Learning device, temperature history prediction device, welding system and program

Abstract

To provide a learning device that can highly accurately predict temperature history of a laminate at high speed, a temperature history prediction device, a welding system and a program.SOLUTION: A learning device 200 includes: a temperature distribution acquisition section 33 that determines first temperature distribution indicating temperatures of a plurality of unit elements of a laminate at a specific time and second temperature distribution indicating temperatures of a plurality of unit elements at a time when a predetermined time passes from the specific time; and a learning section 35 that performs machine learning by associating relation between the first temperature distribution and the second temperature distribution determined by the temperature distribution acquisition section 33 with the specified time to generate a prediction model.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present invention relates to a learning device, a temperature history prediction device, a welding system, and a program. [Background technology]

[0002] In recent years, there has been an increasing need for parts manufacturing through additive manufacturing using 3D printers, and research and development is underway to commercialize this manufacturing using metallic materials. For example, Patent Document 1 proposes a thermal fluid analysis method for determining the temperature history of a molded object (layered body) obtained by moving a heat source to form a bead on a substrate with high accuracy and in a short amount of time required for analysis. In the technology of Patent Document 1, the thermal fluid analysis is performed not in a Lagrangian coordinate system but in an Eulerian coordinate system in which the heat source is fixed and the substrate is moved relatively to form the bead. This makes it possible to reduce the number of areas with small element sizes in the element division model used for the analysis, thereby shortening the calculation time. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-44541 Summary of the Invention [Problem to be solved by the invention]

[0004] The finite element method (FEM) is usually used in temperature prediction using conventional analysis methods such as that described in Patent Document 1. In this case, calculations are performed according to the laws of physics for all meshes into which the shape of the laminate is divided, and therefore the larger the size of the laminate, the more massive the calculations required, making it difficult to speed up predictions.

[0005] SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a learning device, a temperature history prediction device, a welding system, and a program that can predict the temperature history of a laminate with high accuracy and high speed. [Means for solving the problem]

[0006] The present invention comprises the following configurations. (1) A learning device that generates, by machine learning, a prediction model that predicts the temperature history of a laminate during the formation of the laminate by each unit element obtained by dividing the shape of the laminate, when the laminate is formed by a weld bead formed by melting and solidifying a filler material by moving a heat source along a predetermined path, a temperature distribution acquiring unit that acquires a first temperature distribution representing temperatures of the plurality of unit elements at a specific time of the laminated body and a second temperature distribution representing temperatures of the plurality of unit elements at a time when a specified time has elapsed since the specific time; a learning unit that performs machine learning on the relationship between the first temperature distribution and the second temperature distribution obtained by the temperature distribution acquisition unit, by associating the relationship with the specified time, and generates the prediction model; A learning device comprising: (2) The prediction model generated by the learning device according to (1); an input data acquisition unit that acquires input data including information on a temperature distribution of the plurality of unit elements at an initial time when the laminated body is modeled; a prediction unit that obtains a predicted temperature distribution by predicting a temperature distribution after a specified time has elapsed from the state of the temperature distribution included in the input data using the prediction model; a prediction control unit that inputs information about the predicted temperature distribution as the input data back into the prediction unit and further obtains a predicted temperature distribution after the specified time has elapsed; A temperature history prediction device comprising: (3) The temperature history prediction device according to (2), a welding device that builds the laminate based on building conditions that are determined using information on the temperature distribution predicted by the temperature history prediction device; and A welding system comprising: (4) A program for generating, by machine learning, a prediction model that predicts the temperature history of a laminate during the formation of the laminate by each unit element obtained by dividing the shape of the laminate, when the laminate is formed by a weld bead formed by melting and solidifying a filler material by moving a heat source along a predetermined path, On the computer, a function of calculating a first temperature distribution representing the temperatures of the plurality of unit elements at a specific time of the laminate, and a second temperature distribution representing the temperatures of the plurality of unit elements at a time when a specified time has elapsed since the specific time; a function of performing machine learning on the relationship between the first temperature distribution and the second temperature distribution and the specified time to generate the prediction model; A program to achieve this. [Effects of the Invention]

[0007] According to the present invention, the temperature history of a laminate can be predicted with high accuracy and high speed. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is an explanatory diagram showing how a filler metal is melted and solidified to form a weld bead. [Figure 2] FIG. 2 is a schematic perspective view showing an example of a laminate. [Figure 3A] FIG. 3A is an explanatory diagram showing an example of a procedure for determining a path for forming a laminated body. [Figure 3B] FIG. 3B is an explanatory diagram showing an example of a procedure for determining a path for forming a laminated body. [Figure 4] FIG. 4 is a schematic diagram showing the overall configuration of the welding device. [Figure 5] FIG. 5 is a block diagram of the learning device. [Figure 6] FIG. 6 is a flowchart showing the procedure for machine learning a temperature history prediction model by the learning device. [Figure 7] FIG. 7 is an explanatory diagram showing the relationship between the temperature distribution at a specific time and the temperature distribution at a time after a specified time has elapsed for a plurality of times. [Figure 8] FIG. 8 is an explanatory diagram showing a specific example of temperature distribution information to be input to the learning unit. [Figure 9] FIG. 9 is a block diagram of the temperature history prediction device. [Figure 10] FIG. 10 is a flowchart showing the procedure for predicting the temperature history of a laminate when the laminate is manufactured. [Figure 11] FIG. 11 is an explanatory diagram showing the formation of a weld bead and the temperature distribution of each unit element in a time series. [Figure 12] FIG. 12 is an explanatory diagram showing the formation of a weld bead and the temperature distribution of each unit element in a time series. [Figure 13] FIG. 13 is an explanatory diagram showing how the element size of a unit element is changed. [Figure 14A] FIG. 14A is an explanatory diagram showing how the temperature changes when the predicted temperature is kept constant for a specified time until the predicted temperature approaches the reference temperature. [Figure 14B] FIG. 14B is an explanatory diagram showing how the temperature changes when the specified time is changed before the predicted temperature approaches the reference temperature. [Figure 15A] FIG. 15A is an explanatory diagram showing a first temperature distribution and a second temperature distribution used in machine learning. [Figure 15B] FIG. 15B is an explanatory diagram showing the data sets of the first temperature distribution and the second temperature distribution used in machine learning. [Figure 16] FIG. 16 is an explanatory diagram showing a state in which a plurality of unit elements obtained by dividing the shape of a laminate are made to resemble the shape of a weld bead. [Figure 17] FIG. 17 is a model diagram showing the shape of the laminate used in Verification Example 1. As shown in FIG. [Figure 18] FIG. 18 is a graph showing test results showing the predicted temperature values ​​at positions P1, P2, P3, and P4 on the two-dimensional cross section shown in FIG. 17 and the time change in the temperature determined by numerical analysis. [Figure 19] FIG. 19 is a model diagram showing the shape of the laminate used in Verification Example 2. [Figure 20]FIG. 20 is a cross-sectional view showing the predicted position of the temperature distribution in the model shown in FIG. [Figure 21] FIG. 21 is a graph showing test results showing the predicted temperature values ​​at positions P1, P2, P3, and P4 on the two-dimensional cross section shown in FIG. 19 and the time change in the temperature determined by numerical analysis. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments of a learning device and a temperature history prediction device according to the present invention will be described in detail with reference to the drawings. In this embodiment, when a laminated body having a desired shape is manufactured by stacking weld beads, the temperature history during the manufacturing process of the laminated body is predicted. This temperature history prediction is performed using a prediction model generated by machine learning.

[0010] 1 is an explanatory diagram showing how a filler metal is melted and solidified to form a weld bead B. To form the laminate, a torch 11, which generates an arc Ak as a heat source at its tip, is moved along predetermined paths PS1, PS2, PS3, ..., while continuously supplying a filler metal (not shown) to melt and solidify it with the arc Ak. As a result, a weld bead B, which is a molten solidified body of the filler metal, is formed on a base 13, and the weld beads B are sequentially stacked to obtain the laminate.

[0011] Fig. 2 is a schematic perspective view showing an example of the laminate W. For example, to form a rectangular parallelepiped structure, a plurality of weld beads B as shown in Fig. 2 are laminated, and the laminated laminate W is machined as necessary to form a structure of the desired shape. In this case, the path (movement trajectory of torch 11) for forming the weld beads B is determined by creating a molding plan based on shape data (CAD data, etc.) that represents the shape of the desired structure and the type and specifications of the welding equipment to be used.

[0012] 3A and 3B are explanatory diagrams showing an example of a procedure for determining a path for building a laminate. Specifically, as shown in FIG. 3A, the path is determined by slicing the shape of the laminate W given by the shape data to a predetermined thickness. Then, as shown in FIG. 3B, the path for forming a weld bead and the welding conditions are determined so that the shape of each sliced ​​layer Ly1, Ly2, Ly3, Ly4 is filled with a bead shape model BM of a predetermined width. The algorithm for creating such a building plan is not particularly limited and may be a conventionally known one.

[0013] The temperature history during the formation of the laminate is used to improve the manufacturing quality of the laminate W, such as by evaluating the thermal distortion that occurs in the laminate when the parts that became hot during the formation of the weld bead B are cooled, and by creating a welding plan that suppresses thermal distortion.

[0014] <Welding equipment configuration> Here, a welding device used to form the laminate W will be described. FIG. 4 is a schematic diagram showing the overall configuration of the welding device.

[0015] The welding apparatus 100 includes a shaping control device 15, a manipulator 17, a filler metal supply device 19, a manipulator control device 21, and a heat source control device 23.

[0016] The manipulator control device 21 controls the manipulator 17 and the heat source control device 23. A controller (not shown) is connected to the manipulator control device 21, and an operator can instruct any operation of the manipulator control device 21 via the controller.

[0017] The manipulator 17 is, for example, an articulated robot, and a torch 11 attached to the tip shaft supports a filler material (welding wire) M so that the filler material M can be continuously supplied. The torch 11 holds the filler material M protruding from the tip. The position and posture of the torch 11 can be set arbitrarily in three dimensions within the range of the degrees of freedom of the robot arm constituting the manipulator 17. The manipulator 17 preferably has six or more degrees of freedom, and is preferably one that can arbitrarily change the axial direction of the heat source at the tip. The manipulator 17 may take various forms, such as a four- or more-axis articulated robot as shown in FIG. 4, or a robot equipped with angle adjustment mechanisms on two or more orthogonal axes.

[0018] The torch 11 has a shield nozzle (not shown), through which shielding gas is supplied. The shielding gas blocks the atmosphere and prevents oxidation and nitridation of the molten metal during welding, thereby suppressing welding defects. The arc welding method used in this configuration may be either a consumable electrode type such as shielded metal arc welding or carbon dioxide gas arc welding, or a non-consumable electrode type such as TIG (Tungsten Inert Gas) welding or plasma arc welding, and is appropriately selected depending on the laminate W to be formed. Here, gas metal arc welding will be used as an example. In the case of a consumable electrode type, a contact tip is disposed inside the shield nozzle, and a filler material M to which current is supplied is held by the contact tip. The torch 11 holds the filler material M and generates an arc from the tip of the filler material M in a shielding gas atmosphere.

[0019] The filler material supply device 19 supplies the filler material M toward the torch 11. The filler material supply device 19 includes a reel 19a around which the filler material M is wound, and a feeding mechanism 19b that feeds the filler material M from the reel 19a. The filler material M is fed to the torch 11 by the feeding mechanism 19b while being sent in the forward or reverse direction as needed. The feeding mechanism 19b is not limited to a push type that is arranged on the filler material supply device 19 side and pushes out the filler material M, but may also be a pull type or a push-pull type that is arranged on a robot arm or the like.

[0020] The heat source control device 23 is a welding power source that supplies the power required for welding by the manipulator 17. The heat source control device 23 adjusts the welding current and welding voltage supplied when forming a bead by melting and solidifying the filler material M. In addition, the filler material supply speed of the filler material supply device 19 is adjusted in conjunction with the welding conditions such as the welding current and welding voltage set by the heat source control device 23.

[0021] The heat source for melting the filler metal M is not limited to the arc described above. Other heat sources may be used, such as a heating method that combines an arc and a laser, a heating method that uses plasma, or a heating method that uses an electron beam or a laser. When heating with an electron beam or a laser, the amount of heat can be controlled more precisely, and the state of the formed bead can be maintained more appropriately, contributing to further improving the quality of the laminated structure. The material of the filler metal M is also not particularly limited. The type of filler metal M used may vary depending on the properties of the laminate W, such as mild steel, high-tensile steel, aluminum, aluminum alloy, nickel, or nickel-based alloy.

[0022] The welding apparatus 100 configured as described above operates in accordance with a building program created based on a building plan for the laminated body W. The building program is composed of numerous command codes and is created based on an appropriate algorithm depending on various conditions, such as the shape, material, and heat input of the object to be built. According to this building program, the torch 11 is moved while the supplied filler material M is melted and solidified, and a linear weld bead B, which is a molten solid of the filler material M, is formed on the base 13. That is, the manipulator control device 21 drives the manipulator 17 and the heat source control device 23 based on a predetermined program provided by the building control device 15. In response to commands from the manipulator control device 21, the manipulator 17 moves the torch 11 while melting the filler material M with an arc to form the weld bead B. By sequentially forming and stacking the weld beads B in this manner, a laminated body W having the desired shape is obtained.

[0023] The welding apparatus 100 may be configured as a welding system equipped with a learning device and a temperature history prediction device, which will be described below. In this case, a building plan for a laminate can be created based on an accurately predicted temperature distribution, thereby enabling the building of a high-quality laminate.

[0024] <Learning device> Next, a learning device that generates, by machine learning, a prediction model that predicts the temperature history during the manufacturing of the laminate W manufactured by the welding device 100 will be described. 5 is a block diagram of the learning device 200. The learning device 200 includes an element dividing unit 31, a temperature distribution acquiring unit 33, a learning unit 35, and a simulation unit 37. The simulation unit 37 is used when learning data is obtained by numerical analysis.

[0025] Each of the above units will be described in detail later, but roughly speaking, each unit functions as follows. Shape data representing the shape of the laminate to be built, and information such as welding conditions based on a predetermined building plan for the laminate, are input to the learning device 200. The element dividing unit 31 divides the shape of the laminate W in the shape data into multiple unit elements to generate a shape model. This division into unit elements corresponds to, for example, dividing the shape of the laminate shown in FIG. 3B into multiple bead shape models BM, but it may also be divided into meshes of finer size. Note that if a shape model divided into multiple unit elements is input to the learning device 100, the element dividing unit 31 is unnecessary.

[0026] The temperature distribution acquisition unit 33 acquires a temperature distribution (first temperature distribution) representing the temperature of the plurality of unit elements of the laminate at a specific time, and a temperature distribution (second temperature distribution) representing the temperature of the plurality of unit elements at a time when a specified time Δt has elapsed since the specific time. The specified time Δt can be set to any time, such as 10 seconds. Each temperature distribution may be calculated by the simulation unit 37, or may be calculated by fabricating a test sample and measuring it. The simulation unit 37 calculates the second temperature distribution from information about the initial temperature distribution (first temperature distribution) through numerical analysis, such as thermal fluid analysis using the finite element method (FEM) using the divided plurality of unit elements. The procedure for calculating each temperature distribution through numerical analysis by the simulation unit 37 will be described below. The simulation unit 37 may be configured with dedicated software or circuits, or may be configured to perform simulations by setting various additive manufacturing conditions in general-purpose thermal fluid analysis software (e.g., FLOW-3D (registered trademark) by Flow Science, Inc.).

[0027] The learning unit 35 performs machine learning on the relationship between the first temperature distribution and the second temperature distribution obtained by the temperature distribution acquisition unit 33, associating it with a specified time Δt, and generates a prediction model. In addition to the first temperature distribution and the second temperature distribution, information regarding the movement of the heat source and the generation of the weld bead may also be included in the learning data. As shown in FIG. 1, information regarding the heat source includes the amount of heat input Q supplied per unit time, the direction and speed v of movement of the heat source, etc., and information regarding the generation of the weld bead includes, for example, the bead generation volume V and the bead cross-sectional area S.

[0028] The learning device 200 is configured with hardware using an information processing device such as a PC (Personal Computer). Each function of the learning device 200 is realized by a control unit (not shown) reading and executing a program having a specific function stored in a storage device (not shown). Examples of the storage device include memory such as RAM (Random Access Memory), which is a volatile storage area, and ROM (Read Only Memory), which is a non-volatile storage area, as well as storage such as HDD (Hard Disk Drive) and SSD (Solid State Drive). Examples of the control unit include a processor such as a CPU (Central Processing Unit) or MPU (Micro Processor Unit), or a dedicated circuit.

[0029] <Learning Procedure> FIG. 6 is a flowchart showing the procedure for machine learning a temperature history prediction model by learning device 200. Learning device 200 sets initial conditions, such as inputting a shape model consisting of multiple unit elements and initial conditions such as welding conditions into simulation unit 37 (step 1: S11). Simulation unit 37 obtains the temperatures of the multiple unit elements at a specific time t0 when a layered body is manufactured by forming a weld bead. Temperature distribution acquisition unit 33 obtains the obtained temperature distribution of the unit elements as a first temperature distribution at the specific time t0 (S12). The specific time t0 here may be the time when formation of the weld bead begins, or may be a time during bead formation. In either case, the temperature distribution at the specific time t0 is assumed to be known.

[0030] Next, a second temperature distribution is calculated (S13), which represents the temperatures of the plurality of unit elements at a time (t0+Δt) when a specified time Δt has elapsed since the specific time t0. The second temperature distribution is a temperature distribution obtained after the specified time Δt has elapsed since the first temperature distribution, when heat input from the heat source (arc) for bead formation has propagated to the surrounding unit elements and cooled. The second temperature distribution is analytically calculated by the simulation unit 37 from the temperature information of the first temperature distribution and information such as the physical property values ​​of the shape model. The temperature distribution acquisition unit 33 acquires the second temperature distribution calculated by the simulation unit 37.

[0031] The temperature distribution acquisition unit 33 then outputs the acquired information on the first temperature distribution and second temperature distribution to the learning unit 35 as learning data. The learning unit 35 also receives various information based on the shape data and layering plan information input to the simulation unit 37, such as the amount of heat input supplied from the heat source when forming a weld bead, the direction and speed of movement of the heat source, the volume of the weld bead formed per unit time, status information indicating the presence or absence of a weld bead for each unit element, and size information of the unit element. The learning unit 35 may also learn information about the addition of a weld bead and the amount of heat input associated with the addition of a weld bead. The learning unit 35 performs machine learning on the relationship between the input first temperature distribution and the second temperature distribution, associating it with a specified time Δt (S14). The learning unit 35 may also perform machine learning on the heat source information, bead information, and element size information described above.

[0032] By including information about the heat source in the learning data for learning unit 35, it is possible to more accurately predict the temperature distribution when a weld bead is formed as the heat source moves. Also, by learning status information indicating the presence or absence of a weld bead, it is possible to predict the presence or absence of a weld bead generated as the heat source moves, along with the temperature distribution. Furthermore, by including size information about the unit element in the learning data, it is possible to make predictions at any unit element size (mesh size).

[0033] 7 is an explanatory diagram showing the relationship between the temperature distribution at a specific time (t=t0) and the temperature distribution at a time (t=t0+Δt) after a specified time Δt has elapsed, for multiple times. The learning device 200 sequentially obtains pairs of a first temperature distribution at a specific time (t=t0) shown in FIG. 7 and a second temperature distribution at a time (t=t0+Δt) after a specified time Δt has elapsed, along the path that forms the weld bead. In other words, the temperature distribution acquisition unit 33 acquires information on pairs of the first temperature distribution and the second temperature distribution in the order of t1, t2, t3, ..., until the formation of the laminated body is completed along the predetermined path, and outputs the information to the learning unit 35 (S15).

[0034] Furthermore, the temperature distribution acquisition unit 33 may acquire information on the temperature distribution at other times and provide it to the learning unit 35 as learning data. i to the next specific time t i+1 The time until the specified time t is set to be longer than the specified time Δt. For example, the specified time Δt is 1 second, and the specified time t i From t i+1 The time until the start of the process may be set to 10 seconds, etc., and each time can be set arbitrarily depending on the purpose.

[0035] Here, the time when a specified time Δt has elapsed since the specific time t0 is defined as ta, and the time from time ta to the next specific time t1 is defined as tb. In this case, the relationship between the second temperature distribution at time ta, when the specified time Δt has elapsed since the specific time t0, and the first temperature distribution after time Δtb has elapsed since time ta, i.e., at the next specific time t1, can also be learned. Specifically, the second temperature distribution at time ta is defined as the "first temperature distribution," and the first temperature distribution at time Δtb after time ta, i.e., the next specific time (t0 + Δt + Δtb = t1), is defined as the "second temperature distribution" and output to the learning unit 35. Similarly, information on each second temperature distribution after the specified time Δt has elapsed from the subsequent specific times t2, t3, . . . (treated as the "first temperature distribution") and each first temperature distribution at each specific time after the specific time t2 (treated as the "second temperature distribution") are output to the learning unit 35. The change in temperature distribution in this case is a temperature change accompanied by heat input, which makes it relatively easy to learn the cycle of cooling and heat input.

[0036] That is, the temperature distribution acquisition unit 33 acquires a first data set of a plurality of first temperature distributions in a time range including a plurality of specific times different from each other, and a second data set of a plurality of second temperature distributions in a time range including a specific time a predetermined time Δt has elapsed since each of the plurality of specific times, and also acquires a third data set of the second temperature distribution at a specific time ta in the second data set, and a fourth data set of the first temperature distribution at a specific time (ta+Δtb) in the first data set corresponding to the next time heat is input to the laminate after the specific time ta.

[0037] The learning unit 35 performs machine learning on the input information to generate a prediction model (S16). For example, in the above case, the learning unit performs machine learning on the relationship between the first data set and the second data set by associating them with the respective specified times Δt. Also, the learning unit performs machine learning on the relationship between the third data set and the fourth data set by associating them with the time difference (Δtb) between the second temperature distribution and the first temperature distribution. In this way, the learning unit 35 generates a prediction model.

[0038] When information on a first temperature distribution is input from the outside, this prediction model functions to predict a second temperature distribution after a specified time Δt has elapsed and output the prediction result.

[0039] FIG. 8 is an explanatory diagram showing a specific example of temperature distribution information input to the learning unit 35. The temperature distribution information input to the learning unit 35 is set for each unit element. If a weld bead is present in a unit element, a flag of "1" is set; if not, the flag is set to "0." For each unit element shown in FIG. 8, a unit element without a weld bead is flagged with a flag of "0," and a unit element with a weld bead is flagged with a flag of "1" and the temperature of the unit element. As a specific example, the temperature of the unit elements (six unit elements) at the position where a new weld bead is formed is 900°C, and the temperature of the unit elements (six unit elements) adjacent to that at the position where the previous weld bead was formed is 90°C. Furthermore, the temperature of the unit elements below these unit elements has dropped from 70°C to 20°C due to cooling over time. This temperature distribution information is provided to the learning unit 35 as learning data for machine learning.

[0040] Examples of machine learning methods for generating a predictive model include decision trees, linear regression, random forests, support vector machines, Gaussian process regression, and convolutional neural networks. Multiple predictive models may be generated for each type of filler metal. When consolidating the data into a single predictive model, some or all of the information on the components of the filler metal may be added to the training data for training.

[0041] When acquiring the second temperature distribution at a time (t = t0 + Δt) after the specified time Δt has elapsed, the temperature distribution acquisition unit 33 may repeatedly acquire the second temperature distribution by passing the specified time Δt until the temperature of the unit element corresponding to the formed weld bead falls below a predetermined reference temperature. When forming a weld bead, if the base weld bead is high in temperature and viscosity (fluidity), the desired bead height may not be achieved due to dripping of molten metal from the newly formed weld bead on the base or crushing of the base weld bead. Therefore, by waiting until the temperature of the base weld bead falls below a predetermined reference temperature Tp and then forming a new weld bead once it falls below the reference temperature Tp, it becomes easier to achieve the designed bead height. This reference temperature Tp can be considered the "interpass time" required to cool the stacked weld beads. The simulation unit 37 analytically determines the time required for the temperature to fall below the reference temperature Tp, allowing the learning unit 35 to learn the interpass time, thereby enabling accurate prediction of the temperature distribution and interpass time. This makes it possible to set the inter-pass time even at the stage of predicting the temperature distribution using a predictive model before the bead is actually formed, and the laminate manufacturing plan can be easily adjusted so that the upper layer weld bead is formed after the temperature of the underlying weld bead (the temperature of a specific part of the second temperature distribution or the maximum temperature) reaches the reference temperature Tp.

[0042] The prediction model is generated by machine learning the relationship between a first temperature distribution representing the temperature of multiple unit elements of the laminate at a specific time and a second temperature distribution representing the temperature of multiple unit elements at a time a specific time has elapsed since the specific time, by associating the relationship with the specific time. Therefore, a large amount of data from different times can be obtained from the calculation results of numerical analysis, and this large amount of data can be used as training data. Furthermore, by using a prediction model trained on a large amount of training data, the temperature distribution of the laminate during construction can be predicted with high accuracy and speed. Furthermore, because the temperature is calculated for each unit element, the temperature at any position on the laminate can be predicted.

[0043] <Temperature history prediction> Next, a temperature history prediction device that predicts the temperature distribution of a stack using a prediction model generated by machine learning using the learning device 200 described above will be described. 9 is a block diagram of the temperature history prediction device 300. The temperature history prediction device 300 includes a prediction model 41 generated by the learning device 200, an input data acquisition unit 43, a prediction unit 45, and a prediction control unit 47. The temperature history prediction device 300 has the same hardware configuration as the learning device 200 described above. The temperature history prediction device 300 may also be configured as an integrated part of the learning device 200.

[0044] Each of the above units will be described in detail later, but they generally function as follows: The input data acquisition unit 43 acquires input data including information on the temperature distribution (first temperature distribution) of multiple unit elements at an initial time when building a laminate. The prediction unit 45 uses the prediction model 41 to predict the temperature distribution after a specified time Δt has elapsed from the temperature distribution state included in the input data. The prediction control unit 47 inputs the information on the predicted temperature distribution back into the prediction unit 45 as input data, and determines the predicted temperature distribution after another specified time has elapsed. In this way, the temperature distribution after the elapse of specified times such as 2Δt, 3Δt, 4Δt, ... can be continuously obtained as a temperature history.

[0045] 10 is a flowchart showing a procedure for predicting the temperature history of a laminate when the laminate is molded. First, the input data acquisition unit 43 obtains a first temperature distribution at an initial time (t=t0) when the laminate is molded from the input data, and inputs information about the first temperature distribution to the prediction unit 45 (S21). Next, the prediction unit 45 predicts a second temperature distribution after a specified time Δt has elapsed, using the prediction model 41, from the input information about the first temperature distribution (S22). Then, the prediction of the temperature distribution is repeated until bead formation for one pass is completed (S23). When the prediction is repeated, the predicted second temperature distribution is input again to the prediction model 41 as the first temperature distribution (S24). As a result, the prediction model 41 again predicts the temperature distribution after the specified time has elapsed, based on the input temperature distribution (second temperature distribution).

[0046] Figure 11 is an explanatory diagram showing the formation of a weld bead and the temperature distribution of each unit element in a chronological order. The left column of Figure 11 shows the state of the weld bead on base 13, and the right column shows the temperature distribution in each state. At a specific time (initial time) t0, the unit element (shown in light hatching) corresponding to base 13 is at a constant temperature. The unit element above base 13 is a unit element (shown in a white frame) in which no weld bead B exists, and its flag is set to "0."

[0047] Next, when weld bead B is formed on base 13, the flag of the unit element (shown in dark hatching) at the position corresponding to weld bead B is set to "1," and the temperature distribution is predicted using prediction model 41. As time passes from this state, the temperature of weld bead B gradually decreases and becomes approximately the same temperature as the other existing weld beads B. Then, after the existing weld bead reaches a certain temperature (for example, the aforementioned reference temperature Tp), the next weld bead B is formed adjacent to the existing weld bead B. The flag of the unit element (shown in dark hatching) at the position corresponding to the newly formed weld bead B is set to "1," and the temperature distribution is predicted using prediction model 41.

[0048] As described above, the temperature change during the addition of the weld bead and the subsequent cooling is predicted using prediction model 41. Prediction model 41 is capable of accurate temperature prediction because it has learned the relationship between the temperature at the time of weld bead formation and the temperature distribution after a specified time.

[0049] When the formation of a weld bead in one pass is completed, it is determined whether the temperature of the unit element at the position corresponding to the weld bead formed in the current pass has reached the reference temperature Tp in the second temperature distribution last predicted by the prediction model 41 (S25). If the reference temperature Tp has not been reached, the temperature distribution after a specified time has elapsed is predicted. If the predicted temperature reaches the reference temperature Tp as a result, the time required to reach the reference temperature Tp is set as the interpass time described above. Here, prediction is repeated until the reference temperature Tp is reached, but it may be repeated for a predetermined period of time. When the predicted temperature reaches the reference temperature Tp, the prediction control unit 47 sets the time required to reach the reference temperature Tp as the interpass time and outputs it to the prediction unit 45. The prediction unit 45 outputs the interpass time together with information on the predicted temperature distribution as output information (S26).

[0050] The above process is repeated until all passes (passes (k): k=1 to N, N is the number of passes) for forming the laminated body are completed (S27, S28).

[0051] According to this, by predicting the temperature history during the manufacturing of the laminate using the prediction model 41, the calculation processing can be significantly reduced compared to when the temperature distribution is determined by repeated numerical analysis, and the temperature history can be determined accurately in a short time.

[0052] Figure 12 is an explanatory diagram showing the formation of a weld bead and the temperature distribution of each unit element in a time series, step by step. The temperature distribution prediction using prediction model 41 shown in Figure 12 is performed only for the process in which weld bead B is added and then cooled. The temperature rise caused by adding weld bead B is predicted by numerical analysis. In this way, the amount of calculation increases compared to when the entire heating and cooling processes are predicted by prediction model 41. However, by predicting the heat dissipation phenomenon, which is a relatively simple phenomenon, using prediction model 41 and calculating the complex process involving metal melting and solidification by numerical analysis, it is possible to achieve a good balance between improving overall prediction accuracy and reducing calculation time.

[0053] Generally, instead of performing numerical analysis such as the above-mentioned FEM, a method of prediction using a neural network or the like, as in this prediction method, is known as surrogate modeling. This surrogate modeling enables faster prediction than numerical analysis. By using a prediction model instead of the above-mentioned numerical analysis of temperature distribution, faster and more accurate temperature prediction can be easily performed.

[0054] FIG. 13 is an explanatory diagram illustrating how the element size of unit elements is changed. When performing machine learning on the learning unit 35 shown in FIG. 5 using element size information along with temperature distribution information, the learning unit 35 may learn the relationship between temperature distributions when the element size of unit elements varies. For example, if the temperature change over time in a specific temperature range is more drastic than in other ranges, the element size of unit elements included in that temperature range is reduced. This allows for more accurate understanding of temperature change, and learning with unit elements of small element sizes enables more accurate temperature prediction. Furthermore, for regions with particularly small changes over time, the element size of unit elements included in that region is increased. This allows for consolidation of regions with particularly little effect on temperature, reducing the amount of calculation without reducing accuracy. In other words, the prediction unit 45 outputs a second temperature distribution with an element size different from the first temperature distribution by adjusting the element size of the unit elements finer or coarser depending on temperature or other factors. This also achieves a good balance between improved prediction accuracy and reduced calculation time.

[0055] FIG. 14A is an explanatory diagram showing temperature changes when the specified time Δt is maintained constant until the predicted temperature approaches the reference temperature Tp. FIG. 14B is an explanatory diagram showing temperature changes when the specified time Δt is changed until the predicted temperature approaches the reference temperature Tp. When repeatedly predicting the temperature distribution over time, the specified time Δt is set to a constant time as shown in FIG. 14A in the prediction procedure described above. However, the specified time Δt is not limited to this. As shown in FIG. 14B, the time interval Δt over which the prediction control unit 47 repeatedly calculates the predicted temperature distribution may be set longer as the time goes back before the weld bead temperature in the predicted temperature distribution reaches the predetermined reference temperature Tp. In this case, calculations for intermediate stages when the predicted temperature reaches a temperature close to the reference temperature Tp can be omitted. Furthermore, the time interval Δt over which the prediction control unit 47 repeatedly calculates the predicted temperature distribution may be set shorter as the weld bead temperature in the predicted temperature distribution approaches the predetermined reference temperature Tp. In this case, the time at which the predicted temperature reaches the reference temperature Tp can be more accurately predicted. In learning the temperature distribution, it is preferable to learn a temperature distribution in a temperature range that is sufficiently higher than the reference temperature (interpass temperature) Tp, from the viewpoint of accurately determining the timing when the temperature falls below the interpass time.

[0056] In the machine learning of the learning device 200 described above, i The first temperature distribution and the specific time t i The time t is the time when the specified time Δt has passed from i+1 The first temperature distribution and the second temperature distribution were obtained by machine learning. FIG. 15A is an explanatory diagram showing the first temperature distribution and the second temperature distribution used in machine learning. FIG. 15B is an explanatory diagram showing the respective data sets of the first temperature distribution and the second temperature distribution used in machine learning. As shown in FIG. 15A, at a specific time t i Temperature distribution and time t i+1 When machine learning is performed using a set of information on the temperature distribution of i The temperature distribution at time t i+1 The temperature distribution at time t i+1 Therefore, as shown in FIG. 15B, it may be difficult to predict the temperature distribution at a specific time t i ,ti+1 ,···,t n-1 The temperature distribution data set and the time t i +Δt,t i+1 +Δt, ,t n The relationship between the temperature distribution data set of +Δt and each specified time Δt is associated with each other and machine learning is performed. In this case, the generated prediction model is i+1 Since it has also learned about times other than these, we can expect more accurate predictions.

[0057] In addition, the prediction model is i The second temperature distribution (treated as the "first temperature distribution") at time ta after a specified time Δt has elapsed from time ta, and the second temperature distribution (treated as the "first temperature distribution") at time ta after a specified time Δt has elapsed from time ta, that is, the next specific time t i+1 If the relationship between the temperature distribution and the first temperature distribution (treated as the "second temperature distribution") in the temperature distribution is also learned, predictions can be made that take into account temperature changes due to heat input. In addition, more information on temperature distribution obtained through numerical analysis can be used as learning data, which can efficiently improve prediction accuracy.

[0058] FIG. 16 is an explanatory diagram showing a state in which multiple unit elements obtained by dividing the shape of a laminate are made to resemble the shape of a weld bead. In the creation stage of a modeling plan for forming a laminate, a path (the movement path of the torch 11) PS indicating the formation trajectory and formation order of the weld bead is set, and multiple unit volume models 51, which are unit elements, are arranged so that they appear in time sequence along this path PS. Each unit volume model 51 preferably has a shape that resembles the shape of an actual weld bead. Furthermore, each unit volume model 51 is set to a size according to the welding conditions. That is, a weld bead of a specific path is simulated based on information about the movement direction and speed of the heat source included in the input data input to the learning device 200, and a bead shape model 53 is generated, which is a collection of multiple unit volume models 51 that appear in time sequence along the path. The temperature of the bead shape model 53 at the time of appearance is set according to the amount of heat input supplied from the heat source when the weld bead is formed.

[0059] By predicting the temperature distribution using the bead shape model 53 that is simulated to the shape of an actual weld bead in this way, it is possible to make an accurate prediction that minimizes the difference from the temperature distribution in the actual laminate. [Example]

[0060] <Verification example 1> The temperature distribution predicted using the temperature history prediction device described above was compared with the temperature distribution obtained by numerical analysis. 17 is a model diagram showing the shape of the laminate used in Verification Example 1. This model has a shape in which five layers of bead-shaped models BM made of voxel-like hexahedral elements (rectangular parallelepipeds) simulating weld beads are laminated on the top surface of a flat base 13. When this model is molded with weld beads, the temperature distribution in a two-dimensional cross section during air cooling after the five layers of weld beads are laminated was predicted.

[0061] The dimensions of the model are as follows: L1: 100mm L2: 100mm L3:6mm L4: 4mm L5:4mm The two-dimensional cross section used to determine the temperature distribution was a cross section of the model cut perpendicular to the longitudinal direction of the bead-shaped model BM at a position L2 / 2 along the longitudinal direction of the bead-shaped model BM from the end of the model (the hatched surface with the boundary indicated by the dashed line in Figure 17).

[0062] Learning method: For a total of 389 cases with different conditions in which the heat input and interpass temperature were randomly changed, the temperature distribution at the above cross section at time t when air-cooling was performed at 800°C to 100°C and the temperature distribution at time t + 10 seconds were obtained by numerical analysis, and this temperature distribution information was used for machine learning to generate a prediction model. Prediction method: The following steps were repeatedly performed: predicting the temperature distribution after 10 seconds using a prediction model with the initial temperature as an explanatory variable; and predicting the temperature distribution after 20 seconds using the prediction model with the temperature distribution after 10 seconds as an explanatory variable.

[0063] Figure 18 is a graph showing test results showing the time change of the predicted temperature values ​​and the temperatures obtained by numerical analysis at positions P1, P2, P3, and P4 on the two-dimensional cross section shown in Figure 17. At all positions P1 to P4, the predicted temperature values ​​almost matched the temperatures obtained by numerical analysis, demonstrating that it is possible to predict the temperature distribution with sufficient accuracy.

[0064] <Verification example 2> When a block was manufactured using the same procedure as in Verification Example 1, the predicted temperature distribution was compared with the temperature distribution obtained by numerical analysis. 19 is a model diagram showing the shape of the laminate used in Verification Example 2. This model has a shape with a total of 50 bead-shaped models BM of voxel-like hexahedral elements (rectangular parallelepipeds) simulating weld beads, with 10 rows of each layer stacked on top of a flat base 13. When this model is molded with a weld bead, the temperature distribution in a two-dimensional cross section during air cooling after the five weld bead layers are stacked was predicted.

[0065] The dimensions of the model are as follows: L1: 100mm L2: 100mm L3:6mm L4: 10mm L5:4mm The two-dimensional cross section used to determine the temperature distribution was taken at a position L2 / 2 along the longitudinal direction of the bead-shaped model BM from the end of the model, as shown by the dashed line, and cut the model perpendicular to the longitudinal direction of the bead-shaped model BM.

[0066] Learning method: For a total of 74 cases (49 cases in the high temperature range above 800°C, 25 cases in the low temperature range between 100°C and 800°C) with different conditions in which the heat input and inter-pass temperature were randomly changed, the temperature distribution at the above cross section at time t when air-cooling from 800°C to 100°C and the temperature distribution at time t + 10 seconds were determined by numerical analysis, and this temperature distribution information was used for machine learning to generate a prediction model. Prediction method: The following steps were repeatedly performed: predicting the temperature distribution after 10 seconds using a prediction model with the initial temperature as an explanatory variable; and predicting the temperature distribution after 20 seconds using the prediction model with the temperature distribution after 10 seconds as an explanatory variable.

[0067] Figure 20 is a cross-sectional view showing predicted positions of temperature distribution in the model shown in Figure 19. Here, one end of the lower layer in the rectangular cross section is P1, the other end is P2, the middle is P3, and the middle of the upper layer is P4. Figure 21 is a graph showing test results showing the time change of the predicted temperature values ​​and the temperatures obtained by numerical analysis at positions P1, P2, P3, and P4 on the two-dimensional cross section shown in Figure 19. At any of positions P1 to P4, there was almost no difference in the time change between the predicted temperature values ​​and the temperatures obtained by numerical analysis, and the standard deviation of the difference between the two was 6.2°C. Furthermore, the calculation time for obtaining the temperature distribution by numerical analysis was 3,850 seconds, but the calculation time for prediction using the prediction model was reduced to 13.4 seconds.

[0068] As such, the present invention is not limited to the above-described embodiments, and the present invention also contemplates the mutual combination of the various components of the embodiments, as well as modifications and applications by those skilled in the art based on the description in the specification and well-known techniques, and these modifications and applications are included in the scope of protection sought.

[0069] As described above, the present specification discloses the following: (1) A learning device that generates, by machine learning, a prediction model that predicts the temperature history of a laminate during the formation of the laminate by each unit element obtained by dividing the shape of the laminate, when the laminate is formed by a weld bead formed by melting and solidifying a filler material by moving a heat source along a predetermined path, a temperature distribution acquiring unit that acquires a first temperature distribution representing temperatures of the plurality of unit elements at a specific time of the laminated body and a second temperature distribution representing temperatures of the plurality of unit elements at a time when a specified time has elapsed since the specific time; a learning unit that performs machine learning on the relationship between the first temperature distribution and the second temperature distribution obtained by the temperature distribution acquisition unit, by associating the relationship with the specified time, and generates the prediction model; A learning device comprising: This learning device can acquire a large amount of learning data from the results of numerical analysis or actual measurements of the relationship between the first temperature distribution and the second temperature distribution after a specified time has elapsed. Therefore, machine learning of a large amount of learning data can generate a prediction model with high prediction accuracy. Furthermore, because the temperature is calculated for each unit element, a prediction model that can accurately predict the temperature at any position in the stack can be obtained.

[0070] (2) The learning device described in (1), wherein the learning data used for the machine learning includes the amount of heat input supplied from the heat source when the weld bead is formed, the direction and speed of movement of the heat source, the volume of the weld bead formed per unit time, and status information indicating the presence or absence of the weld bead for each unit element. This learning device, which includes information about the heat source, can obtain a prediction model that can predict the temperature distribution when a weld bead is formed while the heat source is moving. Also, by learning status information indicating the presence or absence of a weld bead, it is possible to predict the generation of a weld bead as the heat source moves, along with the temperature distribution.

[0071] (3) The learning device according to (2), wherein the learning data further includes size information of the unit elements. According to this learning device, since the size information of the unit element is included in the input data, prediction can be made at any size of unit element (mesh size).

[0072] (4) A learning device described in any one of (1) to (3), wherein the temperature distribution acquisition unit calculates the temperatures of the plurality of unit elements during the formation of the weld bead by numerical analysis to obtain the first temperature distribution and the second temperature distribution. This learning device calculates temperature distribution through numerical analysis, making it possible to generate large amounts of temperature information and easily obtain large amounts of learning data, which makes it easier to improve the prediction accuracy of the prediction model.

[0073] (5) A learning device described in any one of (1) to (4), wherein the temperature distribution acquisition unit repeatedly determines the second temperature distribution by allowing the specified time to pass until the temperature of the unit element corresponding to the weld bead after formation becomes equal to or lower than a predetermined reference temperature. This learning device repeatedly determines the second temperature distribution until the formed weld bead falls below the reference temperature, thereby generating a prediction model that can accurately predict the temperature distribution up to the time corresponding to the inter-pass time when the next weld bead can be formed on the weld bead.

[0074] (6) The temperature distribution acquisition unit acquires a plurality of pairs of the first temperature distribution and the second temperature distribution by changing at least one of the specific time and the specified time, The learning device described in any one of (1) to (5) generates the prediction model by machine learning the relationship between the first temperature distribution and the second temperature distribution of each of the multiple sets, by correlating it with the change amount at the specific time and the change amount over the specified time. According to this learning device, by acquiring a plurality of sets in which at least one of the specific time and the specified time is different, the amount of learning data can be increased, thereby improving prediction accuracy.

[0075] (7) The temperature distribution acquisition unit acquires a first data set of a plurality of the first temperature distributions in a time range including a plurality of specific times different from each other, and a second data set of a plurality of the second temperature distributions in a time range including specific times each of which is a predetermined time after each of the plurality of specific times, and also acquires a third data set of the second temperature distributions at the specific times of the second data set and a fourth data set of the first temperature distributions at the specific times of the first data set corresponding to the next time heat is input to the laminate after the specific times, The learning device described in any one of (1) to (5), wherein the learning unit performs machine learning on the relationship between the first data set and the second data set by associating them with the respective specified times, and performs machine learning on the relationship between the third data set and the fourth data set by associating them with the time difference between the second temperature distribution and the first temperature distribution, thereby generating the prediction model. According to this learning device, the generated prediction model learns the temperature change due to cooling over a specified time and heating when the next heat input occurs, so the cooling and heat input cycle can be learned relatively easily, and a prediction model can be generated that is expected to make more accurate predictions.

[0076] (8) The prediction model generated by the learning device according to any one of (1) to (7); an input data acquisition unit that acquires input data including information on a temperature distribution of the plurality of unit elements at an initial time when the laminated body is modeled; a prediction unit that obtains a predicted temperature distribution by predicting a temperature distribution after a specified time has elapsed from the state of the temperature distribution included in the input data using the prediction model; a prediction control unit that inputs information about the predicted temperature distribution as the input data back into the prediction unit and further obtains a predicted temperature distribution after the specified time has elapsed; A temperature history prediction device comprising: This temperature history prediction device can obtain a large amount of learning data, for example, from calculation results by numerical analysis or actual measurements, on the relationship between the temperature distribution at a specific time and the temperature distribution at a time a specified time has elapsed since the specific time. By using a prediction model that uses machine learning on a large amount of learning data, the temperature distribution of a laminated body during construction can be predicted with high accuracy and speed. Furthermore, by inputting the temperature distribution after a specified time into the prediction model, the temperature distribution after the specified time can be continuously predicted. Furthermore, since the temperature is calculated for each unit element, the temperature at any position on the laminated body can be predicted.

[0077] (9) The temperature history prediction device according to (8), wherein the time interval at which the prediction control unit repeatedly calculates the predicted temperature distribution is set to a constant interval. According to this temperature history prediction device, the temperature history of the stack can be determined sequentially in time series.

[0078] (10) The temperature history prediction device described in (8), wherein the time interval at which the prediction control unit repeatedly calculates the predicted temperature distribution is set longer the further back in time it goes before the time at which the temperature of the weld bead in the predicted temperature distribution reaches a predetermined reference temperature. According to this temperature history prediction device, it is possible to omit calculations during the intermediate stages until the predicted temperature reaches a temperature close to the reference temperature.

[0079] (11) The temperature history prediction device described in (8), wherein the prediction control unit repeats the prediction of the predicted temperature distribution by the prediction unit until the temperature of the unit element corresponding to the weld bead in the predicted temperature distribution falls below a predetermined reference temperature, and outputs the time required for the temperature to fall below the reference temperature as the inter-pass time. According to this temperature history prediction device, by outputting the inter-pass time, it is possible to efficiently create a building plan for a laminate based on the predicted temperature distribution results and the inter-pass time without actually carrying out building.

[0080] (12) The input data acquisition unit acquires information on the amount of heat input supplied from the heat source when forming the weld bead, the direction and speed of movement of the heat source, the volume of the weld bead formed per unit time, and the element state information indicating the presence or absence of the weld bead for each unit element; The temperature history prediction device described in (8), wherein the prediction control unit, after the formation of the weld bead along the specific path, determines the temperature of the unit element corresponding to the weld bead from the information on the predicted temperature distribution output from the prediction unit, and when the temperature falls below a predetermined reference temperature for the first time, inputs the information on the predicted temperature distribution into the prediction unit as the input data to determine the predicted temperature distribution. This temperature history prediction device acquires information about the heat source and predicts the temperature distribution, thereby accurately predicting the temperature distribution when a weld bead is formed while the heat source is moving. Furthermore, by including status information indicating the presence or absence of a weld bead in the input data, the generation of a weld bead as the heat source moves can be accurately reproduced, improving the accuracy of the temperature distribution prediction. Furthermore, by predicting the temperature distribution when the temperature of the unit element corresponding to the weld bead falls below the reference temperature, the temperature distribution after the interpass time has elapsed can be predicted.

[0081] (13) The temperature history prediction device described in any one of (8) to (12), wherein the prediction unit simulates the weld bead of a specific path based on information on the movement direction and movement speed of the heat source included in the input data, generates a bead shape model that is a collection of multiple unit volume models in which unit volume models appear in chronological order along the path, and calculates the temperature at the time of appearance of the unit volume model depending on the amount of heat input supplied from the heat source when the weld bead is formed. According to this temperature history prediction device, the accuracy of predicting the temperature distribution can be further improved by forming a bead shape model according to the amount of heat input supplied from the heat source.

[0082] (14) The input data acquisition unit acquires size information of the unit elements, The temperature history prediction device according to any one of (8) to (13), wherein the prediction unit outputs the second temperature distribution having an element size different from that of the first temperature distribution. According to this temperature history prediction device, the element size of the unit element can be set to an appropriate balance between the required prediction accuracy and the amount of calculation, thereby making it possible to efficiently predict temperature distribution.

[0083] (15) A learning device according to any one of (1) to (7). Temperature history prediction device. This temperature history prediction device can predict the temperature history of a molded object with high accuracy and high speed.

[0084] (16) A temperature history prediction device according to any one of (8) to (15), a welding device that builds the laminate based on building conditions that are determined using information on the temperature distribution predicted by the temperature history prediction device; and A welding system comprising: This welding system enables the production of high-quality laminated bodies by implementing more appropriate manufacturing plans.

[0085] (17) A program for generating, by machine learning, a prediction model that predicts, when a laminate is manufactured by a weld bead formed by melting and solidifying a filler material by moving a heat source along a predetermined path, a temperature history of the laminate during the manufacturing process for each unit element obtained by dividing the shape of the laminate, the program comprising: On the computer, a function of calculating a first temperature distribution representing the temperatures of the plurality of unit elements at a specific time of the laminate, and a second temperature distribution representing the temperatures of the plurality of unit elements at a time when a specified time has elapsed since the specific time; a function of performing machine learning on the relationship between the first temperature distribution and the second temperature distribution and the specified time to generate the prediction model; A program to achieve this. This program allows for the acquisition of a large amount of training data from the results of numerical analysis or actual measurements of the relationship between the first temperature distribution and the second temperature distribution after a specified time has elapsed. Therefore, machine learning of a large amount of training data allows for the generation of a prediction model with high prediction accuracy. Furthermore, because the temperature is calculated for each unit element, a prediction model that can accurately predict the temperature at any position in the laminate can be obtained.

[0086] (18) Using the prediction model generated by the learning device according to any one of (1) to (7), On the computer, a function of acquiring input data including information on the temperature distribution of the plurality of unit elements at an initial time when the laminate is modeled; and a function of calculating a predicted temperature distribution by predicting, using the prediction model, a temperature distribution after a specified time has elapsed from the state of the temperature distribution included in the input data; a function of using the predicted temperature distribution as the input data and again using the prediction model to further predict a temperature distribution after the specified time has elapsed; A program to achieve this. This program uses a prediction model that has learned the relationship between the temperature distribution at a specific time and the temperature distribution at a time a specified time has elapsed since the specific time, making it possible to quickly and accurately predict the temperature distribution of a laminated body during modeling. Furthermore, by inputting the temperature distribution after a specified time into the prediction model, it is possible to continuously predict the temperature distribution after the specified time. [Explanation of symbols]

[0087] 11 Torch 13. Bass 15. Modeling control device 17 Manipulator 19 Filler metal supply device 19a Reel 19b Feeding mechanism 21 Manipulator control device 23 Heat source control device 31 Element division section 33 Temperature distribution acquisition section 35 Learning Department 37 Simulation Department 41 Predictive Models 43 Input data acquisition unit 45 Prediction Department 47 Predictive control unit 51 Unit Volume Model 53 Bead shape model B Weld bead Ly1,Ly2,Ly3,Ly4 layer M filler metal W laminate 100 welding equipment 200 Learning Device 300 Temperature History Prediction Device

Claims

1. A learning device that generates, by machine learning, a prediction model that predicts the temperature history of a laminate during the formation of the laminate by each unit element obtained by dividing the shape of the laminate, when the laminate is formed by a weld bead formed by melting and solidifying a filler material by moving a heat source along a predetermined path, a temperature distribution acquiring unit that acquires a first temperature distribution representing temperatures of the plurality of unit elements at a specific time of the laminated body and a second temperature distribution representing temperatures of the plurality of unit elements at a time when a specified time has elapsed since the specific time; a learning unit that performs machine learning on the relationship between the first temperature distribution and the second temperature distribution obtained by the temperature distribution acquisition unit, by associating the relationship with the specified time, and generates the prediction model; Equipped with The learning data used for the machine learning includes the amount of heat input supplied from the heat source when the weld bead is formed, the direction and speed of movement of the heat source, the volume of the weld bead formed per unit time, and state information indicating the presence or absence of the weld bead for each unit element. Learning device.

2. the learning data further includes size information of the unit elements. The learning device according to claim 1 .

3. The learning device according to claim 1 or 2, wherein the temperature distribution acquisition unit calculates the temperatures of the plurality of unit elements during the formation of the weld bead by numerical analysis to obtain the first temperature distribution and the second temperature distribution.

4. the temperature distribution acquisition unit repeatedly obtains the second temperature distribution by allowing the specified time to elapse until the temperature of the unit element corresponding to the weld bead after formation becomes equal to or lower than a predetermined reference temperature. The learning device according to any one of claims 1 to 3.

5. the temperature distribution acquisition unit acquires a plurality of pairs of the first temperature distribution and the second temperature distribution by changing at least one of the specific time and the specified time, the learning unit performs machine learning on the relationship between the first temperature distribution and the second temperature distribution for each of the plurality of pairs, by associating the relationship with the amount of change at the specific time and the amount of change over the specified period of time, and generates the prediction model. The learning device according to any one of claims 1 to 4.

6. the temperature distribution acquisition unit acquires a first data set of a plurality of the first temperature distributions in a time range including a plurality of specific times different from each other, and a second data set of a plurality of the second temperature distributions in a time range including specific times each of which is a predetermined time after each of the plurality of specific times, and also acquires a third data set of the second temperature distributions at the specific times of the second data set and a fourth data set of the first temperature distributions at the specific times of the first data set corresponding to the next time heat is input to the laminate after the specific times; the learning unit performs machine learning on the relationship between the first data set and the second data set by associating them with the respective specified times, and performs machine learning on the relationship between the third data set and the fourth data set by associating them with the respective time differences between the second temperature distribution and the first temperature distribution, thereby generating the prediction model. The learning device according to any one of claims 1 to 4.

7. The prediction model generated by the learning device according to any one of claims 1 to 6; an input data acquisition unit that acquires input data including information on a temperature distribution of the plurality of unit elements at an initial time when the laminated body is modeled; a prediction unit that obtains a predicted temperature distribution by predicting a temperature distribution after a specified time has elapsed from the state of the temperature distribution included in the input data using the prediction model; a prediction control unit that inputs information about the predicted temperature distribution as the input data back into the prediction unit and further obtains a predicted temperature distribution after the specified time has elapsed; A temperature history prediction device comprising:

8. The time interval at which the prediction control unit repeatedly calculates the predicted temperature distribution is set to a constant interval. The temperature history prediction device according to claim 7 .

9. a time interval at which the prediction control unit repeatedly obtains the predicted temperature distribution is set to be longer as the time goes back in time before the temperature of the weld bead in the predicted temperature distribution reaches a predetermined reference temperature; The temperature history prediction device according to claim 7 .

10. the prediction control unit repeats prediction of the predicted temperature distribution by the prediction unit until the temperature of the unit element corresponding to the weld bead in the predicted temperature distribution falls below a predetermined reference temperature, and outputs the time required for the temperature to fall below the reference temperature as an interpass time. The temperature history prediction device according to claim 7 .

11. the input data acquisition unit acquires information on the amount of heat input supplied from the heat source when forming the weld bead, the direction and speed of movement of the heat source, the volume of the weld bead formed per unit time, and element state information indicating the presence or absence of the weld bead for each unit element; the prediction control unit, after the weld bead is formed along the specific path, determines the temperature of the unit element corresponding to the weld bead from the information on the predicted temperature distribution output from the prediction unit, and when the temperature falls below a predetermined reference temperature for the first time, inputs the information on the predicted temperature distribution to the prediction unit as the input data to determine the predicted temperature distribution. The temperature history prediction device according to claim 7 .

12. the prediction unit simulates the weld bead of a specific path based on information on the moving direction and moving speed of the heat source included in the input data, generates a bead shape model that is an aggregate of a plurality of unit volume models in which unit volume models appear in time sequence along the path, and determines the temperature at the time of appearance of the unit volume model according to the amount of heat input supplied from the heat source when the weld bead is formed; The temperature history prediction device according to any one of claims 7 to 11.

13. the input data acquisition unit acquires size information of the unit elements; the prediction unit outputs the second temperature distribution having an element size different from that of the first temperature distribution. The temperature history prediction device according to any one of claims 7 to 12.

14. The temperature history prediction device according to any one of claims 7 to 13; a welding device that builds the laminate based on building conditions that are determined using information on the temperature distribution predicted by the temperature history prediction device; and A welding system comprising:

15. Using the prediction model generated by the learning device according to any one of claims 1 to 6, On the computer, a function of acquiring input data including information on the temperature distribution of the plurality of unit elements at an initial time when the laminate is modeled; and a function of calculating a predicted temperature distribution by predicting, using the prediction model, a temperature distribution after a specified time has elapsed from the state of the temperature distribution included in the input data; a function of using the predicted temperature distribution as the input data and again using the prediction model to further predict a temperature distribution after the specified time has elapsed; A program to achieve this.

Images