JP2025005720A - Welding waveform control device and welding waveform control method - Google Patents

Abstract

To provide a welding waveform control device and welding waveform control method that can generate a welding waveform to obtain the desired welding result in a short time and easily while evaluating welding.SOLUTION: A welding waveform control device 100 includes: a welding waveform acquisition unit 110 that acquires the welding waveform during the welding; a welding evaluation unit 130 that evaluates the welding on the basis of the acquired welding waveform; an output unit 140 that outputs the evaluation result; an adjustment information acquisition unit 150 that acquires adjustment information of the desired welding for the evaluation result; and a welding waveform generation unit 160 that generates the welding waveform on the basis of the acquired welding waveform and the adjustment information of the desired welding.SELECTED DRAWING: Figure 2

Description

The present invention relates to a welding waveform control device and a welding waveform control method.

In recent years, many robots have come into use in industry. These robots are used, for example, for assembling, welding, and transporting electronic and mechanical parts, and are helping to streamline and automate factory production lines.

For welding robots, there is a demand to shorten welding times and tact times to improve the efficiency of production lines, but at the same time, welding quality must be ensured, taking into account the bead finish in arc welding.

Generally, the desired welding results (bead finish, etc.) are obtained by adjusting welding conditions such as welding speed, welding time, welding current, and welding voltage.

For example, Patent Document 1 discloses a technology that shortens cycle time and stabilizes welding results by appropriately adjusting welding conditions using a neural network.

JP 2021-159969 A

However, while the technology disclosed in Patent Document 1 can shorten the cycle time and stabilize the welding results in a good condition, users have various detailed requests regarding the finish of the welding, and further improvements are required.

Methods of adjusting welding waveforms, such as the welding current waveform and/or welding voltage waveform, during welding to meet the specific needs of a user regarding the desired welding results are also known. Adjusting such welding waveforms is highly dependent on the skill of the worker, and even for an experienced worker, it can take a huge amount of time to achieve the desired welding results, which is not easy.

The present invention aims to provide a welding waveform control device and a welding waveform control method that can easily generate a welding waveform to obtain a desired welding result in a short time while evaluating the welding.

A welding waveform control device according to one aspect of the present invention includes a welding waveform acquisition unit that acquires a welding waveform during welding, a welding evaluation unit that evaluates the welding based on the acquired welding waveform, an output unit that outputs the evaluation result, an adjustment information acquisition unit that acquires desired welding adjustment information for the evaluation result, and a welding waveform generation unit that generates a welding waveform based on the acquired welding waveform and the desired welding adjustment information.

According to this aspect, the welding evaluation unit evaluates the welding based on the welding waveform at the time of welding acquired by the welding waveform acquisition unit. The output unit outputs the result evaluated by the welding evaluation unit, and the adjustment information acquisition unit acquires desired welding adjustment information for that result. Then, the welding waveform generation unit generates a welding waveform based on the welding waveform acquired by the welding waveform acquisition unit and the desired welding adjustment information. This makes it possible to easily generate a welding waveform for obtaining a desired welding result in a short time while evaluating the welding. As a result, the user can perform the welding desired by rewelding using the generated welding waveform.

In the above aspect, the welding evaluation unit may further include a feature extraction unit that extracts feature values based on the acquired welding waveform, the welding evaluation unit evaluates the welding based on the extracted feature values, and the welding waveform generation unit may generate a welding waveform based on the extracted feature values and the desired welding adjustment information.

According to this aspect, the feature extraction unit extracts feature values based on the welding waveform acquired by the welding waveform acquisition unit, the welding evaluation unit evaluates the welding based on the feature values, and the welding waveform generation unit generates a welding waveform based on the feature values and the desired welding adjustment information. This allows the welding to be appropriately evaluated and the welding waveform desired by the user to be generated while reducing the processing load in the welding evaluation and welding waveform generation. Furthermore, the output welding evaluation results and factors and influences on the generated welding waveform can be confirmed and analyzed.

In the above aspect, the welding evaluation unit may evaluate the welding based on the extracted features using a first machine learning model that uses the features as input data and outputs an evaluation of the welding as output data.

According to this aspect, the welding evaluation unit evaluates the welding using a machine learning model based on the feature amounts, and the output unit outputs the evaluation results. Since feature amounts are extracted from the welding waveform and the welding is evaluated using a machine learning model based on these, the processing load can be reduced while the welding is appropriately evaluated.

In the above aspect, the welding waveform generation unit may generate a welding waveform based on the extracted features and the desired welding adjustment information using a second machine learning model that uses features based on the welding waveform and welding adjustment information as input data and waveform control parameters for generating the welding waveform as output data.

According to this aspect, the welding waveform generation unit uses the feature quantities and welding adjustment information as input data and generates a welding waveform using a machine learning model with waveform control parameters as output data. As a result, by reducing the dependency on the skill of the worker in adjusting the welding waveform and reducing the need for repeated welding and adjustment, a welding waveform for obtaining a desired welding result can be easily generated in a short time.

The welding waveform control method according to one aspect of the present invention is a welding waveform control method executed by a welding waveform control device that controls the welding waveform during welding, and includes a welding waveform acquisition step of acquiring the welding waveform during welding, a welding evaluation step of evaluating the welding based on the acquired welding waveform, an output step of outputting the evaluation result, an adjustment information acquisition step of acquiring desired welding adjustment information for the evaluation result, and a welding waveform generation step of generating a welding waveform based on the acquired welding waveform and the desired welding adjustment information.

According to this aspect, in the welding evaluation step, the welding is evaluated based on the welding waveform obtained in the welding waveform acquisition step. In the output step, the result of the evaluation in the welding evaluation step is output, and in the adjustment information acquisition step, adjustment information for the desired welding corresponding to that result is acquired. Then, in the welding waveform generation step, a welding waveform is generated based on the welding waveform obtained in the welding waveform acquisition step and the adjustment information for the desired welding. This makes it possible to easily generate a welding waveform for obtaining a desired welding result in a short time while evaluating the welding. As a result, the user can perform the desired welding by rewelding using the generated welding waveform.

The present invention aims to provide a welding waveform control device and a welding waveform control method that can easily generate a welding waveform to obtain a desired welding result in a short time while evaluating the welding.

1 is a system schematic diagram showing a welding robot system 10 according to an embodiment of the present invention. 1 is a functional block diagram showing each function of a welding waveform control device 100 according to an embodiment of the present invention. 4 is a diagram showing feature quantities extracted from a welding current waveform acquired by a welding waveform acquisition unit 110. FIG. 13 is a diagram showing how an evaluation result is output using a machine learning model (first machine learning model) with features extracted from a welding current waveform as input data. FIG. FIG. 13 is a diagram showing how a welding current waveform is used as input data, and multiple evaluation items related to welding are output as output data using one machine learning model. A figure showing how waveform control parameters are output using a machine learning model (second machine learning model) with features extracted from a welding current waveform and desired welding adjustment information as input data. FIG. 13 is a diagram showing how a welding current waveform and welding adjustment information are used as input data, and a waveform control parameter is output using one machine learning model. FIG. 13 is a diagram showing an adjusted welding current waveform generated by a welding waveform generating unit 160. 1 is a flowchart showing a welding waveform control method M100 executed by a welding waveform control device 100 according to an embodiment of the present invention. 4 is a diagram showing feature amounts extracted from a welding voltage waveform acquired by a welding waveform acquisition unit 110. FIG.

The following describes an embodiment of the present invention in detail with reference to the drawings. Note that the embodiment described below is merely a specific example for implementing the present invention, and is not intended to limit the interpretation of the present invention. In addition, to facilitate understanding of the description, the same components in each drawing are given the same reference numerals as much as possible, and duplicate descriptions may be omitted.

<One embodiment>
[Welding robot system overview]
1 is a system schematic diagram showing a welding robot system 10 according to an embodiment of the present invention. In FIG. 1, the welding robot system 10 includes a welding robot 20, a teach pendant 30, a robot controller 40, and a power source 50.

The welding robot 20 is connected to the robot control device 40 via a cable, and performs arc welding based on operation commands from the robot control device 40. The welding robot 20 is equipped with a welding torch 21 at the tip of its arm, and performs arc welding by feeding a welding wire from the tip of the welding torch 21 and generating an arc between the welding wire and the base material (workpiece made of a metal material) to be welded.

The welding torch 21 is connected to a power source 50 via a cable, and receives a welding voltage and a welding current to the welding wire. In arc welding, when the welding wire is momentarily brought into contact with a metal material and electricity is passed through it, an arc discharge occurs between the welding wire and the metal material, and the heat of the generated arc melts the welding wire and the metal material, thereby performing welding.

The teach pendant 30 accepts input from the worker performing the welding work regarding welding-related teaching information for the welding robot 20. The worker inputs optimal welding-related teaching information using the teach pendant 30 while checking the state of the arc.

Here, welding-related instruction information refers to information related to the welding performed by the welding robot 20, and includes instruction information that instructs the operation of the welding robot 20 and welding execution conditions.

The instruction information for the welding robot 20 includes information regarding the operation of the arm of the welding robot 20, information regarding the position and posture of the welding robot 20, information regarding the protruding length of the welding wire fed from the tip of the welding torch 21, etc.

The welding conditions also include the welding voltage applied to the welding wire, the value of the welding current flowing through the welding wire, the welding speed and stop time representing the movement speed of the welding torch 21 in the direction of the welding line during welding, and may further include the wire feed speed controlled by the wire feed device.

The robot controller 40 is a device that controls the welding robot 20. The robot controller 40 is connected to the teach pendant 30 and can acquire welding-related teaching information input to the teach pendant 30. The robot controller 40 controls the welding robot 20 and the power source 50 based on the welding-related teaching information.

The power supply 50 is connected to the welding robot 20 via a cable and supplies welding voltage and welding current to the welding torch 21 in the welding robot 20 based on commands from the robot control device 40.

In FIG. 1, the teach pendant 30 is connected to the robot control device 40 via a cable, but it may be connected wirelessly. That is, the teach pendant 30 and the robot control device 40 may be provided with a communication unit that performs wireless communication. By connecting the robot control device 40 and the teach pendant 30 wirelessly, the worker can input welding-related teaching information while moving freely without being bothered by the presence of a cable or being restricted in the range of movement by the length of the cable.

In this way, in the welding robot system 10, arc welding is performed by operating the welding robot 20 using the robot control device 40.

[Configuration of welding waveform control device]
Hereinafter, a welding waveform control device that controls the welding waveform during welding performed as described above will be described.

Figure 2 is a functional block diagram showing each function of the welding waveform control device 100 according to one embodiment of the present invention. As shown in Figure 2, the welding waveform control device 100 includes a welding waveform acquisition unit 110, a feature extraction unit 120, a welding evaluation unit 130, an output unit 140, an adjustment information acquisition unit 150, a welding waveform generation unit 160, and a memory unit 170. All or part of the units having these functions in the welding waveform control device 100 may be provided in the above-mentioned welding robot 20, teach pendant 30, robot control device 40, other peripheral devices, and computers such as personal computers and display devices such as monitors, and the welding waveform control device 100 may be configured by the units having these functions.

The welding waveform acquisition unit 110 acquires the welding waveform during welding. For example, the welding waveform acquisition unit 110 acquires the welding current waveform and/or the welding voltage waveform during the period during which welding is performed by the welding robot 20. The period during which welding is performed may be at predetermined time intervals or for each welding operation in a predetermined section.

The feature extraction unit 120 extracts features based on the welding waveform acquired by the welding waveform acquisition unit 110. For example, the feature extraction unit 120 may extract predefined elements as features from the welding current waveform and/or the welding voltage waveform.

Figure 3 is a diagram showing feature quantities extracted from the welding current waveform acquired by the welding waveform acquisition unit 110. As shown in Figure 3, the following elements are extracted as feature quantities from the welding current waveform: (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time.

When extracting each of these elements as features from the welding current waveform, the features may be extracted from the welding current waveform, for example, at a predetermined time during the stable period excluding the start and end of welding. At the start and end of welding, different control is used than during the stable period, and the welding current waveform is different, so there is a possibility that appropriate features cannot be extracted.

Furthermore, each element extracted as a feature may be calculated using its average value, standard deviation, and variance. For example, the pulse interval (c) shown in FIG. 3 is defined as one cycle, and the elements (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time may be extracted from a welding current waveform for multiple cycles, and the average value of each may be calculated to determine the feature value of each element. This reduces the variation in each element, and allows each element to be more appropriately extracted as a feature.

In addition, the elements are defined as (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time, but are not limited to these. For example, only one or more of these elements may be extracted as features, or other elements (e.g., rise slope, fall slope, short circuit current, short circuit recovery current, short circuit time, number of short circuits, etc.) may be extracted as features as long as they can be extracted from the welding current waveform and can affect the evaluation of welding. The elements to be extracted may be selectable and set by the user in advance.

The feature extraction unit 120 may also extract features using a machine learning model. For example, the welding current waveform may be used as input data, and a machine learning model may be used to output each of the above-mentioned elements such as (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time as features. Here, the machine learning model may also be used to determine which elements to output. By using the fed-back trained model, each element can be extracted as a more appropriate feature.

The welding evaluation unit 130 evaluates the welding using a machine learning model (first machine learning model) based on the features extracted by the feature extraction unit 120. For example, the welding evaluation unit 130 evaluates the welding using the machine learning model with the input data being each of the elements (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time, which are the features described in FIG. 3.

Figure 4 is a diagram showing how the feature quantities extracted from the welding current waveform are used as input data, and the evaluation results are output using a machine learning model (first machine learning model). As shown in Figure 4, the input data are each of the elements (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time, and the machine learning model is used to output evaluation results for multiple evaluation items related to welding, such as (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount. For each evaluation item, for example, a score on a five-level scale from "1" to "5" ("1" being the worst, "3" being the standard, and "5" being the best) may be output as the evaluation result, or further, a single score may be output by combining them as the evaluation result.

The machine learning model used here is not particularly limited, but for example, LightGBM, random forest, and support vector machine reduce the processing load compared to algorithms using neural networks.

In addition, when multiple evaluation results are output as shown in FIG. 4, a machine learning model capable of outputting multiple evaluation results may be used, or multiple machine learning models each outputting one evaluation result may be used.

The evaluation items are (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount, but are not limited to these. For example, only one or more of these elements may be used as the evaluation items, or other evaluation items (e.g., penetration, blowholes, etc.) that can evaluate welding may be output as evaluation items. The user may be able to select and set in advance which evaluation items are to be output, in which case it is sufficient to have a machine learning model corresponding to them.

Furthermore, the welding waveform control device 100 may be equipped with a welding condition acquisition unit. The welding condition acquisition unit acquires the welding conditions (welding conditions set by the operator) when the welding robot 20 performs welding. The welding conditions include, for example, the welding current value, welding voltage value, welding speed, stop time, arc generation time, arc stop time, pitch interval, and cooling time, and may also include the wire feed speed controlled by the wire feed device. The values of the welding conditions acquired here may be actual measured values, or, if it is difficult to acquire measured values, they may be setting values set by the operator or output command values from the welding robot 20.

The welding evaluation unit 130 may evaluate the welding using a machine learning model based on the features extracted by the feature extraction unit 120 and the welding conditions acquired by the welding condition acquisition unit. Specifically, in the machine learning model shown in FIG. 4, in addition to the features extracted from the welding current waveform, the welding conditions acquired by the welding condition acquisition unit (for example, welding current value, welding voltage value, welding speed, stop time, arc generation time, arc stop time, welding execution conditions such as pitch interval and cooling time) may be used as input data, and multiple evaluation items related to welding such as (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount may be output as welding evaluation results. For this, it is sufficient to have a machine learning model that uses the features extracted by the feature extraction unit 120 and the welding conditions acquired by the welding condition acquisition unit as input data and outputs multiple evaluation items related to welding such as (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount. As a result, welding can be evaluated more appropriately.

Note that here, the feature extraction unit 120 extracts features from the welding current waveform, and the welding evaluation unit 130 evaluates the welding using a machine learning model with the extracted features as input data and multiple evaluation items related to welding as output data. In addition, when the feature extraction unit 120 extracts features, this also includes extracting features using a machine learning model with the welding current waveform as input data, but for example, the welding may be evaluated using one machine learning model with the welding current waveform as input data and multiple evaluation items related to welding as output data.

Figure 5 is a diagram showing how a welding current waveform is used as input data and a single machine learning model is used to output multiple evaluation items related to welding. As shown in Figure 5, the welding current waveform is used as input data and a machine learning model is used to output multiple evaluation items related to welding, such as (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount, as welding evaluation results.

Here, one machine learning model is, for example, a recurrent neural network (RNN) such as a long short-term memory (LSTM), and includes a feature extraction layer. The feature extraction layer may be configured to be able to extract the following elements as intermediate data: (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time.

For example, a machine learning model may be constructed using a fully connected layer in a feature extraction layer in which the welding current waveform is used as input data and feature amounts (each element of (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time) are extracted as data, and the machine learning model may output multiple evaluation items related to welding, such as (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount.

Note that, although RNN is used as one machine learning model here, this is not limited to this, and as long as welding can be evaluated using the welding current waveform as input data and multiple evaluation items related to welding as output data, for example, LightGBM, random forest, and support vector machine may also be used as one machine learning model.

Returning to the explanation of FIG. 2, the output unit 140 outputs the results evaluated by the welding evaluation unit 130. For example, the results may be output to a monitor or to the display screen of a tablet or dedicated terminal.

The adjustment information acquisition unit 150 acquires adjustment information for the desired welding. For example, the adjustment information acquisition unit 150 receives adjustment information for the welding desired by a user who has confirmed the welding results performed by the welding robot 20 (the welding evaluation results output by the output unit 140), and acquires the adjustment information for the welding.

Specifically, when the user checks the welding result performed by the welding robot 20 (the welding evaluation result output by the output unit 140), if the user wants to make the bead width a little wider, the user inputs the adjustment item "bead width" and the adjustment amount "+1" as the adjustment information for the desired welding, and the adjustment information acquisition unit 150 acquires the adjustment information for the welding. As an example of the input mode of the adjustment amount, when the user wants to make the bead width a little wider, the adjustment amount "+1" is given, but if the user wants to make it wider, it may be "+2" or "+3", and if the user wants to make it thinner, it may be "-1", "-2", "-3", etc. according to the degree of the adjustment. The adjustment amount per step may be preset according to the welding object. Also, the user may input "+10%", "-10%", "+1 cm", "-1 cm", etc., and the adjustment information acquisition unit 150 may acquire a more specific adjustment amount.

Note that the adjustment item is not limited to bead width, and may be, for example, bead height, spatter amount, undercut, penetration, blowholes, etc. For example, if the user wishes to reduce the amount of spatter, the user sets the adjustment item to "spatter amount" and inputs an adjustment amount of "-1", "-2", "-3", "-10%", "-20%", "-30%", etc. according to the degree of reduction, and the adjustment information acquisition unit 150 acquires adjustment information for that welding (adjustment item "spatter amount", adjustment amount "-1", etc.). As a specific example, if the welding evaluation unit 130 evaluates the amount of spatter as "2" out of a five-level rating ("1" very high, "2" high, "3" normal, "4" low, "5" very low), and the user wishes to reduce the amount of spatter, the user may input the adjustment item "spatter amount" and adjustment amount "-1 (meaning to reduce)" as welding adjustment information, or may input the adjustment item "spatter amount" and rating "3 (meaning "3" on a five-level rating scale, normal). In other words, the user may input the difference from the evaluation of the welding by the welding evaluation unit 130, or may input the desired target value (evaluation value).

Furthermore, the user can select multiple adjustment items. For example, if the user wants to make the bead width a little wider and reduce the amount of spatter, the user can input the adjustment item "bead width" and the adjustment amount "+1", and further input the adjustment item "amount of spatter" and the adjustment amount "-1". The adjustment information acquisition unit 150 then acquires the multiple adjustment items and their corresponding adjustment amounts as welding adjustment information.

The welding waveform generation unit 160 generates a welding waveform using a machine learning model (second machine learning model) based on the features extracted by the feature extraction unit 120 and the welding adjustment information acquired by the adjustment information acquisition unit 150.

Here, the machine learning model takes as input data the feature quantities based on the welding waveform and welding adjustment information, and outputs waveform control parameters for generating a welding waveform, which are stored in advance in the storage unit 170. For example, the machine learning model takes as input data the elements (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time as the feature quantities described in FIG. 3, as well as welding adjustment information including adjustment items and adjustment amounts for the adjustment items, and outputs waveform control parameters for generating a welding current waveform.

The adjustment items as input data are, for example, bead width, bead height, amount of spatter, undercut, penetration, blowholes, etc., and the adjustment amount is indicated by "-3" to "+3" etc. corresponding to each adjustment item. The waveform control parameters as output data are parameters for generating a welding waveform, and are the elements of "adjusted pulse current", "adjusted base current", "adjusted pulse interval", "adjusted pulse time", "adjusted rise time", and "adjusted fall time" adjusted from the welding current waveform shown in Figure 3 as input data. More specifically, the waveform control parameters are parameters for configuring the shape of the welding waveform.

Figure 6 is a diagram showing how the feature quantities extracted from the welding current waveform and the desired welding adjustment information are used as input data, and the machine learning model (second machine learning model) is used to output waveform control parameters. As shown in Figure 6, the input data are the elements of (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time, and the welding adjustment information desired by the user (adjustment item "bead width" and the adjustment amount "+1" for that adjustment item), and the machine learning model is used to output the adjusted elements of (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time as waveform control parameters.

Here, the user who has checked the welding results (the welding evaluation results output by the output unit 140) performed by the welding robot 20 wants to make the bead width a little thicker, and the adjustment information acquisition unit 150 acquires the adjustment information for the welding. Then, the welding waveform generation unit 160 uses a machine learning model based on the feature amounts (each element of (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time) extracted by the feature amount extraction unit 120 and the welding adjustment information (adjustment item "bead width" and adjustment amount "+1") acquired by the adjustment information acquisition unit 150 to generate a welding current waveform using each element of the adjusted (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time as waveform control parameters. In other words, the welding current waveform generated by the welding waveform generation unit 160 is a welding current waveform that has been adjusted from the welding current waveform acquired by the welding waveform acquisition unit 110 so that the bead width is slightly wider.

The machine learning model used here is not particularly limited, but for example, LightGBM, random forest, and support vector machine reduce the processing load compared to algorithms using neural networks.

In addition, when multiple elements are output as waveform control parameters as shown in FIG. 6, a machine learning model capable of outputting multiple elements may be used, or multiple machine learning models each outputting one element may be used.

Furthermore, when the welding waveform control device 100 is provided with the above-mentioned welding condition acquisition unit, in addition to the feature amount extracted by the feature amount extraction unit 120 and the welding adjustment information acquired by the adjustment information acquisition unit 150, a machine learning model may be used to generate a welding current waveform based on the welding conditions acquired by the welding condition acquisition unit, with each element adjusted from the welding current waveform during welding as a waveform control parameter. Specifically, in the machine learning model shown in FIG. 6, in addition to the feature amount extracted from the welding current waveform during welding and the welding adjustment information, the welding conditions acquired by the welding condition acquisition unit (e.g., welding current value, welding voltage value, welding speed, stop time, arc generation time, arc stop time, pitch interval, cooling time, etc.) are used as input data, and a welding current waveform is generated using each element of the adjusted (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time as waveform control parameters. This can be achieved by providing a machine learning model that uses the features extracted by the feature extraction unit 120, the welding adjustment information acquired by the adjustment information acquisition unit 150, and the welding conditions acquired by the welding condition acquisition unit as input data, and outputs the adjusted elements (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time as waveform control parameters.

Here, the feature extraction unit 120 extracts features from the welding current waveform, the adjustment information acquisition unit 150 acquires the welding adjustment information desired by the user, and the welding waveform generation unit 160 generates a welding current waveform using the extracted features and the acquired welding adjustment information as input data and a machine learning model with waveform control parameters as output data. In addition, when the feature extraction unit 120 extracts features, this also includes extracting features using a machine learning model with the welding current waveform as input data, but for example, a welding current waveform may be generated using one machine learning model with the welding current waveform and welding adjustment information as input data and waveform control parameters as output data.

Figure 7 is a diagram showing how a welding current waveform and welding adjustment information are used as input data, and a single machine learning model is used to output waveform control parameters. As shown in Figure 7, the welding current waveform and welding adjustment information are used as input data, and a machine learning model is used to output each element, such as (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time, as waveform control parameters.

Here, one machine learning model is, for example, a recurrent neural network (RNN) such as a long short-term memory (LSTM), and includes a feature extraction layer. The feature extraction layer may be configured to be able to extract the following elements as intermediate data: (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time.

For example, the machine learning model may be configured using a fully connected layer in a feature extraction layer in which the welding current waveform and welding adjustment information are used as input data, and the features of the welding current waveform (each of the elements (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time) are used as extracted data, and the machine learning model may output each element (waveform control parameters) such as (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time.

Note that, although RNN is used as one machine learning model here, this is not limited to this, and as long as a welding current waveform can be generated using the welding current waveform and welding adjustment information as input data and the waveform control parameters as output data, for example, LightGBM, random forest, and support vector machine may also be used as one machine learning model.

In this way, the welding waveform generated by the welding waveform generating unit 160 may be output, for example, to a monitor or to the display screen of a tablet or dedicated terminal.

Figure 8 is a diagram showing the adjusted welding current waveform generated by the welding waveform generation unit 160. As shown in Figure 8, a welding current waveform adjusted to make the bead width slightly wider is generated from the welding current waveform acquired by the welding waveform acquisition unit 110. This indicates that the user can check the welding result performed by the welding robot 20 (the welding evaluation result output by the output unit 140) and input the desired welding adjustment information (adjustment item "bead width" and adjustment amount "+1"), and an adjusted welding current waveform that has been appropriately adjusted to meet the user's requirements is generated.

The welding current waveform whose features have been extracted by the feature extraction unit 120 and the welding current waveform generated by the welding waveform generation unit 160 may be displayed on the same screen. The user can check and compare the welding current waveforms before and after adjustment, and can recognize how the welding current waveform changes by inputting the desired welding adjustment information (adjustment item "bead width" and adjustment amount "+1"), for example.

Returning to the explanation of FIG. 2, the memory unit 170 stores various set conditions and machine learning models. For example, the memory unit 170 stores information related to the features extracted by the feature extraction unit 120 from the welding waveform acquired by the welding waveform acquisition unit 110, and the feature extraction unit 120 refers to this information. In addition, the welding evaluation unit 130 stores the machine learning models shown in FIG. 4 and FIG. 5 stored in the memory unit 170, and evaluates welding using the machine learning models.

In addition, the memory unit 170 stores information related to the welding adjustment information acquired by the adjustment information acquisition unit 150, such as regulations related to adjustment items and the adjustment amounts for the adjustment items. In addition, the welding waveform generation unit 160 generates a welding waveform from the waveform control parameters output using the machine learning models shown in Figures 6 and 7 stored in the memory unit 170.

Note that the items stored in the memory unit 170 are not limited to these, and other data required for the processing of the welding waveform control device 100 may also be stored, and data may also be stored temporarily during the processing process.

[Welding waveform control method]
Next, a welding waveform control method according to one embodiment of the present invention will be described in detail. Fig. 9 is a flow chart showing a welding waveform control method M100 executed by welding waveform controller 100 according to one embodiment of the present invention. As shown in Fig. 9, welding waveform control method M100 includes steps S110 to S160, and each step is executed by a processor included in welding waveform controller 100.

In step S110, the welding waveform acquisition unit 110 acquires the welding waveform during welding (welding waveform acquisition step). As a specific example, the welding waveform acquisition unit 110 acquires the welding current waveform during welding performed by the welding robot 20.

In step S120, the feature extraction unit 120 extracts features based on the welding waveform acquired in step S110 (feature extraction step). As a specific example, the feature extraction unit 120 extracts each of the following elements from the welding current waveform as features, as shown in FIG. 3: (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time.

In step S130, the welding evaluation unit 130 evaluates the welding using a machine learning model based on the feature quantities extracted in step S120 (welding evaluation step). As a specific example, as shown in FIG. 4, the welding evaluation unit 130 uses the elements (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time as input data, and outputs, using a machine learning model, evaluation results for multiple evaluation items related to welding, such as (A) bead width, (B) bead height, (C) undercut, and (D) spatter amount.

In this embodiment, the welding evaluation unit 130 evaluates the welding using a machine learning model based on the feature quantities, but this is not limited to this. For example, the welding evaluation unit 130 may evaluate the welding using a preset calculation formula (function) or conversion table based on the welding current waveform and/or the feature quantities extracted from the welding current waveform shown in FIG. 3, without using a machine learning model.

In step S140, the output unit 140 outputs the results of the evaluation in step S130 (output step). As a specific example, the output unit 140 outputs the evaluation items related to welding, such as (A) bead width, (B) bead height, (C) undercut, and (D) amount of spatter, evaluated in step S130, to a monitor or to the display screen of a tablet or dedicated terminal.

In step S150, the adjustment information acquisition unit 150 acquires adjustment information for the desired weld (adjustment information acquisition step). As a specific example, if the user checks the welding result performed by the welding robot 20 (the welding evaluation result output in step S140) and wants to make the bead width a little wider, the user inputs, for example, the adjustment item "bead width" and the adjustment amount "+1" as the adjustment information for the desired weld. Then, the adjustment information acquisition unit 150 acquires the adjustment information for that weld (adjustment item "bead width" and adjustment amount "+1").

Furthermore, the user can select multiple adjustment items. For example, if the user wants to make the bead width a little wider and reduce the amount of spatter, the user can input the adjustment item "bead width" and the adjustment amount "+1", and further input the adjustment item "amount of spatter" and the adjustment amount "-1". The adjustment information acquisition unit 150 can acquire multiple adjustment items and their corresponding adjustment amounts as adjustment information.

In step S160, the welding waveform generating unit 160 generates a welding waveform using a machine learning model based on the feature amount extracted in step S120 and the welding adjustment information acquired in step S150 (welding waveform generating step). As a specific example, as shown in FIG. 6, the welding waveform generating unit 160 uses the elements of (a) pulse current, (b) base current, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time, as well as the welding adjustment information desired by the user (adjustment item "bead width" and adjustment amount "+1") as input data, and generates a welding current waveform using the adjusted elements of (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time as waveform control parameters using the machine learning model.

When multiple adjustment items are acquired in step S150 with respect to the welding adjustment information, a machine learning model may be used to generate a welding waveform, with the feature values extracted in step S120 and the adjustment information including the multiple adjustment items acquired in step S150 and their corresponding adjustment amounts as input data. The machine learning model used here may be set in advance to take adjustment information including the multiple adjustment items and their corresponding adjustment amounts as input data.

Then, the welding robot 20 performs welding using the welding current waveform (waveform control parameters) generated in step S160. This allows the user to perform the welding desired.

It is also conceivable that the user may wish to further adjust the welding. Typically, after rewelding, the above-described welding waveform control method M100 can be repeated based on the welding waveform at the time of rewelding, but, for example, the welding current waveform may be adjusted before rewelding.

The welding current waveform generated in step S160 is displayed by a display unit (not shown) (display step), and a judgment unit (not shown) judges whether the welding current waveform should be further adjusted (welding waveform adjustment judgment step). As a specific example, the display unit may output the welding current waveform generated in step S160 using the adjusted elements (A) adjusted pulse current, (B) adjusted base current, (C) adjusted pulse interval, (D) adjusted pulse time, (E) adjusted rise time, and (F) adjusted fall time as waveform control parameters to a monitor or to a display screen of a tablet or dedicated terminal. Furthermore, the display unit may display the welding current waveform from which the feature values were extracted in step S120 and the welding current waveform generated in step S160 on the same screen.

Next, the determination unit determines that the welding current waveform generated in step S160, displayed on a monitor or the like, is to be adjusted if the waveform is to be further adjusted. For example, the welding current waveform generated in step S160 is displayed on a monitor or the like for a user to check, and if the user wishes to further adjust the welding current waveform, the determination unit accepts welding adjustment information from the user and determines that the welding current waveform is to be adjusted. Here, the adjustment items may be the same as the previous adjustment items, or may be different.

If the determination unit determines that the welding current waveform should be adjusted, the process returns to step S120, and the feature extraction unit 120 extracts features based on the welding current waveform generated in step S160. Then, in step S130, the welding evaluation unit 130 evaluates the welding using a machine learning model based on the features extracted in step S120, and in step S140, the output unit 140 outputs the results of the evaluation in step S130. The evaluation of the welding here is based on the welding current waveform during actual welding when the welding is performed, or the welding current waveform generated in step S160 (without actually performing welding).

Next, in step S150, the adjustment information acquisition unit 150 acquires the desired welding adjustment information that the user wishes to further adjust, and in step S160, the welding waveform generation unit 160 generates a welding current waveform using a machine learning model based on the feature amount extracted in step S120 and the welding adjustment information acquired in step S150. Then, when displaying the generated welding current waveform, the previous welding current waveform, or the previous one if multiple repetitions have been performed, may be displayed on the same screen.

In addition, if the user selects multiple items as adjustment items and, for example, inputs the adjustment item "bead width" and the adjustment amount "+1" and also the adjustment item "spatter amount" and the adjustment amount "-1," the welding current waveform may be adjusted the first time to correspond to the adjustment item "bead width" and the adjustment amount "+1" (step S160), and then the judgment unit may judge that the welding current waveform should be adjusted further (returning to step S120), and the welding current waveform may be adjusted the second time to correspond to the adjustment item "spatter amount" and the adjustment amount "-1."

For example, if the user has input the adjustment item "bead width" and adjustment amount "+3", the welding current waveform is adjusted the first time to correspond to the adjustment item "bead width" and adjustment amount "+1" (step S160), and then the judgment unit judges that the welding current waveform should be adjusted (return to step S120). The welding current waveform is adjusted the second time to correspond to the adjustment item "bead width" and adjustment amount "+1" as in the first time, and the welding current waveform is also adjusted the third time to correspond to the adjustment item "bead width" and adjustment amount "+1". In other words, the adjustment item "bead width" and adjustment amount "+1" are repeated three times.

Here, when the user inputs the adjustment items and adjustment amounts multiple times, the difference from the previously generated welding current waveform (evaluation result) may be input as the adjustment information, or the accumulated amount from the original welding current waveform (evaluation result) may be input as the adjustment information, or a reference value (absolute value) such as an evaluation value may be input as the adjustment information.

As a specific example, if the spatter amount is rated as "2: high" on a 5-point scale for welding evaluation, an adjustment amount of "-1 (meaning to reduce)" to reduce the spatter amount is input the first time, resulting in a rating of "3: normal." If it is desired to reduce the spatter amount even further, an adjustment amount to reduce the spatter amount can be input the second time, but here, the adjustment amount can be "-1" as the difference from the first adjustment, or "-2" as the cumulative amount. In this way, after two adjustments, the rating becomes "4: low."

Also, the user may input a target value (evaluation value) of "3: normal" the first time and a target value (evaluation value) of "4: low" the second time.

As described above, according to the welding waveform control device 100 and the welding waveform control method M100 according to one embodiment of the present invention, the feature extraction unit 120 extracts feature values based on the welding current waveform acquired by the welding waveform acquisition unit 110 during welding. The welding evaluation unit 130 evaluates the welding using the feature values as input data and a machine learning model with multiple evaluation items related to welding as output data, and the output unit 140 outputs the evaluation result to a monitor or the like. The adjustment information acquisition unit 150 acquires adjustment information (adjustment items and adjustment amounts) of welding desired by the user, and the welding waveform generation unit 160 generates a welding current waveform using a machine learning model based on the feature values extracted by the feature extraction unit 120 and the welding adjustment information acquired by the adjustment information acquisition unit 150. This makes it possible to easily generate a welding current waveform for obtaining a desired welding result in a short time while evaluating the welding. As a result, the welding robot 20 performs welding desired by the user using the welding current waveform (waveform control parameters) generated by the welding waveform generation unit 160.

In addition, if the elements extracted by the feature extraction unit 120 are used as input data for each machine learning model, the elements can be confirmed, allowing the worker to, for example, confirm and analyze the factors and effects on the output welding evaluation results and the generated welding current waveform (waveform control parameters).

In addition, in this embodiment, the welding waveform acquired by the welding waveform acquisition unit 110 during welding is the welding current waveform, but this is not limited to this and may be, for example, a welding voltage waveform. As the welding waveform during welding, a welding current waveform may be acquired, a welding voltage waveform may be acquired, or even a welding current waveform and a welding voltage waveform may be acquired.

Figure 10 is a diagram showing feature quantities extracted from the welding voltage waveform acquired by the welding waveform acquisition unit 110. As shown in Figure 10, in the case of a welding voltage waveform, the feature quantity extraction unit 120 may extract each of the following elements from the welding voltage waveform as feature quantities: (a) pulse voltage, (b) base voltage, (c) pulse interval, (d) pulse time, (e) rise time, and (f) fall time. The welding evaluation unit 130 then uses these elements as input data to evaluate the welding using a machine learning model, and the welding waveform generation unit 160 uses these elements as input data to generate a welding voltage waveform (waveform control parameters) using a machine learning model.

In addition, in this embodiment, the welding waveform control device 100 has been described as an example of a case where welding is performed by a welding robot 20, but this is not limited thereto, and can also be applied, for example, to a case where an operator performs welding manually. When welding manually, for example, the welding torch is provided with a sensor such as a gyro sensor, and the welding speed, torch angle, and tip-base material (work) distance are estimated from the measurement values of the sensor, and these values are acquired by the welding condition acquisition unit as welding execution conditions, and are used to evaluate the welding and generate the welding waveform.

The above-described embodiments are intended to facilitate understanding of the present invention, and are not intended to limit the present invention. The elements of the embodiments, as well as their arrangements, materials, conditions, shapes, sizes, etc., are not limited to those exemplified, and may be modified as appropriate. Furthermore, configurations shown in different embodiments may be partially substituted or combined.

10...welding robot system, 20...welding robot, 21...welding torch, 30...teach pendant, 40...robot control device, 50...power source, 100...welding waveform control device, 110...welding waveform acquisition unit, 120...feature extraction unit, 130...welding evaluation unit, 140...output unit, 150...adjustment information acquisition unit, 160...welding waveform generation unit, 170...storage unit, M100...welding waveform control method, S110 to S160...each step of welding waveform control method M100

Claims (5)

a welding waveform acquisition unit for acquiring a welding waveform during welding;
a welding evaluation unit that evaluates the welding based on the acquired welding waveform;
an output unit that outputs the evaluation result;
an adjustment information acquisition unit that acquires desired welding adjustment information for the evaluation result;
A welding waveform generating unit that generates a welding waveform based on the acquired welding waveform and the desired welding adjustment information.
Welding waveform control device. Further comprising a feature extraction unit that extracts a feature based on the acquired welding waveform,
The welding evaluation unit evaluates the welding based on the extracted feature amount,
The welding waveform generation unit generates a welding waveform based on the extracted feature amount and the desired welding adjustment information.
The welding waveform control device according to claim 1 . The welding evaluation unit evaluates the welding based on the extracted feature amount using a first machine learning model that uses the feature amount as input data and outputs an evaluation of the welding as output data.
The welding waveform control device according to claim 2 . The welding waveform generation unit uses a second machine learning model in which a feature amount based on a welding waveform and welding adjustment information are input data, and a waveform control parameter for generating a welding waveform is output data, to generate a welding waveform based on the extracted feature amount and the desired welding adjustment information.
The welding waveform control device according to claim 2 or 3. A welding waveform control method executed by a welding waveform control device that controls a welding waveform during welding,
A welding waveform acquisition step for acquiring a welding waveform during welding;
a welding evaluation step of evaluating the welding based on the acquired welding waveform;
an output step of outputting the evaluation result;
an adjustment information acquisition step of acquiring desired welding adjustment information for the evaluation result;
and generating a welding waveform based on the acquired welding waveform and the desired welding adjustment information.
Welding waveform control method.

Images