JPH0452221A - Device for deciding whether production of iron and steel products is possible or not - Google Patents

Abstract

PURPOSE:To automatically decide whether production is possible or not for the conditions on the particulars of the orders for which the coinciding actual results do not exist in the past by previously allowing the device to learn the weight coeffts. of network in accordance with the past records of the decision, inputting the specifications of the iron and steel products in the input layer of the network and taking out the result of the decision from the output layers. CONSTITUTION:This device is provided with a means which is provided with plural network means including means having plural input means and at least one output terminals and holding the plural weight coeffts. to determine the correlations between the signal levels inputted to the input terminals and the level of the output terminal and inputs the information on the sizes and material quality conditions relating to the specifications of the iron and steel products to be produced; a selecting means which identifies as to which of the predetermined plural groups the contents of the inputted specifications belong and select one among the plural network means; a means which executes a learning the weight coeffts, of the above-mentioned network weans according to the past deciding results.; and a means which delivers the outputs of the network means selected by the above-mentioned selecting means as the result of the decision.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to determining whether or not steel products can be manufactured based on product specifications at the time of receiving an order and considering the order before receiving the order.

[Prior Art] Steel products have been the most widely used material for a long time, and depending on the type of use, various product specifications are requested by users to manufacturers in the form of order details. Ru.

The order details include, for example, tensile strength conditions, impact guarantee conditions, and test piece collection locations.

Specimen sampling direction 1, etc. are specified.

Therefore, manufacturers of steel products determine whether or not each product can be manufactured based on the order details. In other words, it is determined whether it is unconditionally manufacturable, manufacturable with two-condition materials, it cannot be determined without conducting a certainty test, or it is impossible to manufacture, and a reply is sent to the orderer. Although this type of determination is extremely difficult, it is extremely important for manufacturers to provide reliable customer service in a short period of time.

Systems that support this type of determination have existed and are actually in use. That is, this system is a combination of a database related to past manufacturing feasibility studies and a device that searches for data that matches the contents of the order details. If the order specification conditions completely match the searched past performance, it is used as a reference performance to determine whether manufacturing is possible. Therefore, when the order specification conditions are in accordance with the standard, there is a high probability that matching past data exists, and the search results of system t1 play a major role in determining whether manufacturing is possible.

[Problem to be solved by the invention] However, if the order details conditions requested by the user to the manufacturer include conditions other than those defined in the standard, they may not be completely consistent with past manufacturing feasibility review results. Matches are extremely rare, and in such cases, conventional search systems are of no use other than allowing experts to find examples of similar order details. will rely on the estimation of

Furthermore, when making a judgment based on human estimation, it takes a long time to make the judgment, and if the expert's skill level is low, the range of answers that can be answered is limited, which increases the frequency at which accuracy tests are required for judgment. Conducting accuracy tests leads to increased product costs.

Therefore, it is an object of the present invention to provide a system that can automatically determine whether or not a product can be manufactured even for order specification conditions that have not been matched in the past.

[Means for Solving the Problems 1] In order to solve the above problems, the first configuration of the present invention has a plurality of input terminals and at least one output terminal, and a voltage is applied to the plurality of input terminals. A plurality of network means including means for holding a plurality of weighting coefficients for determining the correlation between the signal level and the level of the output terminal, and inputting at least size information and material condition information regarding the specifications of the target steel product. means for identifying which of a plurality of predetermined groups the contents of the input specifications belong to, and selecting one from the plurality of network means according to the identification result; past determination; Means for learning weighting coefficients of the network means based on actual performance; and means for outputting the output of the network means selected by the selection means as a determination result are provided.

[Operation 1] In this configuration, a network means, that is, a neural network is used to automatically determine whether or not the product can be manufactured. The weighting coefficients of the network are trained in advance based on past judgment results, and the specifications of the steel product, that is, information on the size and material conditions, are input to the input layer of the network, and the judgment results are extracted from the output layer. .

Neural networks change their connection structure (strength of connections) depending on what they have learned, but they can output results with fairly high accuracy even for information they have not learned. Therefore, even if the information input as the specifications of the steel product does not exist as past results, the determination result can be obtained reliably and in a short time.

By the way, the judgment accuracy of a neural network varies greatly depending on the learning method, so it is very important whether it is easy to learn7. , for example H
For shaped steel, this is the flange thickness.

Tensile strength properties, impact guarantee conditions, test piece collection locations.

There are many things such as the direction of specimen collection, etc., but if all of this information is applied to the input layer of a single network, the network structure will become very complex and the learning process will be slow. It will also take up a lot of time.

Therefore, in the present invention, the content of information input as specifications is classified into a plurality of categories, and a plurality of network means are used, so that an independent network means is used for each category. Therefore, the configuration of each network becomes simple, and the time required for learning it is also shortened. Also, when acquiring knowledge about manufacturability judgment from experts for actual learning purposes.

It is more efficient to divide judgment cases into multiple classifications, and the learning process becomes easier.

Note that "judgment performance" here refers to performance information that is a combination of the order specification information at that time and the actual manufacturing results for it, and a combination of the order specification information at that time and the results of accuracy tests, etc. for it. This means both actual performance information.

[Means for Solving the Problem 2] In order to solve the above problem, the second configuration of the present invention includes a database in which past results are accumulated,
A primary judgment means that inputs information regarding the specifications of the target steel product and outputs a judgment result therefor; and when the primary judgment means requests a secondary judgment, inputs information on the specifications and outputs a secondary judgment result. and a secondary determination means that has a plurality of input terminals and at least one output terminal, and has a signal level applied to the plurality of input terminals and a level of the eight-power terminal. A network means including means for holding a plurality of weighting coefficients for determining the correlation of the network means, and a means for learning the weighting coefficients of the network means based on past judgment results are provided.

[Effect 2] In this configuration, in order to automatically determine whether manufacturing is possible, there is a primary determination means that performs determination based on the contents of a database in which past results are accumulated, and a primary determination means that performs determination based on the contents of a database in which past results are accumulated; A secondary determination means is provided to make a determination when the value cannot be obtained, and a network means, that is, a neural network is used for the secondary determination means.

As the primary determination means, a conventional search system or a device that performs determination processing based on predetermined determination rules can be used. The primary determination means has no ambiguous elements, so it can be expected to have high determination accuracy and high reliability. However, since the range that can be determined is limited to a certain extent, there is a high probability that products outside of this range will be determined to require "probability testing," for example, even if they are actually manufacturable. When conducting a certainty test, it takes a long time to obtain a correct judgment result.

In the present invention, when the determination result of the primary determination means is insufficient, the result is outputted by the secondary determination means.

Since the secondary judgment means uses a neural network, the judgment accuracy is lower than that of the primary judgment means, but
Even when special conditions that do not exist in the database are input, the judgment results can be output immediately. This result can be fully used as provisional information until the results of the confirmation test are obtained.

Other objects and features of the present invention will become clear from the following description of embodiments with reference to the drawings.

[Example 1] FIG. 1 shows the configuration of a manufacturing feasibility determination device according to an example. Referring to FIG. 1, an input/output device (IO) including a keyboard and a CRT display is connected to the expert system ES. The expert system ES is composed of a computer main body for data processing, hardware including a storage device, a neural network group 30, and various software stored on the storage device. The software part consists of database 10. It consists of an inference engine 2° and a knowledge base 30. In this example, a host computer CPU is also connected to the expert system computer.

Although the details will be described later, the database 40 is a database that organizes and registers a large amount of performance information regarding manufacturing cases and manufacturing feasibility examination cases of steel products, and the knowledge base 10 is a collection of rules required for actual manufacturing feasibility determination. The inference engine 20, in which information is registered, is a program that searches and judges information on the database 40 according to rules written on the knowledge base 10, and derives judgment results.

In other words, when a user inputs the order details conditions regarding steel products from the input/output device IIO, the expert system
S starts the operation of the inference engine 20, repeatedly judges various rules and searches for performance data, and outputs the final judgment results to the input/output device! Output to IO.

When the devices shown in FIG. 1 are functionally arranged, the configuration is as shown in FIG. 2. As can be seen from Figure 2, this device has -
A system 100 for examining the possibility of next production and a system 200 for examining whether or not secondary production is possible are provided.

The primary manufacturing feasibility examination system 100 is activated first, and if sufficient results cannot be obtained thereby, the secondary manufacturing feasibility examination system 200 is subsequently activated.

The secondary manufacturing feasibility examination system 200 is configured by a neural network distribution rule 11 and a neural network group 30. In this example, the neural net group is A,
Although only four, B, C, and D are shown, there are actually many more, as will be described later. Each neural network is connected to a storage device that stores the respective learning knowledge. Each neural network is independent, and rule 11 determines which network to use and the distribution of neural networks.

FIG. 9 shows an outline of the processing of the expert system ES. Note that steps 81 to S5 belong to the primary manufacturing feasibility examination system 100.

The contents of each processing step in FIG. 9 will be explained below.

In step S1, when information on required specifications for steel products is inputted as order specification conditions from the input device 2, the process proceeds to the next step S2. Here, the database 40 is searched for manufacturing examples and manufacturing feasibility study examples that satisfy part of the conditions. One item of data on the database 40 is shown in Table 1 below.

On the database 10, as shown in Table 1, various items of information are registered for each manufacturing case or manufacturing feasibility study case. Note that manufacturing cases are related to products that have actually been manufactured, and manufacturing feasibility study cases are related to products that have actually been manufactured.
It has been confirmed through testing and analysis that it can be manufactured, and it can be treated as 100% manufacturable and past results. On the database 10, there are many cases (manufacturing cases and manufacturing feasibility examination cases: the same applies hereinafter).
is registered.

When searching for data in step S2, various search condition rules existing on the knowledge base 30 are referred to to extract data that satisfy a predetermined condition. Next, typical examples of search condition rules are shown for each item. In the following explanation, "requirement specifications" refers to specifications developed by developing order details conditions based on database items.In addition, requirements specification values that do not exist are not considered during the search.

(1) Size and shape items: The case to be searched must be the same as the product name in the required specifications.

The case to be searched must be the same as the series of requirements specifications.

The case to be searched must be thicker than the required specification flange thickness.

(2) Tensile test items: If the required specifications require two tensile test directions, the case to be searched must be in two directions.

(3) Impact test items: If the impact test direction in the required specifications is two directions, the case to be searched must be in two directions.

If the required specification impact test sampling area is 1/2 part of the flange, the case to be searched must be part of the corner, 1/4 part of the flange, or 1/6 part of the flange.

In the example to be searched, the tensile test sampling part of the required specification is Web 1.
/3 parts, it must be part of the corner, 1/4 part of the flange, or 176 parts of the flange.

If the proof stress classification of the required specifications is the upper yield point, the case to be searched must be at the upper yield point.

The case to be searched must have a tensile strength lower than the upper limit of the required specifications.

The case to be searched has a required specification of impact test piece type of 2 mm.
For V, it must be 2 mm V.

The case to be searched must be below the impact test temperature value of the required specifications.

(4) Bending test items: The case to be searched is 2 if the bending test direction of the required specifications is Z.
Must be in the direction.

The case to be searched must be a part of the corner if the bending test sampling part of the required specification is part of the corner.

The case to be searched must have a bending angle value greater than or equal to the required specification.

(5) Weldability item: The case to be searched is A if the component CEQ formula of the required specification is A.
or C or D or E or F or G or I or J or K or L or M or N or P or Q or R or T or U.

(Definition of symbols of CE Q formula) A: C+Mn/6 B: C+Mn/10 C: C+Mn/6+Si/24 D: C+Mn/6+Si/24+Cr15+V/
14T: C+Mn/6+(Ni+Cu)/15U
: C+Kn/6+ (Ni+Cu)/15+ (
Cr+No+V)15 The case to be searched must be less than or equal to the component CEQ upper limit of the required specification.

In the next step S3, as a result of the search in step S2, 1
Identify whether three or more referable (valid) cases have been extracted. If a referenceable case is extracted, the process advances to step S4.

In step S4, a single case is selected as preprocessing for determining whether manufacturing is possible. That is, if only one example is extracted in the process of step S2, the determination is made based on it, but if multiple referenceable examples are extracted, the most appropriate one among them is selected. and make a judgment based on the selected cases.

Actually, for each extracted case, a value called a counter number indicating the degree of satisfaction is determined, and the case with the smallest number of counters is selected first. The number of counters is a number that indicates how many items the case does not satisfy with respect to the order details condition. 6 Here, in order to find the number of counters, the rules for calculating the number of counters are created on the knowledge base 30. The inference engine calculates the counter number without referring to the calculation rule. Typical examples of calculation rules are shown below, categorized by item.

(1) Component item: If the C lower limit of the searched case is smaller than the C lower limit of the required specification, set it as a counter (add 1).

If the C upper limit of the searched case is larger than the C upper limit of the required specification, it is set as a counter.

If the Mn/C lower limit of the searched case is smaller than the Mn/C lower limit of the required specification, it is set as a counter.

(2) Base material heat treatment method: If there is no designation for the base material heat treatment in the searched case, if there is a designation in the base material heat treatment in the required specifications, it will be counted as a counter.

If the base material heat treatment method in the searched case is normalizing, and the required specification is not normalizing, use the counter.

In addition, if multiple cases with the minimum number of counters calculated based on the above rules are detected, in order to select one from them, the conditions of the case with the lowest manufacturing cost and the latest case are added. and select. by this,
Eventually only one instance will be detected.

In step S5, it is determined whether manufacturing is possible or not based on one case detected in step S4. First, from the detected case data, extras, manufacturing time, necessity and period of accuracy testing, ingredients (C, Si, etc.) are examined.

Mn, P, S, Cu, Ni, Cr, Mo, Nb.

V, Ti, B, At, Ca, Nt H+ Mn/C,
Nb/V.

The target value of upper and lower limits (V/N, etc.), the values of items related to the manufacturing method such as the base material heat treatment method, deoxidizing conditions, and controlled rolling method are extracted. These values will remain as they are and will be the resulting values for the current order details conditions. Next, in the same manner as in step S4, the value of the current order detail condition and the resulting value are compared for the item related to the manufacturing method, and a counter value for the current order detail condition is derived. For example, if the Mn upper limit of the current order details condition is 1.35% and the resulting Mn upper limit target value is 1.60%, the counter value of the Mn upper limit will be 1.60%. Further, if the C upper limit of the current order specification condition is 0.23% and the resulting C upper limit target value is 0.17%, there is no C upper limit counter value.

If there is no item for which the counter value derived as described above exists and no certainty test is required, the manufacturability determination result will be "manufacturable"; conversely, the item for which the counter value exists If there is even one, or if a certainty test is required, the manufacturing feasibility determination result will be "manufacturable with conditions", and the conditions are the period of the certainty test and the retrieved counter value. An example of the results of the manufacturing feasibility determination is shown below.

“The result of determining whether or not it is possible to manufacture for this order details condition is that it is possible to manufacture.” “The result of determining whether or not manufacturing can be done for this order details condition is Mn
Manufacture is possible on the condition that the upper limit is increased from 1.35% to 1.60%. ” “The manufacturing feasibility determination results for this order details condition are that it is possible to manufacture on the condition that normalizing treatment is performed. Also, X yen/Sh On is required as heat treatment extra.” These manufacturing feasibility determination results are as follows. , is registered in the database 10 and used for the next manufacturing feasibility determination.

In step S2, if case data that satisfies the minimum order details conditions is not found, step S2
3 and step S6. Proceed to S7.

In step S6, a neural network is used to output the determination result. This result can be obtained in a short time.

Step S7 is a task that requires human intervention, and either a more detailed judgment is performed using material test data (scatter diagram data for tensile tests, transition diagram data for impact tests, etc.), or a certainty test is actually performed. By doing so, it is determined whether or not it can be manufactured. Therefore, it takes a long time to obtain the results. This result is registered in the database 1o as a new track record, and will be used in the subsequent manufacturing feasibility determination.

Next, the determination of whether or not secondary manufacturing is possible in step S6 will be explained in detail.

When determining whether or not a steel product can be manufactured, the input product specifications include the flange thickness in the case of H-beam steel.

Tensile strength properties, impact guarantee conditions, test piece collection locations.

There are many options such as the direction in which the specimen is taken. If these were all input parameters to a single neural network, the network structure would become extremely complex. As the structure of the network becomes more complex, the calculations become more complex and the number of parameters increases during the learning process, which requires an enormous amount of time. In neural networks, the quality of learning is very important.
In order to obtain highly accurate results, it is better to simplify the network structure.

For example, in the case of H-beam steel, the order details regarding flange thickness and tensile strength characteristics include: ■ Contents of the basic components carbon, silicon, and manganese; ■ Type of slab before rolling (IC
~Bulking slab or continuous casting slab), shape, and rolling method. Furthermore, if the impact guarantee conditions, test piece sampling location, and test piece sampling direction become stricter, it becomes necessary to change the refining process for adding special elements and to change the cooling conditions during rolling, which has a large effect on whether or not IJij can be produced. Therefore, even when considering whether or not H-beam steel of the same size can be manufactured, it is necessary to change the judgment method depending on the presence or absence of these conditions, so it is desirable to classify them in advance and handle them as a system.

Therefore, in this embodiment, a large number of independent neural networks are provided, and in order to selectively use any of them, the neural network distribution is made according to rule 1.
1 exists. The contents of rule 11 are as shown in FIG. That is.

There are independent neural networks depending on the size of the flange, the presence or absence of tensile strength conditions, the presence or absence of impact guarantee conditions, differences in the test specimen collection location, and differences in the test specimen collection direction, and depending on the input specifications at that time. The selected network will now be selected. Therefore, each neural network has a relatively simple configuration.

The configuration of one neural network is shown in FIG.

Referring to FIG. 3, in this example, the network has a three-layer structure: an input layer, a middle layer, and an output layer, and each layer has 16, 11, and 5 units, respectively. . Input layer numbers 1 to 5, 6 to 10, and 11 to
No. 15 has tensile strength Ts and flange thickness tf, respectively.
Information on the shock guarantee temperature Tiv is input. Each unit in the output layer is given the following meaning.

1: Cannot be manufactured 2: Cannot be manufactured 3: Confirmation test required 4: Can be manufactured 5: Can be manufactured The information on the tensile strength Ts, flange thickness tf, and guaranteed impact temperature Tiv applied to the input layer is based on the respective values. The information is classified into one of five ranks as shown in FIG. 4 depending on the size of the data, and the pattern of information applied to each unit of the input layer changes depending on the classification position. For example, if the tensile strength Ts is 45, it is within the range of 40 to 50, so unit No. 2 of the input layer will be at the target level, and units No. 1, 3.4, and 5 will be at the non-target level. . Target level Yi
teeth.

It is obtained from the following formula.

Y i = ((X i −X1M1n)/
(Xi ax -X 1M1n)) + AHowever, X
i: Input level X1M1n: Lower limit value XiMa of the range of unit i
x: Lower limit value of the range of (i+1) unit A: Bias value (0, 5) For example, if the input level Xi of tensile strength Ts is 52,
Since 1M1n is 50 and X iMax is 55, (
(52-50)/(55-50))+0.57) By the calculation, 0.9 is applied to the target unit (No. 3) of the input layer. The bias value A is provided to clarify the differentiation between input layer units that have non-target levels regarding the same factor. Furthermore, a minute value (0, 1) is applied as a bias value to the input layer unit at the non-target level. In other words, if the input level Xi of tensile strength Ts is 52, then 1
, 0.1 is applied to the input crab units of numbers 2, 4, and 5. The same applies to flange thickness and impact temperature.

In this embodiment, regarding the secondary manufacturing feasibility examination system, a manufacturing feasibility determination mode and a knowledge update mode are provided. When the learning of the neural network group 30 has been completed in advance, when the manufacturing availability determination mode is started, one neural network selected according to the rules as shown in FIG. The order details conditions are converted into the pattern described above and input to the network as input information to the input layer.

As a result, the determination result is output to the output layer of the neural network, and is output as a provisional determination result. In other words, when using neural networks, there is some ambiguity, just like human judgment.
Since 100% reliability cannot be expected, it is ultimately necessary to perform detailed human examination and accuracy testing (S7). In knowledge update mode, enter performance data and
Automatically learn the weight parameters (coupling strength) of each part on the network. A well-known backpropagation method (pack propagation) is used as a learning algorithm.

The basic operating principle of the neural network used here is the same as a known one, and will be explained below.

The basic structure of one of the units making up the network is shown in Figure 6a. See Figure 6a. Unit j in the S layer receives input information from multiple units in the (s −1) layer below (closer to the input layer than the S layer),
It functions as a multi-input-output element that performs conversion based on certain rules within the unit and outputs the result. Furthermore, in the unit coupling parts, variable coupling weights W a j
Set i. , Waji takes a positive, 0, or negative value, where positive corresponds to an excited state and negative corresponds to an inhibited state. If it is 0, it is the same as there being no connection between units.

Sum of inputs NETsj of unit number j in hierarchy S
is expressed by the following equation.

NETsj=Σ(Wsji ・O(s −1)i) ・
・”(1) However, i: The number output Osi of the lower layer unit of unit j can be obtained by converting NETsj by the input/output function f.Here, the function f is a sigmoidal (S-shaped) logistic function 0sj = f (
NETsj) = 1 / (1+exp (-NETsj)) ・
” (2) (see Figure 7c). This function is a monotonically bounded, continuous (n homogeneous) nonlinear function,
The value range is [0, 1], and as the input value increases, it approaches 1,
As it becomes smaller, it approaches 0. In addition, by adding a threshold value θaj (i+1) to the following equation and giving separate input/output functions to each unit, the degree of freedom of memory stored in the network is increased.

○5j== f (N E Tsj) = 17 (1+exp (−NETsj + θ5j(i
+ t )))...(2b) In reality, for each gradation, a unit (
This function is realized by adding a dummy unit (see Figure 3).

In this way, for any input data pattern P, the output value Osj for each unit j of the intermediate layer is obtained, and by propagating this to the units of the upper layer in sequence, the output value Osj is finally output to the output layer (layer Unit 2 with number 0)
Obtain the output value Ooz for (forward propagation: see Figure 6c)
.

In this embodiment, a well-known back propagation (pack propagation) method is employed as a learning method.

This is to sequentially change the connection weights backwards (from the output layer to the input layer) so that the obtained output value of the output layer approaches the desired output value (teacher data). .

Now, in order to evaluate the difference between the obtained output value of the output layer and the desired output value, in any input data pattern P, the actual output value of the output layer unit 2 is Ooz, and the desired output value is Tpz, and the error is calculated. The function Epz is defined as follows.

Epz=Σ ((Tpz-Opz)2/2) =・(
3) Here, the connection weight W sj for pattern P
The amount of change in i is ΔWsji=nee ηlδE pz/ a Wsji)
...(4). This is to obtain the optimal solution using the steepest descent method. By further transforming this, the following equation is obtained.

ΔWsji= 71 ・(δEpz/θWii)=-
rt-CδEpz/dNErsj) (60y/δW
sji)=-η・(δE pz/δNEr5j)・(
θ/δWsji)Σ(Wsji・O(s −1)i)=
-rt-CδEpz/δNETsj)○C5-t)i=
−η・δphantom・0(s−1)i
, ...(5) However, δsj=δEpz/δNET
Let it be sj.

δsj in Equation (5) differs depending on whether the layer S is an output layer or an intermediate layer, and can be expressed as follows.

When S is the output layer (s=o): δoz=(Tpz−○oz)・f'oz(NEToz
) - (6) When S is the middle layer: δSj = f' sj (NETsj) - Σ(δ(s
+1)k-W(s+1)kj)...(7) However, k represents the unit number of the (s+1) layer.

Further, δsj is a recursive function.

In this way, equation (6) or equation (7) can be converted into equation (5).
By substituting into the equation, the amount of change in the weight of the connection ΔWsj
i is changed from the output layer to the input layer. However, the initial connection weights before learning are set randomly. Furthermore, the output value of the output layer is determined for other input data patterns (forward propagation), and the connection weights are sequentially updated based on the training data pattern for that pattern and the output value just obtained (backpropagation). two-back propagation)
. This is done for all the training data patterns to complete one round of learning.

Learning is performed repeatedly until the error function converges. A generalized expression for the connection weight ΔWsji is shown below.

ΔWsji(n+1)=η・δsj・o(s−1)i+
a ・ΔWsji(n) ... (8) However, n:
Learning frequency η in Equation (8) is called a learning coefficient, and if this value is made smaller, the learning accuracy improves, but the learning speed becomes slower. On the other hand, if it is made too large, there is a high possibility that vibration will occur. This oscillation is often a problem in minimization methods such as steepest descent, and is caused by the fact that the aspect of the function to be minimized has a deep valley shape. In order to avoid this vibration, a method is often used in which inertia is given in the correction direction by the second term of equation (8). This makes it easier for students to proceed in the direction of the mountain stream without having to go back and forth between both sides of the valley. Here, α is called an inertia coefficient. In this example, in most cases η=0.1, α=0.
0 was used. Please refer to FIG. 6d for an overview of the learning process of the three-layer neural network.

By the way, in this embodiment, as shown in Figure 3, the output value is always 1 in the input layer and middle layer of the neural network.
One dummy unit is provided for adjusting the logistic function threshold value.

In the equation (1) above, when one dummy unit is added to the (s-1) layer, the equation (1) is expressed as follows.

NETsj=Σ(Wsji・0(s−t)i)+Ws
j(i+t)O(s t )(i+t )=Σ(W
sji・O(s 1 )i)+Wsj(i+1)・
1...(9) This equation (9) and the above equation (
Referring to equation 2b), it can be seen that Wsj (i+ 1 ) represents a threshold value.

In other words, the threshold value can be updated every time learning is performed.

In fact, we tested the network structures shown in Figures 8a and 8b in addition to Figure 3, but in terms of learning stability and efficiency, the structure shown in Figure 3 was the same. The best results were obtained when the number of units in the intermediate layer was about half the sum of the number of units in the input layer and the number of units in the output layer, which was efficient for learning.

In this example, when data is applied to the input layer, one unit in the output layer outputs a 1, and the other units output a 0. The number of the output unit that outputs 1 is used as the determination result.

A part of the teacher data (performance data) used in the example is shown in Table 2 below.

In the example shown in Table 2, a correct answer rate of 100% was obtained in the judgment of the neural network after learning.

In the above embodiment, the next manufacturing feasibility examination system uses a system that automatically makes judgments based on rule-based knowledge even for product specifications for which there is no matching performance data. The primary manufacturing feasibility examination system may be replaced by a system that includes a database in which track record data is registered and detects by searching whether or not completely matching track record data exists.

In addition, in FIG. 3, only a part of the network connection relationship is shown to make the drawing easier to read. In reality, all units of the input layer are each coupled to all units of the hidden layer, and all units of the hidden layer are each coupled to all units of the output layer.

Also in FIG. 8a, all units of the intermediate layer are actually coupled to all units of the output layer. Chapter 8b
In the figure, in reality, all units of the input layer are coupled to all units of the intermediate layer l, and all units of the intermediate layer 1 are connected to the intermediate layer M2.
All units of the intermediate layer 2 are connected to all units of the output layer.

In addition to the N-layer structure of the embodiment (see FIG. 7a), the network structure may also be an interconnected network structure as shown in FIG. 7b.

[Effects of the Invention] As described above, according to the present invention, even if there is no past performance data that matches the product specifications to be determined,
It is possible to automatically determine whether manufacturing is possible without human intervention. Moreover, network means (30)
Since this method is used, relatively reliable judgment results can be obtained immediately even under conditions where there is little actual data. In addition, since multiple independent network methods are used and a selection means is provided to select the network method to be used from among them, the network structure is simplified and the learning time and efficiency are significantly reduced. Achieves desired results.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an apparatus according to an embodiment, and FIG. 2 is a block diagram showing the apparatus shown in FIG. 1 in terms of functional connections. Figure 3 is a block diagram showing the configuration of the network selected from the neural network group based on the condition N003 of the rule in Figure 5.6 Figure 4 shows the range of input data and the network to which it is applied. This is a map showing the correspondence with the unit numbers of the input layer. FIG. 5 is a map showing the correspondence between input data conditions and selected networks. Figures 6a, 6b, 6c and 6d are block diagrams showing the operation of a portion or the entire neural network. FIGS. 7a and 7b are block diagrams showing an example of a network structure, and FIG. 7c is a graph showing a sigmoid function. FIGS. 8a and 8b are block diagrams showing modified examples of the neural network. FIG. 9 is a flowchart showing an outline of the processing content of the system of the embodiment. 10: Knowledge base 11: Neural network distribution rules (selection means) 2
0: Inference engine 30: Neural network group (network means)
40: Database 100 Secondary manufacturing feasibility examination system 200: Secondary manufacturing feasibility examination system CPU: Host computer ES: Expert system I (Yl=alXIshia2X2+a3X3) (y2-b+
x+ +bzxz +b3x3〕埒4(s=+) Takashi 6c Figure 1 gift (s=2) Genjo (s=o=3) Voice 6b Zuni 1 mark pattern Voice 8a Figure voice 8b Inichi-Ts-ichi Nii- □TiV-J voice 78 figure voice 7b month 7C figure east figure procedure Honnan Seisho (self-published) Display of 1990 case 1990 patent application No. 161789 2, name of invention Determination of whether or not steel products can be manufactured Device 3: Relationship with the amended person case Patent applicant address: 2-6-3 Otemachi, Chiyoda-ku, Tokyo Name
(665) Nippon Steel Corporation Representative Zensaku Shiraki 4, Agent Address: 103 Phone: 03-864-6052 2-27-6-5 Higashi Nihonbashi, Chuo-ku, Tokyo, Invention of the Specification Subject to Amendment Detailed explanation column and Figure 6b (y1=alXI+a2X2+a3X1) [y2Iwb
1x1fb2x20b3x3] 6. Contents of amendment (1) The following pages and lines in the specification will be corrected as "errors". Correct the "Test site" in the item name to "Collection site" in the content. (3) As shown in red on the attached drawing, the number of one of the figures is corrected to Figure 6d. There are two Figures 6b and 7, catalog drawings of attached documents (Figure 6b)... 1 leaf voice 6 valley diagrams part 7

Claims (2)

[Claims] (1) A network having a plurality of input terminals and at least one output terminal and including means for holding a plurality of weighting coefficients for determining the correlation between the signal level applied to the plurality of input terminals and the level of the output terminal. a means for inputting at least size information and material condition information regarding the specifications of the target steel product; a means for identifying to which of a plurality of predetermined groups the contents of the input specifications belong; Selection means for selecting one of the plurality of network means according to the identification result; means for learning weighting coefficients of the network means based on past determination results; A device for determining whether or not a steel product can be manufactured, comprising: means for outputting the output of the network means as a determination result. (2) A primary judgment means that has a database in which past results are accumulated and inputs information regarding the specifications of the target steel product and outputs judgment results therefor; and when the primary judgment means requests a secondary judgment. , a secondary determination means for inputting information on the specifications and outputting a secondary determination result; and the secondary determination means has a plurality of input terminals and at least one
network means having one output terminal and comprising means for holding a plurality of weighting coefficients for determining the correlation between the signal level applied to the plurality of input terminals and the level of the output terminal;
and means for learning weighting coefficients of the network means based on past determination results.