area, and the processed value of each data processing area is output in a preset order. In the plurality of data processing areas, it is possible to set an arbitrary size in an arbitrary position by an empirical unique technique for an arbitrary dimension, and the processing value is set in the plurality of data processing areas. By inputting an empirical value to each of the data processing areas by using the input / hold function, a teacher pattern for each identification type can be obtained, and a plurality of patterns existing in the plurality of data processing areas can be obtained. It is also possible to obtain the input pattern and the teacher pattern from the experimental data by using the function of performing the arithmetic processing for obtaining the representative value for the data of
The calculation processing method for obtaining the representative value can be selected from a plurality of calculation methods (maximum value, average value, minimum value, etc.),
By performing the arithmetic processing method, it is possible to generate a pattern with a small amount of data for experimental data, perform network learning, and then recognize an unknown input pattern.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, an embodiment of a pattern generating method in a neural network system according to the present invention will be described with reference to the accompanying drawings. FIG. 1 is a diagram showing a cylindrical welded part to which a pattern generation method in a neural network system according to the present invention is applied, and is used to determine the type of welding defect occurring in the entire circumference welded part of the cylindrical welded part. One embodiment will be described with reference to FIGS. In FIG. 1, 1 is a first member constituting a cylindrical welded part, 2 is a second member, 3 is an entire circumference welded portion formed between the first member 1 and the second member 2, and 4 is a back surface. This material.

FIG. 2 is a diagram showing a groove shape for defining and displaying and confirming the shape of the entire circumference welded portion of the cylindrical welded part to be recognized by inputting parameters to the computer.
In FIG. 2, 5 is the groove shape of the entire circumference welded portion, and α1 is the first
The groove angle of the member, α2 is the groove angle of the second member, H1 is the groove height of the first member, H2 is the groove height of the second member, H3 is the groove root height, and R is the groove height. It is the radius of the previous root circle. FIG.
FIG. 3 is a diagram showing an embodiment in which the groove shape that is parameter-input to the computer shown in FIG. 2 is divided into a plurality of dividing portions (hereinafter, referred to as “data processing areas”). In FIG. 3, a space 6 (a space in which welding defect echoes are three-dimensionally distributed by ultrasonic flaw detection) including the groove shape 5 of the welded portion shown in FIG. 2 is defined, and an arbitrary space based on experience is defined in the space 6. , And defines a plurality of data processing areas 7 existing at arbitrary locations. In this embodiment, 35 data processing areas 7 are defined.

Next, an embodiment of the pattern generating method in the neural network system according to the present invention will be described with reference to the flow charts shown in FIGS. In addition, in order to make the explanation easy to understand, the terminology when the neural network system according to the present invention is applied to the detection of the welding defect type is added in parentheses after each term. FIG. 4 is a diagram showing a setting flow chart of a feature quantity (hereinafter, referred to as “pattern”) generation method in one embodiment of the present invention, which is common to generation of a teacher pattern and recognition of an unknown pattern. FIG. 5 is a diagram showing a teacher pattern generation flowchart in one embodiment of the present invention.

First, the computer processing of the setting flow chart of the pattern generation method shown in FIG. 4 will be described.
When the flow is started, in step S1, the [shape] of the recognition target (welded portion) is defined and stored in the memory. In step S2, the inside of the shape of the recognition target defined in step S1 is divided into a plurality of partition parts (hereinafter referred to as data processing areas).
The data is divided into [N], and the name [Name] of the aggregate formed by this data processing area is set and stored in the memory. The [shape] of step S1 and the data processing area [N] of step S2 are normally stored in the memory after being displayed and confirmed on the display. In step S3, the start point position [X (i), Y (i), Z (i)] and the size of the data processing area [x ( i), y (i), z (i)] are set and stored in the memory. In step S4, the representative value in the data processing area (representative value of ultrasonic echo level)
Set the calculation method [S]. That is, if the representative value is the maximum value, S = 1, and if the representative value is the average value, S = 2,
If the representative value is the minimum value, S = 3 is set and stored in the memory. In step S5, the entire data processing area [i
= 1 to N], the order of outputting representative values [D (j = 1
~ N)] is set and stored in the memory. The setting flow of the pattern generation method is completed by the above steps S1 to S5.

Next, the computer processing of the teacher pattern generation flowchart shown in FIG. 5 will be described. When the flow is started, in step S11, the [shape] of the recognition target (welding part) stored in the computer memory in step S1 of FIG. 4 and the aggregate [Name, N] including the data processing area are selected. Read into the processing unit. In step S12, as in step S11, the start point position [X] is set for all data processing areas [i = 1 to N] saved in the computer memory in step S2 of FIG.
(I), Y (i), Z (i)], size [x (i), y
(I), z (i)], a representative value calculation method [S] in the data processing area, and an order [D (j = 1) for outputting representative values to all data processing areas [i = 1 to N]. ~ N)] is selected and read into the arithmetic processing unit. After being read into this arithmetic processing unit, the entire data processing area [J = 1 to N] and the order of outputting representative values [D (J = 1 to N)] are represented. Step S1
In 3, the name (blowhole etc.) [Kname] of the identification type (welding defect type) is set. In step S14, the number of created teacher patterns [H = 1 to NN] (for example,
Number of teacher patterns indicating blowholes) is set. In step S15, the output array of the teacher pattern is set. The output array of this teacher pattern is [H = 1 to NN]
And a function [J = 1 to N] [T (H = 1 to NN, J = 1 to
N)]. In step S16, the method of creating the teacher pattern is set either by inputting an experience value or by setting it from experimental data.

When the teacher pattern is set by inputting the experience value, in step S17, the entire data processing area [J
= 1 to N] [empirical value (ultrasonic echo level) =
P] and store it in the output array. This [experience value =
P] is a function of [J = 1 to N] and P = T (J). In step S18, the number of teacher patterns created [H = 1 to N
N], the process of step S17 is repeated for the number of teacher patterns created [H = 1 to NN]. That is, [experience value = P] is [H = 1 to NN] and [J = 1 to 1]
N], P = T (H, J). Step S24
Then, the teacher pattern output array [T (H = 1 to NN, J
= 1 to N], the order of outputting representative values [D (K
= 1 to N)] = [J], the data is output to and stored in the memory of the teacher signal management device 65 shown in FIG. Step S25
Then, the name (blowhole etc.) [Kname] of the identification type (type of welding defect) is added to the [teaching pattern] stored in the memory of the teaching signal management device 65 in step S24 and the result is stored.

The above steps S11 to S18,
Then, in steps 24 and 25, the flow of generating the teacher pattern by inputting the experience value ends. In this way,
With the teacher pattern stored in the memory of the teacher signal management device 65 shown in FIG. 7, unknown pattern recognition (welding defect type detection) can be performed as in the conventional technique shown in FIG.

Further, in the case of setting the teacher pattern from the experimental data, the experimental data (ultrasonic echo level) previously collected and stored in a predetermined memory is read into the arithmetic processing unit and displayed in step S19. Is displayed on the screen, and is selected while confirming whether the experimental data is appropriate as a teacher pattern for the identification type (type of welding defect). In step S20, among the experimental data selected in step S19, [starting point (Xs, Ys, Z
s)] is selected. In step S21, step S2
The starting point [X (J), Y of the data processing area [J = 1 to N] of the aggregate from the [starting point (Xs, Ys, Zs)] selected with 0.
(J), Z (J)] is obtained, and for [experimental data] (ultrasonic echo level) existing within the size [x (J), y (J), z (J)] of the data processing area. , [Representative value = P] (maximum value of ultrasonic echo level) in the data processing area is calculated [S] to obtain [representative value = P] and stored in the output array. In step S22, step S
21 is repeated for all data processing areas [J = 1 to N]. That is, [representative value = P] is [J = 1 to N]
, P = T (J). In step S23, the processes of steps S20 to S22 are repeated for the number of created teacher patterns [H = 1 to NN]. That is, [representative value = P] is a function of [H = 1 to NN] and [J = 1 to N], and P = T (H, J). In step S24, the teacher pattern output array [T (H = 1 to NN, J = 1 to
N)], the order of outputting representative values [D (K = 1
~ N)] = [J], the data is output to and stored in the memory of the teacher signal management device 65 shown in FIG. In step S25, the name (blowhole or the like) [Kname] of the identification type (welding defect type) is added to the [teaching pattern] saved in the memory in step S24 and saved.

The above steps S11 to S16,
Then, steps 19 to 25 terminate the flow of generating a teacher pattern by inputting experimental data. In this manner, unknown pattern recognition (welding defect type detection) can be performed by the teacher pattern stored in the memory of the teacher signal management device 65 shown in FIG. 7 in the same manner as in the conventional technique shown in FIG. it can.

FIG. 6 is a diagram showing an example of creating a teacher signal corresponding to each type of welding defect. FIG. 6A is a diagram showing a basic shape of an aggregate composed of a plurality of data processing areas, and FIG. FIG. 7A is a diagram showing an example in which a pinhole teacher signal is input to the basic form shown in FIG. 8C, and FIG. 9C is a diagram showing an example in which a blowhole teacher signal is similarly input. FIG. 6A shows a cross section of a welded portion of an assembly of 35 data processing areas stored in a computer memory, each data processing area being numbered. Of the 35 data processing areas shown in (B), each data processing area of the black-painted portion, the area 23,
10 in each of the area 13, the area 7, and the area 19.
By inputting three ultrasonic echo levels of%, 20% and 30%, a total of 4 × 3 = 12 patterns of teacher signals are created. Of the 35 data processing areas shown in (C), each data processing area of the black-painted portion, areas 12, 14,
18,22,24 and areas 7,9,13,17,19
10%, 2 in the areas 14 and 17 and 12 and 19
By inputting four ultrasonic echo levels of 0%, 30% and 40%, a total of 4 × 4 = 16 patterns of teacher signals are created. That is, 12 shown in FIG.
Any input signal corresponding to the pattern teacher signal is determined to be a pinhole, and any input signal corresponding to the 16 pattern teacher signals shown in FIG. 6C is determined to be a blowhole.

[0025]

[Table 1]

Table 1 is a diagram showing an output example of the pinhole teacher signal shown in FIG. 6B. The numerical values of the data in Table 1 are the maximum of each data processing area with respect to the reference echo level (= 127%). Echo level E ratio (= E / 127)
Indicates. That is, No. No. 1 is an output corresponding to the teacher pattern in which the ultrasonic echo level 20% (= 20/127 = 0.157) is input to the area 23, No. 1 2 is area 23
Ultrasonic echo level 30% (= 30/127 = 0.2
36), the output corresponding to the input teacher pattern, No.
3 is the area 7 where the ultrasonic echo level is 10% (= 10/12
7 = 0.079) is the output corresponding to the input teacher pattern.

[0027]

As described in detail above, according to the present invention, the following effects can be obtained. (1) Even if there are many types of identification, a large number of test pieces are not required to create a teacher signal, and the cost for that can be suppressed. (2) In addition, for example, in order to correctly express the characteristics of welding defects such as blowholes scattered in a specific range of the welded portion, a large section is set in the specific range so that the welded portion is densely packed in the specific range. By setting a small section that correctly expresses the characteristics of the welding defect, the characteristics of the welding defect can be correctly expressed. (3) By adopting a maximum value, or in some cases a minimum value or an average value, as the feature amount, a feature can be correctly expressed according to a recognition target, and a representative value is adopted. Thus, the amount of information to be processed can be reduced and the processing speed can be improved.

[Brief description of drawings]

FIG. 1 is a side view showing a partial cross section of a cylindrical part to which a pattern generation method in a neural network system of the present invention is applied.

FIG. 2 is a view showing a groove shape for defining and displaying / confirming the shape of the entire circumference welded portion of the cylindrical part which is a recognition target of the present invention by inputting parameters to a computer.

FIG. 3 is a diagram showing an embodiment in which the space shown in FIG. 2 including a groove shape parameter-input to a computer is divided into a plurality of data processing areas.

FIG. 4 is a diagram showing a setting flowchart of a pattern generation method according to an embodiment of the present invention.

FIG. 5 is a diagram showing a flowchart for generating a teacher pattern according to an embodiment of the present invention.

6A and 6B are diagrams showing an example of creating a teacher signal corresponding to each type of welding defect, FIG. 6A is a diagram showing a basic shape of a plurality of data processing regions, and FIG. 6B is a diagram showing an input example of a pinhole; (C)
FIG. 4 is a diagram showing an input example of a blowhole.

FIG. 7 is a diagram showing a principle configuration in a first conventional technique.

FIG. 8 is a diagram showing a configuration of a welding defect inspection system in a first conventional technique.

FIG. 9 is a diagram showing a partition structure of a welded portion in the first conventional technique.

FIG. 10 is a diagram showing a network data processing function in the first conventional technique.

11A and 11B are detailed explanatory diagrams of a teacher signal in the second conventional technique, where FIG. 11A shows a teacher signal, FIG. 11B shows an input signal, and FIG. 11C shows processing of an input signal. Is.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... 1st member 2 ... 2nd member 3 ... Full circumference welding part 4 ... Backing material 5 ... Groove shape 6 ... Space containing groove shape 5 ... Data processing area α1 ... Grating angle of first member α2 ... Grating angle of second member H1 ... Gating height of first member H2 ... Gating height of second member H3 ... Height of groove root R ... Radius of root circle of groove S1 ... Define and save shape of recognition target S2 ... Set and save name of aggregate and data processing area S3 ... Position and size of start point and size of all data processing areas, save S4 ... Select and save representative value calculation method of data processing area S5 ... Set representative value output order of all data processing areas,
Save S11 ... Select and read shape and aggregate of recognition target S12 ... Read setting values of S3 to S5 S13 ... Set name of identification type S14 ... Set number of teacher patterns created S15. ..Setting of teacher pattern output array S16 ... Selection of teacher pattern creation method S17 ... Definition of experience values of all data processing areas and storage of the experience values in the output array S18 ... S17 Repeat as many as the number of teacher patterns created S19 ... Read / display experimental data and select experimental data suitable for teacher pattern S20 ... Enter the teacher pattern in the data processing area among the experimental data selected in S19 Selection of starting point to be performed S21 ... Obtaining a starting point of the data processing area from the starting point selected in S20 and calculating a representative value of experimental data within the size of the data processing area S22 ... S2 Repeat only for all data processing areas of S23 ... Repeat as many times as the number of teacher patterns created in S20 to S22 S24 ... Output / save the experience values or representative values set in the teacher pattern output array in order to the memory S25 ... Adds / saves the name of the identification type to the teacher pattern saved in S24

Claims (5)

[Claims] 1. A basic unit that receives one or a plurality of inputs and an internal state value to be multiplied by the inputs to obtain a product sum, and transforms the product sum value by a prescribed function to obtain a final output. A basic unit as a basic unit, and using a network configuration data processing means composed of a network connection of the basic unit, learning the network by a preset teacher pattern for each identification type,
Next, in a pattern generation method in a neural network system, which comprises the step of performing pattern recognition for an unknown input pattern using a learned network, a plurality of data processing regions are generated in order to generate a teacher pattern and an input pattern. Has a function of inputting and holding a processed value in each data processing area and a function of performing a calculation processing for obtaining a representative value for a plurality of data existing in the processing area. A method for generating a pattern in a neural network system, wherein the processed value of each of the data processing areas is output. 2. The plurality of data processing areas according to claim 1 can be set in arbitrary dimensions, at arbitrary positions by empirical unique techniques, with respect to arbitrary dimensions. Pattern generation method in neural network system. 3. A plurality of data processing areas according to claim 1, wherein the function of inputting and holding the processing values is used, and empirical values are input to each of the data processing areas, thereby identifying each type of identification. A pattern generation method in a neural network system characterized by obtaining a teacher pattern. 4. A teacher pattern is selectively selected from experimental data as an input pattern by using a function of performing arithmetic processing for obtaining a representative value for a plurality of data existing in a plurality of data processing areas according to claim 1. A method for generating a pattern in a neural network system, characterized in that 5. The calculation processing method for obtaining a representative value according to claim 4, is a plurality of calculation methods (maximum value, average value, minimum value, etc.).
The processing speed at the stage of learning a network by generating a pattern with a small amount of data for experimental data by performing the arithmetic processing and the step of performing pattern recognition for an unknown input pattern next time. A method for generating patterns in a neural network system, which is characterized by improving