JP7764265B2 - Learning device, nozzle maintenance device, welding control device and welding device - Google Patents

Description

The present invention relates to a learning device, a nozzle maintenance device, a welding control device, and a welding device.

In operations involving continuous, long-term bead deposition, such as additive manufacturing or multi-layer welding, spatter accumulates on the welding torch nozzle. This can make it necessary to replace the nozzle midway through the welding process. While maintaining low spatter deposition is desirable from the perspective of maintaining bead quality, removing the deposited spatter or replacing the nozzle temporarily halts bead formation. For example, if the nozzle is cleaned less frequently than appropriate, excessive spatter deposition may occur, potentially preventing an adequate supply of shielding gas. On the other hand, if the nozzle is cleaned more frequently than appropriate, the time required for cleaning increases, reducing manufacturing efficiency. Therefore, from a productivity perspective, there is a desire to minimize removal and replacement work. It is advisable to plan ahead for this spatter accumulation, particularly by predicting the amount of spatter accumulation and appropriately timing nozzle cleaning or nozzle replacement.

Patent Document 1 discloses a technology for detecting nozzle clogging in welding torches such as the one described above. In this technology, the gas flow rate in the gas supply passage within the torch is detected, and the extent of nozzle clogging due to spatter is determined from the level of that gas flow rate.

Japanese Patent Application Publication No. 11-267848

However, because the technology in Patent Document 1 is designed to capture the characteristics of blocked gas supply passages, it is not capable of quantitatively evaluating the level of spatter accumulation. This makes it difficult to use it to detect signs of quality degradation in the bead. Furthermore, the amount of spatter deposition is based on a complex phenomenon that depends on many parameters, such as welding conditions, for example, welding speed, filler metal feed rate, distance from the nozzle tip to the bead formation position (tip-to-work distance), torch angle, and rod operation method (weaving, etc.), so it is generally difficult to predict the amount of deposition.

The present invention aims to provide a learning device, nozzle maintenance device, welding control device, and welding device that enables accurate prediction of difficult-to-predict spatter accumulation amounts by using a prediction model that uses machine learning of specific welding conditions and measured values of spatter accumulation amounts, thereby maintaining the quality of the formed bead.

The present invention comprises the following configurations.
(1) A learning device that generates, by machine learning, a prediction model for predicting an amount of spatter that accumulates on a nozzle of a welding torch when a filler metal is melted and solidified at the tip of the welding torch to form a bead, the learning device comprising:
an information extraction unit that extracts, from a molding plan for molding an object having a desired shape, bead formation information, which is information on a movement path of the welding torch and welding conditions including at least a welding speed, a feed rate of the filler metal supplied to the welding torch, and a distance from a tip of the nozzle to a formation position of the bead;
an accumulation amount measuring unit that measures an accumulation amount of spatter accumulated on the nozzle of the welding torch when the bead is formed;
a model generation unit that performs machine learning on the relationship between the bead formation information and the sputter deposition amount to generate the prediction model that represents the relationship between the bead formation information and the sputter deposition amount corresponding to the bead to be formed;
A learning device comprising:
(2) The learning device according to (1),
a plan receiving unit that receives information about the modeling plan;
a deposition amount estimation unit that extracts the bead formation information from the received information of the manufacturing plan, inputs the extracted bead formation information into the prediction model, and calculates an estimated value of the sputter deposition amount from the prediction model;
a determination unit that determines whether the estimated value of the sputter deposition amount exceeds a predetermined reference value;
a plan change unit that changes the shaping plan when it is determined that the estimated value exceeds the reference value;
A nozzle maintenance support device comprising:
(3) The nozzle maintenance support device according to (2),
a control unit that outputs a control signal for executing arc welding based on the changed modeling plan;
A welding control device comprising:
(4) The welding control device according to (3),
a welding robot that performs arc welding;
a nozzle station for cleaning or replacing the nozzle of the welding torch;
A welding device comprising:

According to the present invention, the amount of spatter accumulation, which is difficult to predict, can be accurately predicted by using a prediction model that uses machine learning of specific welding conditions and measured values of spatter accumulation, thereby maintaining the quality of the formed bead.

FIG. 1 is a diagram showing the overall configuration of a welding apparatus for manufacturing a shaped object. FIG. 2 is a functional block diagram of the nozzle maintenance support device. FIG. 3 is a flowchart showing the procedure for generating a prediction model. FIG. 4A is an explanatory diagram showing the state of the tip of the welding torch during welding. FIG. 4B is an explanatory diagram showing the state of the tip of the welding torch during welding. FIG. 5 is a flowchart showing a procedure for changing the modeling program. FIG. 6 is an explanatory diagram showing an example of changes to the modeling program. FIG. 7 is a graph showing the distribution of predicted values by the prediction model with respect to the actually measured values of the sputter deposition amount.

Embodiments of the present invention will be described in detail below with reference to the drawings. Here, we will explain an example of a configuration in which the present invention is applied to additive manufacturing, in which a welding robot holding a welding torch is used to melt and solidify a filler material supplied to the welding torch to form a bead on a base material. However, the present invention is not limited to this, and can also be applied to general welding such as fillet welding, butt welding, and multi-layer welding within a groove.

<Configuration of welding equipment>
FIG. 1 is a diagram showing the overall configuration of a welding apparatus for manufacturing a shaped object.
The welding apparatus 100 includes a manufacturing unit 11, a control unit 13 that controls the manufacturing unit 11, and a nozzle maintenance support device 15. The control unit 13 and the nozzle maintenance support device 15 function together as a welding control device 16.

The manufacturing unit 11 includes a welding robot 19 having a welding torch 17 on its tip shaft, a robot drive unit 21 that drives the welding robot 19, a filler material supply unit 23 that supplies filler material (welding wire) M to the welding torch 17, a welding power supply unit 25 that supplies welding current and welding voltage to the welding torch 17, a nozzle station 27 that cleans and replaces the shield nozzle (nozzle) of the welding torch 17, and an imaging unit 29 that images the welded area where the bead B is formed.

(Modeling Department)
Welding robot 19 is an articulated robot, and a filler material M is supported at the tip of welding torch 17 attached to the tip shaft of the robot arm. The position and posture of welding torch 17 can be set arbitrarily in three dimensions within the range of the degrees of freedom of the robot arm by commands from robot driver 21. Although not shown, a weaving mechanism for weaving welding torch 17 may be provided at the tip shaft of the robot arm.

Welding torch 17 is a torch for gas metal arc welding, to which shielding gas is supplied from a nozzle (not shown). The arc welding method 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 welding or plasma arc welding, and is selected appropriately depending on the additive manufacturing object to be produced.

For example, in the case of a consumable electrode type, a contact tip is placed inside the nozzle, and the filler material M, to which the welding current is supplied, is held by the contact tip. While holding the filler material M, the welding torch 17 generates an arc from the tip of the filler material M in a shielding gas atmosphere.

The filler material supply unit 23 is equipped with a reel 23a around which filler material M is wound. The filler material M is sent from the filler material supply unit 23 to a feeding mechanism attached to a robot arm or the like, and is fed to the welding torch 17 while being fed forward and backward by the feeding mechanism as needed.

Any commercially available welding wire can be used as the filler metal M. For example, welding wires specified in MAG welding and MIG welding solid wire for mild steel, high-tensile steel, and low-temperature steel (JIS Z 3312), arc welding flux-cored wire for mild steel, high-tensile steel, and low-temperature steel (JIS Z 3313), etc. can be used. Furthermore, filler metals M such as aluminum, aluminum alloys, nickel, and nickel-based alloys can be used depending on the desired characteristics.

The robot driver 21 drives the welding robot 19 to move the welding torch 17. As the welding torch 17 moves, the continuously supplied filler material M is melted by the welding current and welding voltage from the welding power source 25.

In other words, the welding robot 19 is a manipulator that holds a welding torch 17 at the tip of its arm, which generates an arc while melting and solidifying a wire-shaped filler metal M. While the welding torch 17 is moved by driving this manipulator, the filler metal M, which is continuously fed to the welding torch 17, is melted and solidified by the arc, forming a bead B, which is a molten solidified body of the filler metal M, on the base plate BP, which is the base material.

The imaging unit 29 is composed of a camera attached to the welding torch 17 or to a manipulator closer to the base of the welding torch 17, and captures images of the welding area, including the arc and molten pool generated at the tip of the welding torch 17. The imaging unit 29 may be a general optical camera, or may be configured to use other sensors, such as an infrared detection sensor, or a combination of these. The image information obtained by imaging may be still image information or video information. The imaging unit 29 may also be supported separately from the welding torch 17.

The nozzle station 27 cleans or replaces the nozzle of the welding torch 17. The welding robot 19 drives the welding torch 17 to move it to the position of the nozzle station 27, where spatter deposits are removed from the nozzle attached to the welding torch 17 or replaced with a new nozzle.

(Control unit)
The control unit 13 is a computer device including an input/output interface, a calculation unit, and a memory unit (not shown). The input/output interface unit is connected to the welding robot 19, the robot driver 21, the welding power supply 25, the filler metal supply unit 23, the nozzle maintenance support device 15, and the like. The calculation unit includes a processor such as a CPU or an MPU and memories such as a ROM or a RAM. The memory unit has a drive device such as a hard disk or a solid-state drive (SSD), and may include various storage media such as CDs, DVDs, and various memory cards. The memory unit stores various information, such as a driving program (described below), and is capable of inputting and outputting various information. A molding program corresponding to the object to be manufactured is input to the control unit 13 via the various storage media or a communication line such as a network. The molding program is created based on a molding plan that defines the bead formation trajectory (the movement path of the welding torch 17) for forming the bead B and the welding conditions for forming the bead, and is composed of numerous command codes.

The control unit 13 executes the manufacturing program stored in the memory unit, driving each unit, such as the welding robot 19, filler material supply unit 23, and welding power supply unit 25, to form a bead B according to the manufacturing program. In other words, the control unit 13 causes the robot drive unit 21 to drive the welding robot 19, moving the welding torch 17 along the movement path set in the manufacturing program, and controls the filler material supply unit 23 and welding power supply unit 25 according to the set welding conditions to melt and solidify the filler material M at the tip of the welding torch 17 by an arc. In this way, by sequentially stacking beads B based on the manufacturing program, a desired three-dimensional object 31 is manufactured.

The nozzle maintenance support device 15 is composed of a computer device similar to the control unit 13. The nozzle maintenance support device 15 has the function of changing the molding program (molding plan) executed by the control unit 13 in accordance with the amount of spatter deposited on the nozzle of the welding torch 17.

(Nozzle maintenance support device)
FIG. 2 is a functional block diagram of the nozzle maintenance support device 15.
The nozzle maintenance support device 15 includes a learning device 33 , a plan receiving unit 35 , an accumulation amount estimating unit 37 , a determining unit 39 , and a plan changing unit 41 .
During welding, in which filler metal M is melted and solidified at the tip of welding torch 17 to form bead B, learning device 33 generates, by machine learning, a prediction model ML for predicting the amount of spatter scattered at the tip of welding torch 17 and deposited on the nozzle of welding torch 17. Details of this prediction model ML will be described later.

The plan receiving unit 35 receives information about the forming plan in the form of the above-mentioned forming program, etc. The deposition amount estimation unit 37 extracts bead formation information from the received forming plan information, which is information about the movement path of the welding torch 17 and welding conditions including at least the welding speed, the feed rate of the filler metal M supplied to the welding torch 17, and the distance from the nozzle tip to the bead formation position (tip-to-work distance). The extracted bead formation information is then input into a prediction model ML generated by the learning device 33, and an estimated value for the spatter deposition amount corresponding to the bead formation information is obtained from the prediction model ML.

The determination unit 39 determines whether the estimated value of the sputter deposition amount obtained by the deposition amount estimation unit 37 exceeds a predetermined reference value. The plan modification unit 41 modifies the modeling plan if it is determined that the estimated value exceeds the reference value. If the estimated value is equal to or less than the reference value, the modeling plan is not modified. The input modeling plan or the modified modeling plan, i.e., the modeling program, is output to the control unit 13 in Figure 1 and executed by the control unit 13.

When the above-mentioned bead formation information is input, the prediction model ML outputs an estimated value of the sputter deposition amount according to the input. This prediction model ML is generated by the learning device 33 performing the following machine learning.

As shown in Figure 2, the learning device 33 includes an information extraction unit 43, a deposition amount measurement unit 45, and a model generation unit 47. The learning device 33 forms a bead according to the input modeling plan, measures the nozzle deposition amount deposited on the nozzle, and performs machine learning on the measurement results in accordance with the modeling plan. A prediction model ML is generated through this machine learning.

<Procedure for generating a prediction model>
FIG. 3 is a flowchart showing the procedure for generating the prediction model ML.
First, a molding plan for the object to be molded is created according to a predetermined algorithm. For example, slice data is generated by slicing the shape of the object input using CAD data using a predetermined method, and a molding plan is created that defines the bead formation sequence, layering positions, and welding conditions based on the obtained slice data. The molding plan can be created using a conventionally known method.

The information on the created molding plan (molding program) is input to the information extraction unit 43 (S11). The information extraction unit 43 extracts bead formation information from the input molding program information, including the bead formation trajectory (movement path) for forming bead B and the welding conditions for forming the bead (S12).

Furthermore, before forming the bead, the deposition amount measurement unit 45 measures the initial weight of the nozzle of the welding torch 17 (S13). For example, the nozzle is removed from the welding torch 17, and the weight of the nozzle is measured using a measuring device such as an electronic balance. This measurement result is designated as W1. Then, based on the inputted forming program, the learning device 33 outputs a control signal to the control unit 13 for driving the aforementioned forming unit 11 (Figure 1), and executes a welding process to form a bead under predetermined conditions (S14).

As the bead is formed at this time, spatter adheres to the nozzle of the welding torch 17. In other words, as the bead is formed by executing the molding program, the amount of spatter deposited on the nozzle increases. After the bead is formed, the nozzle is removed from the welding torch 17, and the deposition amount measuring unit 45 measures the weight of the nozzle in the same way as measuring the initial weight (S15). This measurement result is designated as W2.

The model generation unit 47 calculates the sputter deposition amount (W2 - W1) of the sputter adhering to the nozzle from the obtained measurement results (S16). Then, machine learning is performed to associate the obtained sputter deposition amount value with the bead formation information that formed that bead (S17). This bead formation and machine learning are repeated (S18), and finally, the model generation unit 47 generates a prediction model ML that represents the correspondence between the bead formation information and the sputter deposition amount (S19).

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

In the case of the above model generation procedure, learning is performed from the results of forming a bead based on a pre-created molding plan, so the necessary information can be obtained efficiently, making it easier to improve the prediction accuracy of the prediction model ML. It is also possible to generate a prediction model ML without using the above-mentioned molding plan. For example, a bead can be formed under appropriate conditions such as time or distance, and the amount of spatter deposition generated during bead formation can be measured in the same manner as above, and a prediction model ML that associates bead formation information with the amount of spatter deposition can be generated. In this case, learning can be performed under any conditions, making it easier to generate a prediction model ML that can be used for general purposes.

The learning data used for machine learning should at least include information about the welding conditions when forming the bead and the amount of spatter accumulation. Specific welding conditions include parameters such as the welding mode of the welding power source (constant voltage welding mode, pulse welding mode, etc.), filler metal feed rate, welding speed, welding current, welding voltage, tip-to-work distance, whether or not weaving is used, weaving conditions, torch angle, arc-on time, and number of arc-on/off cycles. The occurrence of spatter changes depending on the increase or decrease of any of these parameters.

On the other hand, the amount of deposited spatter may be the weight of deposited spatter, or may be a value obtained by detecting a characteristic amount from image information obtained by capturing an image of spatter scattering around the arc.
4A and 4B are explanatory diagrams showing the state of the tip of the welding torch 17 during welding. As shown in FIGS. 4A and 4B, spatter generated from the tip of the welding torch 17 may be extracted by image processing or the like, and a value corresponding to the amount of spatter extracted in the resulting image (e.g., image area, spatter scattering direction, etc.) may be used as learning data. Furthermore, a moving image of the state of spatter scattering during bead formation may be quantitatively evaluated and used as learning data. This allows the progress of spatter deposition to be reflected in the learning.

Furthermore, as shown in Figure 4A, when the welding torch 17 is oriented normal to the workpiece surface, spatter scatters uniformly radially from the molten pool. However, as shown in Figure 4B, when the welding torch 17 is tilted from the normal to the workpiece surface, spatter tends to scatter away from the welding torch 17. Therefore, the amount of spatter deposited on the nozzle varies depending on the torch angle of the welding torch 17. Taking advantage of this, processing can be added to correct the amount of spatter deposited according to the torch angle. Furthermore, when quantitatively evaluating spatter, image processing can be performed in advance to make the spatter stand out, or processing such as enhancing the torch outline can be performed to emphasize the relationship between the orientation of the welding torch 17 and the spatter scattering direction. This allows the spatter scattering direction and the torch orientation to be reflected in the learning, making it easier to accurately predict the amount of spatter deposition. In other words, by learning the torch angle, the relationship between angles at which spatter deposition is favorable and unfavorable can also be reflected in the prediction.

<Estimation of spatter deposition amount and changes to the modeling plan>
Next, a procedure in which the nozzle maintenance support device 15 predicts the sputter deposition amount using the generated prediction model ML and changes the modeling program will be described.
FIG. 5 is a flowchart showing a procedure for changing the modeling program.
First, information on a modeling plan (modeling program) for a model to be modeled is input to the plan receiving unit 35 (S21). The information on the modeling program may be input by the control unit 13 shown in Fig. 1 or may be input by an operator. The plan receiving unit 35 outputs the information on the modeling program that has been received to the deposition amount estimation unit 37.

Similar to the information extraction unit 43 described above, the deposition amount estimation unit 37 extracts bead formation information including multiple movement paths and welding conditions for each movement path from the input molding program (S22). The extracted bead formation information is then input into the generated prediction model ML, and an estimated value for the sputter deposition amount corresponding to the bead formation information is obtained from the prediction model ML (S23).

Information about the estimated sputter deposition amount is output from the deposition amount estimation unit 37 to the determination unit 39. The determination unit 39 determines whether the estimated sputter deposition amount exceeds a predetermined reference value (S24). If the estimated value exceeds the reference value, the plan modification unit 41 modifies the molding program (S25).

FIG. 6 is an explanatory diagram showing an example of changes to the modeling program.
The specific changes to the shaping program by the plan modification unit 41 will be described using an example in which, for example, command codes for bead formation on movement path PA, bead formation on movement path PB, ..., bead formation on movement path PG, and bead formation on movement path PH are present in chronological order, as shown in Figure 6. Here, it is assumed that the determination unit 39 determines that the estimated sputter deposition amount calculated by the deposition amount estimation unit 37 exceeds the reference value after bead formation on movement path PG. In this case, the plan modification unit 41 adds a process (program) for cleaning the nozzle after bead formation on movement path PG to the shaping program. In other words, a program for cleaning the nozzle is added between movement path PG and movement path PH in the shaping program. The nozzle cleaning program added here is a pre-prepared program, and the same program can be used as is even when added at a timing other than the above. Furthermore, instead of the nozzle cleaning program, a nozzle replacement program for replacing the nozzle attached to the welding torch with a new nozzle may be used.

By repeatedly predicting the amount of spatter buildup and adding a nozzle cleaning program or nozzle replacement program based on the predicted value from the first movement pass to the final movement pass (S26, S27), a modeling program can be created in which a nozzle cleaning program or nozzle replacement program is added at the appropriate time. In this way, by calculating the predicted amount of spatter buildup from the planning stage and comparing and determining the predicted value with a reference value, the required number of maintenance tasks, their implementation timing, etc. can be systematically incorporated into the modeling plan.

The modified modeling program is output to the control unit 13 (S28), and the control unit 13 carries out modeling in accordance with the modified modeling program that has been input. This makes the nozzle less susceptible to the effects of spatter adhering to it, and prevents a deterioration in bead quality.

The reference value used by the determination unit 39 as the determination criterion may be a value based on actual measurements of the impact of spatter during previous welding. In this case, it becomes possible to take measures that correspond to the actual degree of impact of spatter. Multiple reference values may also be prepared. For example, if the predicted value exceeds a first reference, the nozzle is cleaned, and if the predicted value exceeds a second reference that is greater than the first reference, the nozzle is replaced. In this way, nozzle cleaning and replacement may be distinguished by setting different reference values.

As mentioned above, by preparing a nozzle replacement program separately from the nozzle cleaning program and inserting it into the modeling program as needed, welding can be performed without being affected by spatter, even in situations where spatter accumulation is difficult to remove with normal cleaning alone.

Furthermore, by inputting the welding conditions corresponding to each of multiple movement passes into the prediction model ML, it is possible to calculate the predicted value of the amount of spatter deposition that may occur for each movement pass. It is also possible to calculate the cumulative value of the predicted amount of spatter deposition for each movement pass. Analyzing the amount of spatter deposition for each movement pass individually in this way enables more accurate responses, which can contribute to improving bead quality.

Furthermore, the welding conditions used as learning data when generating the prediction model ML may be determined by referencing the build plan, but if there is past corresponding track record, information such as spatter scattering images or videos from that time may be added to the learning data as welding conditions. This is expected to further improve prediction accuracy. Also, by including the welding speed and filler metal feed rate in the bead formation information, the amount of spatter generated is reflected in the learning, and by including the tip-to-work distance, the relationship between spatter scattering and the torch position is reflected in the learning. As a result, it becomes possible to quantitatively predict spatter accumulation trends.

A prediction model was generated using the machine learning described above, and the results of estimating the amount of sputter deposition using this prediction model were compared with the results of actually measuring the amount of sputter that occurred.
FIG. 7 is a graph showing the distribution of the predicted value Wt2 by the prediction model relative to the actually measured value Wt1 of the sputter deposition amount.

The explanatory variables for the machine learning used to generate the predictive model were the filler metal M feed rate, welding speed, tip-to-work distance, weaving cycle, weaving width, and torch angle. A total of 33 sets of training data were used. The objective variable was the amount of spatter deposition (amount of deposition [mg] per second of welding). Gaussian process regression was used as the training algorithm, and K-cross validation was used as the accuracy verification method.

As a result, the correlation coefficient between the actual sputter deposition amount Wt1 and the predicted value Wt2 was 0.8517, indicating a high correlation between the two. Therefore, if this prediction model is used to predict sputter deposition amount, it will be possible to form a high-quality bead.

As such, the present invention is not limited to the above-described embodiments, and the scope of protection sought includes 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 disclosures in the specification and well-known technologies.

As described above, the present specification discloses the following:
(1) A learning device that generates, by machine learning, a prediction model for predicting an amount of spatter that accumulates on a nozzle of a welding torch when a filler metal is melted and solidified at the tip of the welding torch to form a bead, the learning device comprising:
an information extraction unit that extracts, from a molding plan for molding an object having a desired shape, bead formation information, which is information on a movement path of the welding torch and welding conditions including at least a welding speed, a feed rate of the filler metal supplied to the welding torch, and a distance from a tip of the nozzle to a formation position of the bead;
an accumulation amount measuring unit that measures an accumulation amount of spatter accumulated on the nozzle of the welding torch when the bead is formed;
a model generation unit that performs machine learning on the relationship between the bead formation information and the sputter deposition amount, and generates the prediction model that represents the relationship between the bead to be formed and the sputter deposition amount;
A learning device comprising:
This learning device uses a prediction model that has been machine-learned in response to bead formation information to determine the amount of spatter deposition on the nozzle, making it possible to quantitatively grasp the trend in spatter accumulation and more accurately predict the amount of spatter deposition.

(2) The learning device according to (1), wherein the welding conditions further include a torch angle of the welding torch.
This learning device can generate a prediction model that takes into account the fact that the amount of spatter deposition varies depending on the torch angle. As a result, the relationship between angles at which spatter deposition is favorable and unfavorable is reflected in the prediction results, improving prediction accuracy.

(3) The learning device according to (1) or (2), wherein the welding conditions include at least one of a weaving amplitude and a weaving frequency.
This learning device can generate a prediction model according to the amplitude and frequency of the weaving, even when the welding torch is weaving, enabling more accurate prediction of the amount of spatter deposition.

(4) The welding torch may further include an imaging unit that images the welding torch and spatter scattered around the welding torch,
The model generation unit associates image information captured by the imaging unit during the formation of the bead with the bead formation information of the bead and performs machine learning on the relationship with the sputter deposition amount to generate the prediction model.
According to this learning device, by using image information of the flying spatter, the progress of the spatter deposition, such as the direction of the flying spatter and the direction of the welding torch, can be reflected in the learning.

(5) A learning device according to any one of (1) to (4),
a plan receiving unit that receives information about the modeling plan;
a deposition amount estimation unit that extracts the bead formation information from the received information of the manufacturing plan, inputs the extracted bead formation information into the prediction model, and calculates an estimated value of the sputter deposition amount from the prediction model;
a determination unit that determines whether the estimated value of the sputter deposition amount exceeds a predetermined reference value;
a plan change unit that changes the shaping plan when it is determined that the estimated value exceeds the reference value;
A nozzle maintenance support device comprising:
This nozzle maintenance support device appropriately calculates an estimated spatter accumulation amount from the planning stage and compares it with a reference value, allowing the necessary number of maintenance work times and timing to be planned accordingly. This makes it possible to create a molding plan that prevents excessive spatter accumulation on the nozzle and maintains bead quality, even when beads are formed continuously and for long periods of time. This improves the efficiency of bead formation and increases the productivity of molded objects.

(6) The nozzle maintenance support device described in (5), wherein the plan change unit inserts a nozzle cleaning process or a nozzle replacement process before or after any of the bead forming processes of the multiple movement paths set in the molding plan.
This nozzle maintenance support device allows the timing of nozzle maintenance to be set at the planning stage, contributing to improved productivity.

(7) A nozzle maintenance support device according to (5) or (6),
a control unit that outputs a control signal for executing arc welding based on the changed modeling plan;
A welding control device comprising:
This welding control device enables welding control that can suppress the accumulation of spatter.

(8) The welding control device according to (7),
A welding robot that performs arc welding;
a nozzle station for cleaning or replacing the nozzle of the welding torch;
A welding device comprising:
With this welding device, spatter that adheres to the nozzle during welding can be removed by cleaning or replacing the nozzle at the appropriate time, allowing beads to be formed with consistently stable quality.

REFERENCE SIGNS LIST 11 Forming unit 13 Control unit 15 Nozzle maintenance support device 16 Welding control device 17 Welding torch 19 Welding robot 21 Robot driving unit 23 Filler metal supply unit 23a Reel 25 Welding power supply unit 27 Nozzle station 29 Imaging unit 31 Formed object 33 Learning device 35 Plan reception unit 37 Deposition amount estimation unit 39 Determination unit 41 Plan change unit 43 Information extraction unit 45 Deposition amount measurement unit 47 Model generation unit 100 Welding device B Bead BP Base plate M Filler metal

Claims (8)

A learning device that generates, by machine learning, a prediction model for predicting an amount of spatter that accumulates on a nozzle of a welding torch when a filler metal is melted and solidified at the tip of the welding torch to form a bead,
an information extraction unit that extracts, from a molding plan for molding an object having a desired shape, bead formation information, which is information on a movement path of the welding torch and welding conditions including at least a welding speed, a feed rate of the filler metal supplied to the welding torch, and a distance from a tip of the nozzle to a formation position of the bead;
an accumulation amount measuring unit that measures an accumulation amount of spatter accumulated on the nozzle of the welding torch when the bead is formed;
a model generation unit that performs machine learning on the relationship between the bead formation information and the sputter deposition amount to generate the prediction model that represents the relationship between the bead formation information and the sputter deposition amount corresponding to the bead to be formed;
A learning device comprising: The welding conditions further include a torch angle of the welding torch.
The learning device according to claim 1 . The welding conditions include at least one of an amplitude of a weaving scan and a frequency of the weaving scan.
The learning device according to claim 1 or 2. An imaging unit is further provided for imaging the welding torch and spatter scattered around the welding torch,
the model generation unit associates image information captured by the imaging unit during the formation of the bead with the bead formation information of the bead, and performs machine learning on the relationship with the sputter deposition amount to generate the prediction model.
The learning device according to any one of claims 1 to 3. A learning device according to any one of claims 1 to 4;
a plan receiving unit that receives information about the modeling plan;
a deposition amount estimation unit that extracts the bead formation information from the received information of the manufacturing plan, inputs the extracted bead formation information into the prediction model, and calculates an estimated value of the sputter deposition amount from the prediction model;
a determination unit that determines whether the estimated value of the sputter deposition amount exceeds a predetermined reference value;
a plan change unit that changes the shaping plan when it is determined that the estimated value exceeds the reference value;
A nozzle maintenance support device comprising: the plan modification unit inserts a nozzle cleaning process or a nozzle replacement process before or after any one of the bead forming processes of the plurality of movement paths set in the molding plan.
The nozzle maintenance support device according to claim 5. a nozzle maintenance support device according to claim 5 or 6;
a control unit that outputs a control signal for executing arc welding based on the changed modeling plan;
A welding control device comprising: The welding control device according to claim 7 ;
A welding robot that performs arc welding;
a nozzle station for cleaning or replacing the nozzle of the welding torch;
A welding device comprising: