Patent Application
20250033133
The present invention relates to a learning device, a defect determination device, a learning method, a defect determination method, a welding control device, and a welding device.
In arc welding, a technique for detecting a defect occurring in a welded structure and determining whether appropriate welding is performed is known. For example, PTL 1 discloses a technique for determining whether a welding state is good or bad by capturing an image of a state of a molten pool with a visual sensor through a filter. In addition, PTL 1 also describes that a welding current, a welding voltage, and a wire feeding speed are also measured at the same time as the image is captured by the visual sensor.
Incidentally, as a means for inspecting the quality of a manufactured object built by depositing beads in a non-contact manner, for example, there is an inspection means using ultrasonic flaw detection. However, in the ultrasonic flaw detection, a probe is brought into contact with a surface of the manufactured object, and thus it is difficult to apply the probe until after a surface of the manufactured object is cut, and there is a possibility that reflected waves of ultrasonic waves cannot be detected smoothly for a complex manufactured object.
A non-contact inspection means using an X-ray CT apparatus is also considered, but the non-contact inspection means is not necessarily practical because a size of an object to be inspected is limited by the X-ray CT apparatus and the X-ray CT apparatus itself is expensive.
Therefore, an object of the present invention is to provide a learning device, a defect determination device, a learning method, a defect determination method, a welding control device, and a welding device that can accurately predict a defect size and reduce occurrence of defects in advance.
The present invention has the following configuration.
(1) A learning device that learns a defect size of an unwelded defect occurring inside an additively manufactured object in which a plurality of beads are stacked in layers on a base metal, and generates an estimation model that outputs the defect size according to input information, the learning device including:
(2) A defect determination device including:
(3) A learning method that learns a defect size of an unwelded defect occurring inside an additively manufactured object in which a plurality of beads are stacked in layers on a base metal, and generates an estimation model that outputs the defect size according to input information, the learning method including:
(4) A defect determination method, including:
(5) A welding control device including:
(6) A welding device including:
According to the present invention, it is possible to accurately predict the defect size and reduce occurrence of defects in advance.
Hereinafter, a configuration example of the present invention will be described in detail with reference to the drawings. Here, additive manufacturing in which an additively manufactured object is built by depositing beads will be described as an example, and the present invention can also be applied to general welding such as fillet welding and butt welding.
A welding system 100 includes a welding device 110 and a welding control device 120. The welding control device 120 includes a control unit 11 and a defect determination device 130.
First, the configuration of the welding device 110 will be described.
The welding device 110 includes a welding robot 13, a robot drive unit 15, a filler metal supply unit 17, a welding power supply unit 19, and a shape detection unit 21. The welding robot 13, the robot drive unit 15, the filler metal supply unit 17, the welding power supply unit 19, and the shape detection unit 21 are connected to the control unit 11 of the welding control device 120.
The welding robot 13 is an articulated robot, and a welding torch 27 is attached to a tip end shaft thereof. The robot drive unit 15 outputs a command to drive the welding robot 13, and freely sets a position and a posture of the welding torch 27 three-dimensionally within a range of degrees of freedom of a robot arm. In addition, a continuously supplied filler metal (welding wire) M is supported at a tip end of the welding torch 27.
The welding torch 27 is a gas metal arc welding torch that has a shield nozzle (not illustrated) and is supplied with shield gas from the shield nozzle. An arc welding method may be a consumable electrode type such as shielded arc welding or carbon dioxide gas arc welding, or a non-consumable electrode type such as TIG welding or plasma arc welding, and is appropriately selected according to a manufactured object (structure) to be manufactured. For example, in the case of the consumable electrode type, a contact tip is disposed inside the shield nozzle, and the filler metal M to which a melting current is supplied is held by the contact tip. The welding torch 27 generates an arc from a tip end of the filler metal M in a shield gas atmosphere while holding the filler metal M.
The filler metal supply unit 17 includes a reel 17a around which the filler metal M is wound. The filler metal M is fed from the filler metal supply unit 17 to a feeding mechanism (not illustrated) attached to the robot arm or the like, and is fed to the welding torch 27 while being fed forward and backward by the feeding mechanism as necessary.
Any commercially available welding wire can be used as the filler metal M. For example, a welding wire specified by solid wires for MAG and MIG welding of mild steel, high tensile strength steel, and low temperature service steel (JIS Z 3312), flux-cored wires for arc welding of mild steel, high tensile strength steel, and low temperature service steel (JIS Z 3313), and the like can be used. Further, it is also possible to use the filler metal M such as aluminum, an aluminum alloy, nickel, or a nickel-based alloy in accordance with desired properties.
The welding power supply unit 19 supplies, to the welding torch 27, a welding current and a welding voltage for generating an arc from the tip end of the torch.
The shape detection unit 21 is provided on or in the vicinity of the tip end shaft of the welding robot 13 and sets the vicinity of the tip end of the welding torch 27 as a measurement region. The shape detection unit 21 may be another detection unit provided at a position different from that of the welding torch 27.
The shape detection unit 21 of this configuration is moved together with the welding torch 27 by the driving of the welding robot 13 and measures shapes of beads B and a portion serving as a base when the beads B are formed. As the shape detection unit 21, for example, a laser sensor that acquires reflected light of irradiated laser light as height data can be used. In addition, other detection units such as a camera for three-dimensional shape measurement may be used as the shape detection unit 21.
According to the welding device 110 having the above-described configuration, a manufacturing program corresponding to the manufactured object to be manufactured is transmitted from the control unit 11 to the robot drive unit 15. The manufacturing program includes a large number of command codes, and is created based on an appropriate algorithm according to various conditions such as shape data (CAD data or the like), a material, and a heat input amount of the manufactured object.
The robot drive unit 15 executes the received manufacturing program, drives the welding robot 13, the filler metal supply unit 17, the welding power supply unit 19, and the like, and forms the beads B according to the manufacturing program. That is, the robot drive unit 15 drives the welding robot 13 to move the welding torch 27 along a trajectory (bead formation trajectory) of the welding torch 27, which is set in the manufacturing program. At the same time, the filler metal supply unit 17 and the welding power supply unit 19 are driven according to set welding conditions, and the filler metal M at the tip end of the welding torch 27 is melted and solidified by the arc. Accordingly, the beads B are formed on a base plate P, which is a base metal, along the trajectory of the welding torch 27. The beads B are formed adjacent to one another to form a bead layer including a plurality of beads B. By depositing a next bead layer on the bead layer, a manufactured object WK having a desired three-dimensional shape is built.
The welding control device 120 is not illustrated and is implemented by a computer device including a processor such as a CPU, a memory such as a ROM and a RAM, and a storage unit such as a hard disk drive (HD) and a solid state drive (SSD). Each component of the welding control device 120 described above operates according to a command from the CPU and performs respective functions. In addition, the welding control device 120 may be disposed away from the welding device 110 and may be connected to the welding device 110 from a remote place via a communication unit such as a network.
The control unit 11 constituting the welding control device 120 has a function of collectively controlling the robot drive unit 15, the filler metal supply unit 17, the welding power supply unit 19, and the shape detection unit 21 illustrated in
The beads B are sequentially formed by the welding torch 27 moving on the base plate P along the bead formation trajectory created in advance. In addition, at the same time as the welding torch 27 is moved, the shape detection unit 21 measures surface shapes of the existing beads B and a bead formation planned surface G. The shape detection unit 21 outputs the surface shapes of the beads B and the bead formation planned surface G (these are collectively referred to as a shape profile) to the welding control device 120.
It is preferable to measure the shape profile at the same time as the beads B are formed. In this case, the shape detection unit 21 may be disposed behind the welding torch 27 in a movement direction. Accordingly, while the beads B are formed by moving the welding torch 27, the shape of the formed beads B can be efficiently measured along a movement path, and thus a takt time can be shortened. The measurement of the shape profile may be performed at a time different from a time when the beads B are formed or may be performed at a desired timing depending on various conditions. Hereinafter, a bead that is to be formed before being formed will also be referred to as a “new bead”, and a bead that has already been formed will also be referred to as an “existing bead”.
The manufactured object WK shown here includes a frame-shaped wall portion Aw formed by the beads B and a filling portion Af that fills a region surrounded by the wall portion Aw with the beads B. The filling portion Af is formed after the wall portion Aw is formed. That is, after the wall portion Aw is formed, the beads B, which become the filling portion Af, are formed inside the wall portion Aw along bead formation trajectories F1 to F3 shown by dotted lines. Thereafter, the bead B is formed along a bead formation trajectory F4. An order of formation of the beads B within the filling portion Af may be any order.
The defect determination device 130 constituting the welding control device 120 includes a learning device 140, a determination unit 151, and a welding plan correction unit 152. The learning device 140 includes a data acquisition unit (information acquisition unit) 141, a learning unit 142, and an estimation model unit (estimation model) 143. The determination unit 151 and the welding plan correction unit 152 are connected to the estimation model unit 143 of the learning device 140.
The learning device 140 learns a defect size of an unwelded defect occurring inside an additively manufactured object in which a plurality of beads B are stacked in layers on the base plate P, and generates an estimation model that outputs the defect size according to input information. The data acquisition unit 141 acquires information on welding conditions when the beads B are deposited, a dimension related to a narrow portion forming a valley portion in a surface shape of the additively manufactured object before the beads B are deposited, a positional relation between the narrow portion and a target position of the bead B, and the defect size of the unwelded defect. Examples of the positional relation between the narrow portion and the target position of the bead B include, but are not limited to, a difference distance between a representative position of a center of the narrow portion or the like and a target position of the bead B to be formed next. The learning unit 142 generates the estimation model by learning a relation between the welding conditions, the dimension related to the narrow portion and the positional relation, and the defect size. The estimation model unit 143 registers the estimation model generated by the learning unit 142.
The determination unit 151 inputs, to the estimation model registered in the estimation model unit 143, information on a welding plan including dimensional information on the narrow portion and the positional relation. An estimated value of the defect size of the unwelded defect output from the estimation model unit 143 is compared with a reference value serving as a predetermined allowable limit.
The welding plan correction unit 152 creates a corrected welding plan by correcting at least one of the welding conditions and the positional relation when the determination unit 151 determines that the estimated value of the defect size exceeds the reference value. The corrected welding plan is transmitted to the estimation model unit 143.
Next, a process of generating training data will be described.
First, training data to be acquired by the data acquisition unit 141 of the learning device 140 is prepared. Specifically, as illustrated in
The narrow portion N between the beads B on the front layer refers to a portion formed in a valley shape between existing beads on a base plate or a lower layer, and at least three features of a bead interval W, a bottom portion interval U, and an average depth H are extracted as the features of the shape of the narrow portion N. That is, when a pair of existing beads B adjacent to each other are formed on a base surface FL representing a surface of the existing beads on the base plate or the lower layer, in addition to the bottom portion interval U and the bead interval W, the average depth H from the base surface FL to a top portion Pt of each bead B is defined as a feature. The average depth H corresponds to a valley depth to a valley bottom of the valley portion formed by the pair of existing beads B in a depositing direction.
As illustrated in
The features W, U, and H may be calculated after applying a model function that simulates a shape of a bead to a shape profile obtained by actual measurement. A sensor that measures the shape of the narrow portion N is preferably a non-contact type sensor as in the shape detection unit 21 of the present example, and more preferably a sensor that is attached in the vicinity of the welding torch 27 and measures the shape of the narrow portion N while scanning a surface of the bead B.
The horizontal distance δ between the welding torch 27 and the narrow portion N is calculated as a difference distance between a representative position of the narrow portion N and a center position at which a bead B is placed next. The representative position of the narrow portion N may be, for example, a position at which the depth is minimum or an intermediate point of a portion evaluated by the feature U.
As data on the defect size, data obtained by directly observing and measuring a cut surface after the additive specimen of the beads illustrated in
Here, when the beads are formed, a recess portion is likely to be formed in root portions of side edge portions of the beads, and a foreign matter is likely to be accumulated in the root portions of the side edge portions. Therefore, as illustrated in
The defect size is a size including the cross-sectional area or a length in the cross section orthogonal to the longitudinal direction of the bead B, and includes, for example, indices such as a diameter when a shape of the defect C is approximated by a true circle, a cross-sectional area of the approximated circle, an area of the defect C to be observed, and a long axis length and a short axis length when the shape of the defect is approximated by an ellipse.
When the prepared training data is input to the data acquisition unit 141 of the learning device 140, the learning unit 142 generates an estimation model including a relation between the training data and the defect size based on the training data. The estimation model generated by the learning unit 142 is transmitted to the estimation model unit 143 and registered in the estimation model unit 143. Examples of the means of generating the estimation model in the learning unit 142 include well-known means such as a decision tree, linear regression, random forest, support vector machine, Gaussian process regression, and neural network. At this time, in addition to the defect size, a probability of occurrence of a defect having a specific size, an occurrence density of defects occurring in a specific region (area or volume), and the like may also be learned.
Next, a process of determining welding conditions in the welding control device 120 will be described.
By moving the welding torch 27 along the bead formation trajectory created in advance, the surface shapes of the existing beads B and the bead formation planned surface G are measured by the shape detection unit 21 provided in parallel to the welding torch 27. Accordingly, the shape of the narrow portion N on the bead formation planned surface G is measured (S1).
The features W, U, and H related to the narrow portion N are calculated based on the measured shape of the narrow portion N on the bead formation planned surface G (S2). The features W, U, and H may be calculated by preparing in advance a model function that simulates the narrow portion N and fitting the same. A smoothing process or the like may be performed on the shape profile obtained by the measurement before fitting the shape profile with the model function.
The calculated features W, U, and H, welding conditions of the beads B to be formed on the bead formation planned surface G, and information on a target position of the welding torch 27 are input to the estimation model unit 143, and an estimated value of the defect size is obtained (S3).
The determination unit 151 compares the estimated value of the defect size obtained by the estimation model unit 143 with a preset allowable value and determines whether the estimated value is equal to or less than the allowable value (S4). The allowable value is a value of an allowable limit size which is an allowable defect size. The allowable limit size may be set for each manufactured object to be built or each material to be used for building.
When the determination unit 151 determines that the estimated value of the defect size exceeds the allowable value (S4: No), the welding plan correction unit 152 searches for a condition under which the defect size is reduced to the allowable limit size or less (S5). For example, the welding plan correction unit 152 corrects a welding plan such as the welding conditions of the beads B to be formed on the bead formation planned surface G and the target position of the welding torch 27, and transmits the corrected welding plan to the estimation model unit 143. Accordingly, the estimated value of the defect size is obtained again (S3), and the estimated value of the defect size obtained again and the allowable value are compared and determined by the determination unit 151. The correction of the welding plan by the welding plan correction unit 152, the estimation of the defect size by the estimation model unit 143, and the determination by the determination unit 151 are repeated.
When the determination unit 151 determines that the estimated value of the defect size is equal to or less than the allowable value (S4: Yes), the welding conditions of the beads B to be formed on the bead formation planned surface G, the target position of the welding torch 27, and the like, which are input to the estimation model unit 143 for obtaining the estimated value, are determined as the welding plan (S6). Thereafter, the beads B are formed on the bead formation planned surface G according to the welding plan.
According to the welding system 100 described above, the estimation model is generated by learning a relation between the welding conditions, the dimension related to the narrow portion and the positional relation, and the defect size. At this time, as a dimension related to the narrow portion N, at least one of the bottom portion interval (bottom width) U of the valley portion of the narrow portion N serving as a base, the opening width (bead interval) W representing an interval between top portions on both sides of the valley portion, both sides constituting the valley portion, and a valley depth H from the top portion to a bottom of the valley portion is used. In this way, by performing machine learning based on the shape of the narrow portion N serving as a base and the conditions of the beads B to be deposited thereon, the defect size can also be accurately predicted when the number of data is relatively small.
The defect size can be accurately predicted by generating the estimation model using the welding conditions including at least one of the feeding speed of the filler metal M, the travel speed, the welding current, the welding voltage, the torch angle α of the welding torch 27, the volume of the bead B, and the cross-sectional area of the cross-section orthogonal to the longitudinal direction of the bead B. In particular, by learning an index related to the volume or a heat input amount of the bead B, a relation of whether the narrow portion N on the front layer is completely filled can be incorporated into machine learning.
Moreover, by comparing an estimated value of the defect size of the defect C with a reference value serving as a predetermined allowable limit, when the occurrence of the defect C is predicted, it is also possible to determine whether the estimated defect C is a harmless defect or a harmful defect and to reduce unnecessary defect handling.
By repeating the correction of the welding plan, the calculation of the estimated value, and the determination of the estimated value, the welding conditions suitable for reducing the defect C and the target position of the welding torch 27 can be extracted.
Accordingly, the arc welding can be executed according to the welding conditions suitable for reducing the defect C and the target position of the welding torch, and the manufactured object can be built while reducing the defect C such as an unwelded defect in the narrow portion N.
In the learning device 140 having the above-described configuration, the learning unit 142 may register, in the estimation model unit 143, a variance of the defect size in addition to the defect size, and the estimation model unit 143 may output the defect size corresponding to the input information and a variance value thereof, according to the input information.
The learning unit 142 registers, in the estimation model unit 143, the variance of the defect size in addition to the defect size, and the estimation model unit 143 can output the defect size and the variance value thereof, and thus it is possible to predict the defect C by taking into account variations occurring depending on the defect size.
As described above, the present invention is not limited to the above-described embodiments, and combinations of the respective configurations of the embodiments and changes and applications made by those skilled in the art based on the description of the specification and well-known techniques are also intended for the present invention and are included in the scope of protection.
As described above, the following matters are disclosed in the present specification.
(1) A learning device that learns a defect size of an unwelded defect occurring inside an additively manufactured object in which a plurality of beads are stacked in layers on a base metal, and generates an estimation model that outputs the defect size according to input information, the learning device including:
According to the learning device, the estimation model is generated by learning the relation between the welding condition, the dimension related to the narrow portion and the positional relation, and the defect size. At this time, as the dimension related to the narrow portion, at least one of the bottom width of the valley portion of the narrow portion serving as a base, the opening width representing the interval between the top portions on both sides of the valley portion, both sides constituting the valley portion, and the valley depth from the top portion to the bottom of the valley portion is used. In this way, by performing machine learning based on a shape of the narrow portion serving as a base and the condition of the beads to be deposited thereon, the defect size can also be accurately predicted when the number of data is relatively small.
(2) The learning device according to (1), in which the welding condition includes at least one of a feeding speed of a welding wire, a travel speed, a welding current, a welding voltage, a torch angle of a welding torch, a volume of the bead, and a cross-sectional area of a cross section orthogonal to a longitudinal direction of the bead.
According to the learning device, the defect size can be accurately predicted by generating the estimation model using the welding condition including at least one of the feeding speed of the welding wire, the travel speed, the welding current, the welding voltage, the torch angle of the welding torch, the volume of the bead, and the cross-sectional area of the cross-section orthogonal to the longitudinal direction of the bead. In particular, by learning an index related to the volume or a heat input amount of the bead, a relation of whether a narrow portion on a front layer is completely filled can be incorporated into machine learning.
(3) The learning device according to (1) or (2), in which the positional relation includes a difference distance between a representative position of the narrow portion and a target position of the bead to be formed next.
According to the learning device, a highly accurate estimation model can be generated using the positional relation including the difference distance between the representative position of the narrow portion and the target position of the bead to be formed next.
(4) The learning device according to any one of (1) to (3), in which the defect size includes the cross-sectional area or a length of the cross section orthogonal to the longitudinal direction of the bead.
According to the learning device, a highly accurate estimation model can be generated using the defect size including the cross-sectional area or the length of the cross section orthogonal to the longitudinal direction of the bead.
(5) The learning device according to any one of (1) to (4), in which the learning unit registers, in the estimation model, a variance of the defect size in addition to the welding condition, the dimension related to the narrow portion, and the defect size corresponding to the positional relation, and
According to the learning device, a reliability of the estimation can be determined by outputting the variance value together with the defect size, and a range of an estimated value that is practically reliable can be determined based on the reliability.
(6) A defect determination device including:
According to the defect determination device, by providing the determination unit that compares the estimated value of the defect size of the unwelded defect with the reference value serving as a predetermined allowable limit, when the occurrence of the defect is predicted, it is also possible to determine whether the estimated defect is a harmless defect or a harmful defect and to reduce unnecessary defect handling.
(7) The defect determination device according to (6), further including:
According to the defect determination device, by repeating the correction of the welding plan, the calculation of the estimated value, and the determination of the estimated value, the welding condition suitable for reducing the unwelded defect and the target position of the welding torch can be extracted.
(8) A learning method that learns a defect size of an unwelded defect occurring inside an additively manufactured object in which a plurality of beads are stacked in layers on a base metal, and generates an estimation model that outputs the defect size according to input information, the learning method including:
According to the learning method, the estimation model is generated by learning the relation between the welding condition, the dimension related to the narrow portion and the positional relation, and the defect size. At this time, as the dimension related to the narrow portion, at least one of the bottom width of the valley portion of the narrow portion serving as a base, the opening width representing the interval between the top portions on both sides of the valley portion, both sides constituting the valley portion, and the valley depth from the top portion to the bottom of the valley portion is used. In this way, by performing machine learning based on a shape of the narrow portion serving as a base and the condition of the beads to be deposited thereon, the defect size can also be accurately predicted when the number of data is relatively small.
(9) A defect determination method, including:
According to the defect determination method, by comparing the estimated value of the defect size of the unwelded defect with the reference value serving as a predetermined allowable limit, when the occurrence of the defect is predicted, it is also possible to determine whether the estimated defect is a harmless defect or a harmful defect and to reduce unnecessary defect handling.
(10) The defect determination method according to (9), further including:
According to the defect determination method, by repeating the correction of the welding plan, the calculation of the estimated value, and the determination of the estimated value, the welding condition suitable for reducing the unwelded defect and the target position of the welding torch can be extracted.
(11) A welding control device including:
According to the welding control device, the arc welding can be executed according to the welding condition suitable for reducing the unwelded defect and the target position of the welding torch.
(12) A welding device including:
According to the welding device, the manufactured object can be built while reducing the defects such as the unwelded defect in the narrow portion.
The present application is based on Japanese Patent Application No. 2022-020686 filed on Feb. 14, 2022, the contents of which are incorporated herein by reference.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2022-020686 | Feb 2022 | JP | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2022/048297 | 12/27/2022 | WO |
