JPH03265064A - Method and device for simulation and production control supporting device and cad device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a simulation method, a device thereof, and a device applying the same, and particularly to a simulation method under various conditions by storing simulation results in a neural network. Simulation method and device, production management support device, and CAD that have suitable means for immediately obtaining results and supporting discovery of optimal conditions, etc.
Regarding equipment.

[Conventional technology]

As described in Japanese Unexamined Patent Publication No. 1217537, in conventional simulation devices, if the user inputs simulation evaluation criteria and a trial-and-error strategy in advance, the simulation results are evaluated and the simulation condition data is changed according to the trial-and-error strategy. Then, the next simulation was repeated and the results were displayed.

In addition, the conventional product price estimation device is
As described in the publication, it is equipped with a file that records still images of product catalogs and information about the products, and when a customer selects a product, it displays the image and information of that product.
Once the quotation conditions were presented, the quotation was output by searching for a standard delivery date and price list.

[Problem to be solved by the invention]

When performing complex and large-scale simulations, the conventional simulation devices described above have problems in that it takes time to obtain simulation results, forcing the operator to wait in front of the terminal. There is a problem in that setting the error strategy requires a high level of knowledge and insight on the part of the operator and is time consuming.

Further, the above-mentioned conventional product price estimating device can only be applied to estimating products whose specifications have been determined in advance to a limited number of types, and it is difficult to estimate delivery dates in consideration of the status of the production line. However, in recent years, customers' needs for products have become more diverse, and there is a strong demand for orders for each product tailored to the customer's preferences. However, in the conventional technology, it was obtained as a result of product design and process design by the designer, and it was impossible to quickly show it to the customer.
There was a problem in that the only way to make an estimate that took into account the status of the production line was to rely on the intuition of the sales person.

The purpose of the present invention is to solve the above-mentioned problems of the prior art, and to instantly obtain simulation results under new input conditions by having a neural network learn in advance the results of various changes in simulation input conditions. In addition, by obtaining an index of the magnitude of the influence that each item of simulation input conditions has on simulation results, it can be used as a guideline when adjusting simulation input conditions, and furthermore, it can be used as a guideline when adjusting simulation input conditions. An object of the present invention is to provide a simulation method and apparatus having an input means that can easily manipulate a plurality of input conditions for simulation at the same time.

Another object of the present invention is to provide a production management support device that has a function of supporting measures such as shortening delivery times (completion times) by adjusting production line conditions and job input conditions by applying the present simulation device. In addition, we have added a function that can provide highly accurate delivery estimates for product specifications that differ from customer to customer as soon as the product specifications are input, thereby assisting customers in changing specifications and making decisions about placing orders. We have decided to provide a CAD device that can be used as a delivery date estimating device, and also has a function that supports obtaining target performance by adjusting design parameters when designing a product by applying this simulation device. be.

[Means to solve the problem]

In order to achieve the above object, the simulation method and apparatus of the present invention combine a conventional simulator with a neural network having learning ability according to the present invention,
By obtaining the results of various simulation input conditions in advance, having the neural network learn the relationship between the input conditions and the simulation results, and using the neural network's interpolation function,
Simulation results based on input conditions that have not been learned can be obtained immediately, allowing operators to evaluate and evaluate simulation results through interactive operations.
This made countermeasures possible.

In addition, in order to obtain a guideline for adjusting the input conditions of the simulation, it is necessary to use the fact that the magnitude of the influence that changes in the individual inputs of the neural network have on the output can be obtained by using the error backpropagation algorithm for the neural network. Accordingly, a factor extraction method is used when adjusting input conditions.

Furthermore, in order to be able to easily operate multiple input conditions for the simulation at the same time, by providing input means such as multiple dials, the multiple input conditions for the simulation will change one after another when these are operated at the same time. Changes in simulation results are displayed in real time.

In order to achieve the other objects mentioned above, the production management support device of the present invention supports adjustment of production line conditions, job input conditions, etc. by applying the present simulation device to a production line simulator. be.

In addition, in order to estimate delivery dates that take into account not only specifications but also line conditions in a make-to-order production system as a delivery date estimation device, the neural After the network is mostly trained using simulation data, learning is performed using actual line data to obtain a more realistic delivery estimate.

Furthermore, the CAD device of the present invention allows the input conditions of the product performance simulator to be varied in various ways by using the present simulation device as a product performance simulator in order to adjust design parameters during product design and obtain target performance. Obtain the results in advance and let the neural network learn the relationship between the design parameters and the performance index obtained as the simulation result, and use the interpolation function of the neural network to calculate the original performance of the design parameters that have not been learned. The index is immediately obtained, allowing the operator to interactively adjust the design parameters.

C Effect] In the above simulation device, when the neural network learns the relationships between various simulation input conditions and simulation results, since the neural network has an interpolation function, it can calculate the simulation results for simulation input conditions that have not been learned. Since it is possible to obtain an approximate value, even when simulating a complex model that takes a long time to perform a single simulation, the neural network can learn and accumulate the results of simulations performed under various conditions in advance. If you do this, the simulation results under new conditions can be immediately obtained, allowing the operator to interactively evaluate the simulation results and take measures such as changing the simulation conditions and making adjustments. becomes possible.

In addition, along with the error backpropagation algorithm, which is a learning algorithm for neural networks, it is possible to obtain the magnitude of the influence that changes in individual inputs of the neural network have on the output. It is possible to extract factors to provide guidelines for adjusting input conditions by looking at the data.

Furthermore, by assigning a dial, mouse, etc. to the simulation input conditions to be adjusted selected by the factor extraction method, you can try changing multiple input conditions at the same time and display the results in real time, making it possible to create an interactive experience. Input conditions can be adjusted.

The above production management support device applies this simulation device to a production line simulator so that training data for the neural network can be obtained from both the production line simulator and the actual production line, so that the production line can start operating. If the neural network is mostly trained using a simulator beforehand, and then trained using actual data, it will be possible to adjust line conditions and job input conditions from the beginning of the production line's operation. By configuring a delivery date estimation device, it is possible to shorten the delivery date (completion time), and to make it possible to estimate the delivery date from the beginning of the production line's operation, and then to be able to estimate the delivery date more accurately as the learning of actual data progresses. Become.

Furthermore, the above-mentioned CAD device applies this simulation device to a product performance simulator and allows the neural network to learn the relationship between the design parameters and the obtained performance index. Since all performance indicators can be obtained instantly, the operator can interactively adjust design parameters to achieve optimal design.

〔Example〕

An embodiment of the present invention will be described below with reference to FIGS. 1 to 9.

FIG. 1 is an overall configuration diagram showing an embodiment of a simulation method and apparatus according to the present invention.

In FIG. 1, a simulator 1 performs simulations under various simulation conditions generated by a simulation condition generating means 2 and generates simulation results. The learning means 3 obtains data pairs of simulation conditions and simulation results, and performs learning so that the input-output relationship of the neural network 4 is the same as in the simulation 1. As a result, simulation results under arbitrary simulation conditions can be immediately obtained by the neural network 4 without operating the simulator 1, so the evaluation/countermeasure means 5 uses this to evaluate the simulation results. , measures are taken to make the simulation results desirable by adjusting the simulation conditions. The control means 7 controls the entire simulation apparatus, and in addition to controlling the operation of each part and passing data to and from the storage device 8 and the input/output device 6, the evaluation/countermeasure means 5 controls the simulation conditions. The adjusted conditions are sent to the simulation condition generating means 2, and the simulator 1 is made to perform a simulation to check the adjustment results, and the neural network 4 is made to perform a more detailed simulation near the simulation conditions.
This will help you make even better adjustments by having your body learn from it. The storage device 8 stores simulation conditions and simulation results, stores internal weight parameters of the neural network 4, and holds the operating program of the simulator 1. The input/output device 6 is equipped with a dial, mouse, keyboard, etc., and has a graphic display function that allows the operator to interactively evaluate and take measures for the simulation using the evaluation and countermeasure means 5, and also controls the entire simulation device. conduct.

With this configuration, it is possible to immediately obtain simulation results under arbitrary simulation conditions and to interactively perform evaluation and countermeasures.

FIG. 2 is a conceptual diagram of the learning and interpolation functions of the neural network 4 of FIG. 1. As shown in Figure 2, the neural network 4 learns the simulation results of the simulator 1 under various simulation conditions, and thereby performs the same input/output operations. Even when the neural network 4 is given simulation conditions that did not exist, a return simulation result can be obtained by the interpolation function of the neural network 4. For convenience, FIG. 2 shows a case where there is only one type of simulation condition and one type of simulation result, but in general, the same learning function and interpolation function are provided even when there are multiple simulation conditions and simulation results.

FIG. 3 is a diagram showing the principle of the internal structure of the neural network 4 shown in FIG. 1. Regarding neural network theory, see Rummelhart, D.E., McClelland, J.L., and Z.P.D.P. Research Group, 1986 Parallel Distributed Processing: MIT Press (
Rumelbart, D. E., 1McCella
nd,, Jru, and The PDP Re5ea
rch group.

1986, Parallel Distribution
dProcessing (formerly known as TPress), FIG. 3 shows a principle diagram of a hierarchical network as an example of a neural network model applied to this embodiment.

Regarding the internal operation principle of the neural network 4, as shown in Figure 3, it has a structure in which m layers each consisting of a large number of units are stacked, and the units included in adjacent layers are connected to each other. The connection is weighted, and the signal input from the first layer, which is the input layer, is transmitted to the next layer according to this weight, and the result is output from the mth layer, which is the output layer. Since the operation of the neural network 4 is determined by this weight, by changing this weight through learning, the input/output operation becomes equal to the relationship between the input signal ii and the teacher signal yi given to the output layer.

Here, the operation and signal flow of each unit included in adjacent layers of the neural network 4 will be explained in detail. Let the input of the 3rd unit of the 1st layer be ik, the output be Qk, and the connection weight parameter between the 1st unit of the 11th layer and the 3rd unit of the 2nd layer be WKil'. The relationship between the output Q Kil of the 1st unit of 11% and the output 0 of the 3rd unit of the 2nd layer is expressed by the following equation (1) and (2
) is expressed as the formula.

0 kidney = f (i day (1) 1y-Σ- to 1
1 wise 0 il (2) However, f (X) is a sigmoid function. According to this relationship, the input i of the first layer
The neural network 4 operates by transmitting signals one after another from the first to the m-th layer output OT.

Next, the learning operation by the learning means 3 of the neural network 4 will be explained. Based on the deviation of the actual output value OT from the teacher signal yj, which is the output value to be learned, this deviation value is propagated backward from the m-th layer to the first layer. Assuming that the error backpropagation value at this time is 17 for the third unit of the first layer, it is expressed as the following equations (3) and (4).

d 'f' = (0'i' Y ff' ) f
' (i 'j) (3) d kidney = (ΣW kidney - dKi□1) V (i!i') (where ≠ m
(4) At this time, learning can be realized by repeatedly changing the value of W"il' according to the product of Q Nil and d17.

For this reason, this learning method is called an error backpropagation algorithm.

As mentioned above, the operation of the neural network 4 is determined by the internal connection weight parameters, and its learning is realized by adjusting the connection weight parameters.If the connection weight parameters are memorized, the simulator 1
Even in the case where another simulation is performed while the operation program of 1 is withheld, the connection weight parameters of the neural network 4 may be loaded from the storage device 8 and used accordingly.

Next, we will perform an evaluation using the above error backpropagation algorithm.
The factor extraction method using countermeasure means 5 will be explained. This method uses error backpropagation values to calculate the degree to which the output QT of the focused output layer unit changes due to changes in the inputs 11 of the input layer units. This is called a factor extraction method because it is possible to determine which input il is effective to change and how much to change when it is desired to bring the output OT closer to the target. In this method, the following equation (5) is derived by using equations (1) to (4) above.

Y = t(OT +”'+ On* O'i' -1,
O=, +・+, OA)) (5) In other words, when error signal 1 is given only to the focused conversion unit, the error backpropagation value d1 propagated to the input unit is the sensitivity of the input to the output. is equal to

Therefore, since inputs with large dl are the main cause, it is sufficient to select several types of inputs and make adjustments to these inputs.

FIG. 4 is an overall configuration diagram showing an embodiment of a production management support device in which the simulation device of FIG. 1 according to the present invention is applied to a production line simulator.

Further, FIGS. 5 and 6 are display example diagrams by the evaluation/countermeasure means 5 shown in FIG. 4, respectively. In FIG. 4, the simulation condition generating means 2 is a line simulator 101.
Randomly generate input conditions for line simulator 1.
01 performs a simulation under these conditions to generate learning data for the neural network 4. The neural network 4 is also capable of learning production performance data of the actual line IO, and both data are selected by the switch 9 and sent to the learning means 3, and the learning means 3 performs learning on the neural network 4. will be held. Since the neural network 4 that has undergone learning here can obtain the original simulation results for arbitrary line conditions and job input conditions, that is, predicted manufacturing performance values, the evaluation/countermeasure means 5 uses this to determine the line conditions and job input conditions. Evaluation and countermeasures are taken to make the simulation results, that is, the predicted manufacturing performance values, favorable by adjusting input conditions. The input/output device 6 is equipped with a graphic display function and a dial, mouse, keyboard, etc. for the operator to use the evaluation/measure means 5 to interactively adjust line conditions and job submission conditions and take measures. . The control means 7 controls the entire production management support device, and the storage bag W8 stores necessary data and holds operating programs.

Here, the internal configuration of the evaluation/countermeasure means 5 will be explained.

First, the bottleneck factor extraction means 501 utilizes the fact that the magnitude of the influence that changes in the individual inputs of the neural network 4 have on the output can be obtained through the error backpropagation algorithm of the neural network 4. Inputs that have a large influence are extracted, and the extracted inputs are sent to the countermeasure means 502 along with their sensitivity. The countermeasure means 502 performs the following processing in cooperation with the input/output device 6. For example, as shown in FIG. 5, each bottleneck factor is assigned to a mouse or a dial, and as the operator simultaneously adjusts a plurality of factors, the simulation results, that is, the changes in the predicted production results are graphically output. As another example, as shown in FIG. 6, the operator selects one or two types of bottleneck factors, and the neural network 4 obtains simulation results when the factors are continuously changed. Plot on a 2D or 3D graph. The line conditions and job submission conditions adjusted by the evaluation/countermeasure means 5 are sent to the simulation condition generation means 2 as necessary, and either a detailed simulation is performed and the results are confirmed, or the conditions are changed in the vicinity of the conditions. By making the neural network 4 learn the results of the simulation, more accurate evaluation and countermeasures can be taken. In addition, the delivery date answering means 503 responds to inquiries about delivery dates for new orders from the sales department, etc. by using the neural network 4 to immediately calculate and respond to delivery date estimates that take into account order specifications and line conditions, and If the product is not satisfied, it has the function of working together with the countermeasure means 502 to consider ways to shorten the delivery period by changing the specifications or taking countermeasures to the line without significantly changing the customer's requirements.

With the production management support device configured as described above, the simulation results of the production line simulator can be trained in advance in the neural network so that they can be immediately retrieved and used, and the data to be learned can be switched to the actual production data of the actual line. This makes it possible to obtain more accurate results from the neural network, which can also be used as a delivery date estimating device with a delivery date response function. These functions are summarized in flowcharts in Figure 7 (a), (b), (C),
Shown in (d).

7(a) to 7(d) are processing flowcharts of the production management support apparatus of FIG. 4. FIG. 7(a) is a schematic processing flowchart, and FIG. 7(b) is an evaluation and processing flowchart of FIG. 7(a).
Processing flowchart of countermeasure processing 1 (processing 720), seventh
Figure (C) shows the evaluation/countermeasure process 2 (process 73) in Figure 7 (a).
0) processing flowchart, FIG. 7(d) is similar to FIG. 7(a).
) is a processing flowchart of the delivery date response processing (processing 740). First, in FIG. 7(a), the switch 9 selects learning data depending on whether the mode is actual line learning mode or not (process 701). When learning the simulation results of the line simulator 101, simulation conditions such as line conditions and job submission conditions are randomly generated by the simulation condition generating means 2 (process 70).
2) Perform line simulation (process 703)
. Furthermore, when learning data from an actual line, the actual production line 10 is observed to obtain learning data (processing 704).
. Next, in either case, learning pattern data is generated by the learning means 3 (step f1705), and the neural network 4 is caused to perform learning (step 706). Here, the above operations are repeated until an evaluation/measure request is issued by the operator (process 707).
The countermeasure processing mode is entered and the operator selects the evaluation/countermeasure processing (process 708). Next evaluation/countermeasure process 1 (
The details of the processing 720), the evaluation/countermeasure processing 2 (process 730), and the delivery date response processing (process 740) will be explained later using examples, but these processes are similar to the evaluation/countermeasure means 5.
By adjusting the simulation conditions, that is, the line conditions and job input conditions, a desirable predicted manufacturing performance value can be obtained. The line simulator 101 is operated again according to the line conditions and job input conditions obtained through these processes, and it is confirmed that a desired predicted manufacturing performance value can be obtained (process 709). When it is desired to carry out more detailed evaluation measures (processing 710), the control means 7 sets the simulation condition generating means 2 to generate simulation conditions close to the line conditions and job input conditions obtained up to processing 709. (Process 711), the process returns to the start and repeats the same process from Process 701.

Next, the detailed flow of the evaluation and countermeasure processing shown in FIG. 7(a) will be explained with reference to FIGS. 7(b), (C), and (d). First, in FIG. 7(b), evaluation/countermeasure processing 1 (processing 72
0) performs the following processing. When the operator selects the line conditions and job input conditions of the simulation conditions to be changed (processing 721), the countermeasure means 502 changes the selected conditions one after another (processing 722), and the neural network 4 outputs the manufacturing result as a simulation result. A predicted actual value is obtained (process 723), and the predicted value is plotted in a graph on the input/output device 6 (process 724), and the processes from process 722 onward are continued until the graph is completed (process 725). In addition, in Figure 7 (C), the evaluation
Countermeasure process 2 (process 730) performs the following process. The net factor extraction means 501 extracts several factors and their sensitivities that have a large influence on the predicted manufacturing performance value of interest (731), and the countermeasure means 502 assigns a dial or mouse to each of the factors. (Processing 732
), the operator uses these operations to change the line conditions and job input conditions that have been extracted as bottleneck factors (processing 7).
33) Input the conditions to the neural network 4 to immediately obtain a predicted manufacturing performance value as a simulation result (process 734), and display graphically how the predicted value changes in real time (process 735). The processes from 733 to process 735 are continued until the operator gives an instruction to finish (process 736).

Furthermore, in FIG. 7(d), the delivery date replying process (processing 740) of the delivery date replying means 503 performs the following process.

The current line conditions and the job input conditions corresponding to the new order are set as input to the neural network 4 (process 741), and the delivery date, which is one item of the predicted manufacturing performance value of the output, is obtained (process 742). Process 743) to respond to the input/output device 6, and when there is a request to shorten the delivery time (process 744), further evaluate/countermeasure process 2 (process 730) is carried out to take line countermeasures or to take measures that do not change customer requirements too much. The specifications are changed to shorten the predicted delivery date and then output to the input/output device 6 (processing 743).

FIG. 8 is a detailed diagram of an example of a model that combines the line simulator 101 and neural network 4 shown in FIG. 4. In FIG. 8, the line simulator 101 connects processes according to the flow of parts, the net working time and the flow of parts change depending on the order specification, and a model is assumed in which there is a certain amount of buffer before each process. The simulation condition generating means 2 generates line condition data 201 and job submission condition data 202 corresponding to a new order for the line simulator 101, and generates the line condition data 201.
includes, for example, the work in process, capacity, and parts inventory for each process, and the job submission condition data 202 corresponding to the order is a collection of jobs corresponding to the order expressed, for example, by product series, variants, options, sizes, etc. . The line simulator 101 uses these condition data 201.202.
The above simulation conditions and the following simulation results are generated as learning data, and the neural network 401 learns this learning data. The data obtained as a result of the above simulation includes, for example, the product delivery date (completion time) of job n, and predicted values of lead time and work in progress in a specific process at a certain point in the future.

With the above configuration, the delivery date (completion time) under arbitrary line conditions and job input conditions corresponding to new orders, and predicted values of lead time and work in progress for a specific process at a certain point in the future can be calculated on the line. This can be obtained immediately without running the simulator 101 every time.

FIG. 9 is an overall configuration diagram showing an embodiment of a CAD device having a design optimization support function in which the simulation device of FIG. 1 according to the present invention is applied to a product performance simulator.

In FIG. 9, the design parameter generation means 203 randomly generates input conditions for the product performance simulator 102, and the product performance simulator 102 performs a simulation under the conditions and performs a performance index which is learning data for the neural network 4. occurs. As an example of this, in the case of motor design, the design parameters include the number of poles, coils,
The number of turns of the coil, the shape of the iron core, the shape of the rotor, the thickness of the winding, etc.
performs electromagnetic field simulation, circuit simulation, mechanical simulation, heat conduction simulation, etc., and outputs performance indicators such as motor output, power consumption, impedance, and temperature rise. The learning means 3 obtains the design parameters and the performance index, and causes the neural network 4 to learn the input-output relationship between the two. If such simulation and learning are performed under various design parameters,
Since the performance index based on any design parameter can be immediately obtained by the neural network 4, the evaluation and countermeasure means 5 uses this to cooperate with the CAD device 11 and the input/output device 6 via the control means 7. , evaluate and take measures to determine what kind of design should be used to achieve the target performance.

As an example of the function of this evaluation/countermeasure measure 5, by extracting several types of design parameters that greatly affect the performance index of interest and changing them simultaneously using a dial, mouse, etc., the performance index changes accordingly. Graphically display the situation in real time. The results of adjusting the design parameters in this way are immediately reflected in the product model of the CAD device, and are also fed back to the design parameter generation means 203 by the control means 7 to perform more detailed simulations and confirm performance. , further optimizing the design.

With the above configuration, a CAD device having a function of interactively optimizing design parameters is realized.

In addition, in the simulation apparatus of the above embodiment, it is also possible to use a pattern conversion means having a learning function instead of the neural network.

〔Effect of the invention〕

Since the present invention is configured as described above, it produces the effects described below.

By having a neural network learn the results of various changes in the input conditions of the simulation device, simulation results under arbitrary input conditions can be immediately obtained, and simulation results can be evaluated and countermeasures can be taken through interactive operations. is possible.

In addition, effective adjustment of simulation conditions is possible using a factor extraction method that uses a neural network learning algorithm and a man-machine interface that uses a dial, mouse, etc., and a graphic display.

Furthermore, by applying the simulation device to a production line simulator, it can be used as a production management support device and a delivery date estimation device, and by using not only simulation results but also actual line data as learning data, it is even closer to reality. You will get results.

Further, by applying the simulation device to a product performance simulator, a CAD device that supports designing a product with target performance by adjusting product parameters through interactive operations can be realized.

[Brief explanation of drawings]

FIG. 1 is an overall configuration diagram showing an embodiment of the simulation method and apparatus according to the present invention, FIG. 2 is a conceptual diagram of the function of the neural network in FIG. 1, and FIG. 3 is a principle diagram of the internal structure. FIG. 4 is an overall configuration diagram showing one embodiment of the production management support device according to the present invention, FIGS. 5 and 6 are display examples of the evaluation and countermeasure means shown in FIG. 4, and FIGS.
d) is a flowchart of the schematic processing and evaluation/countermeasure processing shown in FIG. 4, FIG. 8 is a detailed diagram of an example of a combination model of the simulator and neural network shown in FIG. 4, and FIG. FIG. DESCRIPTION OF SYMBOLS 1...Simulator, 2...Simulation condition generation means, 3...Learning means, 4...Neural network, 5...Evaluation/countermeasure means, 6...I/O device, 7... Control means, 8...Storage device, 9...Switcher, 10...Actual line, 101...Line simulator, 501...Neck factor extraction means, 502.
・Measures for countermeasures, 503 ・Means for answering delivery date, 201 ・・
- Line condition data, 202... Job submission condition data, 102... Product performance simulator, 203...
Design parameter generation means, 11... CAD device.

Claims (1)

[Claims] 1. A simulator and a neural network are combined, the input conditions of the simulator are changed and the results of the simulation are learned by the neural network, and the simulation results for the new input conditions are learned using the neural network. A simulation method characterized by prediction. 2. The simulation method according to claim 1, wherein adjustments to a plurality of input conditions of the simulator are made simultaneously. 3. Select inputs that have a large effect on the output of the neural network by using the fact that the magnitude of the effect that changes in individual inputs of the neural network have on the output can be obtained by using the error backpropagation algorithm for the neural network. 3. The simulation method according to claim 2, wherein items for adjusting input conditions of the simulator are automatically set using a factor extraction method. 4. A simulator, an input condition generating means for generating the input conditions, a neural network for learning the relationship between the input conditions of the simulator and the simulation results, a learning means for this purpose, and a method for generating the simulation results using the learned neural network. A simulation device comprising means for evaluating and taking countermeasures. 5. The means for performing evaluation and countermeasures feeds back the conditions obtained as a result of the evaluation and countermeasures to the input condition generation means, performs more detailed simulations in the vicinity, and performs more accurate evaluations and countermeasures. The simulation device according to claim 4, characterized in that: 6. Simulation using a simulator, an input condition generation means for generating input conditions, a pattern conversion means for learning the relationship between the input conditions of the simulator and simulation results, a learning means for this purpose, and the learned pattern conversion means A simulation device characterized by comprising means for evaluating results and taking countermeasures. 7. The simulation device according to claim 4 is applied to a production line simulator, and a neural network is made to learn the relationship between line conditions, job input conditions, etc. of input conditions of the simulator and lead time, etc. of simulation results, and is used for production management. A production management support device characterized by its use. 8. The production management support device according to claim 7, wherein after the learning of the neural network is mostly performed using data generated by a simulator, further learning is performed by switching to observation data of an actual production line. 9. The production management support device according to claim 7 or 8, wherein the production management support device functions as a delivery date estimating device by including a delivery date answering device in the means for evaluating and taking countermeasures. 10. The simulation device according to claim 4 is applied to a product performance simulator, and a neural network is made to learn the relationship between the design parameters of the input conditions of the simulator and the performance index of the simulation result, and the results are utilized for optimal design of the product. CAD equipment.