TWI770910B - Welding condition adjustment device - Google Patents

Abstract

[Problem] To provide a welding condition adjusting device that automatically adjusts arc welding conditions in at least one welding section between the arc starting step and the arc ending step, and can shorten the cycle of the welding step. [Solution] It is a welding condition adjustment device (5) for adjusting arc welding conditions that is repeated, and includes: an acquisition unit that acquires welding data indicating welding conditions related to the welding step; The welding data acquired by the acquisition unit is used to adjust arc welding conditions in at least one welding section between the arc starting step and the arc ending step.

Description

Welding condition adjustment device

The present invention relates to a welding condition adjusting device for adjusting arc welding conditions in a welding step.

As one of the welding methods, there is a gas shielded arc welding method of consumable electrode type. The gas shielded arc welding method is a method of generating an arc between the welding wire fed to the welded part of the base metal and the base metal, and welding the base metal by the heat of the arc, especially in order to prevent the oxidation of the base metal that becomes a high temperature. , which is the one that sprays shielding gas to the periphery of the welding part and welds.

Patent Document 1 discloses a technique for automatically setting welding conditions by machine learning using image data, appearance data of the bead, and spatter generation amount data, and the image data is obtained by photographing the welding site, and the The appearance data of Lianzhu is obtained by processing the image data.

Patent Document 2 discloses a technique for judging the quality of the welded bead by estimating the physical quantity related to the cross-section of the welded bead based on the measurement values obtained by measuring the welding current and the welding voltage.

[Preceding Patent Documents] [Patent Literature]

[Patent Document 1] Japanese Patent Laid-Open No. 2017-30014

[Patent Document 2] Japanese Patent Application Laid-Open No. 2016-26877

On the other hand, in arc welding, in order to shorten the cycle time while ensuring the welding quality, it is necessary to adjust the welding conditions to stabilize the weldability while increasing the welding speed. In this adjustment, it is necessary for the operator to set the conditions considered to be optimal from the results of many trials by manual work.

An object of the present invention is to provide a welding condition adjusting device that automatically adjusts arc welding conditions in at least one welding section between the arc starting step and the arc ending step, thereby shortening the cycle of the welding step.

The welding condition adjustment device of this form is a welding condition adjustment device for adjusting arc welding conditions, and it includes: an acquisition unit (for example, a welding monitoring data acquisition unit, an image data acquisition unit, a welding condition data acquisition unit, and a state data acquisition unit), acquiring welding data representing a welding state related to the welding step; and an adjusting unit for adjusting and correlating with the cycle of at least one welding interval between the arc starting step and the arc ending step based on the welding data acquired by the acquiring unit arc welding conditions.

According to this aspect, the acquiring section acquires welding data, and the adjusting section adjusts arc welding conditions in at least one welding section between the arc starting step and the arc ending step based on the acquired welding data. The welding step includes an arc starting step, a formal welding step performed in one or a plurality of welding sections, and an arc ending step. The actual welding is the welding performed between the arc start step and the arc end step. In a welding section, welding is performed according to an arc welding condition. When there are a plurality of welding sections, different arc welding conditions may be set in each welding section. In this embodiment, the arc welding conditions are adjusted in at least one welding section during the actual welding step.

Welding data is information showing the state of welding in the welding step and contains information that is helpful for determination. According to the information, the judgment is whether the cycle can be shortened or whether the cycle should be extended by adjusting the arc welding conditions. The adjustment section can adjust the arc welding conditions in at least one welding section between the arc starting step and the arc ending step in a manner of shortening the cycle of the welding step without deteriorating the welding result by using the welding data, and The cycle time of welding steps can be shortened.

In addition, in the case where a plurality of welding power sources are installed in the production line, the welding condition adjusting device adjusts the arc welding conditions for each welding power source, that is, each welding wire. In addition, it goes without saying that a welding condition adjusting device may be provided in each welding power source, or one welding condition adjusting device may be configured to adjust the arc welding conditions of a plurality of welding power sources, respectively.

The welding condition adjustment device of this form has a good or bad judgment part (for example: a first good or bad judgment part, a second good or bad judgment part), and the good or bad judgment part judges whether the welding result is good or not based on the welding data obtained by the acquisition part. Defects caused by the short cycle; the adjustment section determines the content of changes in arc welding conditions associated with the cycle in the following manner, when the quality judgment section judges that the cycle is good, the cycle is shortened, and the quality judgment section judges If the cycle is short, the cycle should be extended.

According to this aspect, when the welding result of the welding step is good, the welding condition adjustment device has the possibility of shortening the cycle of the welding step by shortening the arc welding conditions. Therefore, the arc welding conditions are adjusted to shorten the welding process. cycle of steps. In the case of defects caused by a short cycle, the welding condition adjustment device adjusts the arc welding conditions and prolongs the cycle of the welding step. By this adjustment process, the cycle of welding steps can be shortened so as not to deteriorate the welding result as much as possible.

In the welding condition adjustment device of the present aspect, the adjustment section is based on the shortening of the cycle of the welding step, and when the welding result changes from a good state to a poor state, it is determined based on the arc welding conditions associated with the cycle before the cycle shortening. Adjust and make the memory part memorize the arc welding conditions determined.

According to this form, the cycle of the welding step can be shortened to the shortest, and the memory part is the memory cycle. Shortest arc welding conditions. The shortest arc welding condition is not necessarily the arc welding condition with the shortest period logically. The shortest arc welding condition means the arc welding condition before the cycle is shortened when the welding result of the next welding step changes from a good state to a poor state when the cycle of the welding step is shortened.

After that, by using the arc welding conditions memorized in the memory, the cycle can be shortened immediately.

In the welding condition adjustment device of the present embodiment, the quality judgment unit has a quality judgment neural network, which, when the welding data has been input, outputs a welding result indicating whether the welding data is good or not related to the welding step when the welding data is obtained. And the way the bad data is caused by the short period makes the neural network learn.

According to this aspect, the good or bad judgment neural network is, for example, a deep neural network that has completed learning, and can appropriately judge whether the welding result is good or not. The type of the neural network is not particularly limited. As long as the characteristics of welding data are matched, CNN (Convolutional Neutral Network), RNN (Recurrent Neutral Network), LSTM (Long Short-Term Memory), etc. can be appropriately selected.

In the welding condition adjustment device of this aspect, the adjustment unit has an adjustment neural network, which, when the welding data has been input, outputs data indicating the content of changes in the arc welding conditions associated with the cycle to make the neural network online learning.

According to this aspect, the adjustment neural network is, for example, a deep neural network that has completed learning, and the arc welding conditions can be adjusted appropriately. The type of the neural network is not particularly limited. As long as the characteristics of welding data are matched, CNN, RNN, LSTM, etc. can be appropriately selected.

In the welding condition adjustment device of this aspect, the adjustment neural network outputs data indicating the amount of change in arc welding conditions associated with the cycle.

According to this form, the adjustment of the neural network is not whether the cycle of the welding step should be shortened, but the change amount of the adjustable arc welding conditions can be output. For example, a neural network system can be tuned to output a large amount of change when the welding results are stable, but not when the welding results are good but not In the stable case, output a small amount of change. Therefore, the cycle time of the welding step can be shortened more rapidly.

The welding condition adjustment device of this form includes: a good or bad judgment part, which judges whether the welding result is good or not and the bad caused by the short cycle based on the welding data obtained by the acquisition part; and a learning processing part, which adjusts the welding result according to the The judgment result of the good or bad judgment section obtained after arc welding conditions is used to make the adjustment neural network learn.

According to this aspect, the adjustment neural network uses the data representing the welding result when the arc welding conditions have been adjusted to perform learning. Therefore, the arc welding conditions can be adjusted more effectively and the cycle time of the welding step can be shortened without deteriorating the welding result.

In the welding condition adjustment device of the present aspect, the learning processing unit makes the adjustment neural network learn in the following manner. When the quality judgment unit judges that it is good, the cycle is shortened, and the quality judgment unit judges by the cycle. In case of bad situation caused by short, prolong the cycle.

According to this aspect, the learning of the neural network can be adjusted in the direction of shortening the cycle of the welding step. By this learning, the cycle of the welding step can be minimized.

In the welding condition adjustment apparatus of this aspect, when the welding result is an intermediate state between good and bad, the learning processing unit makes the adjustment neural network learn while maintaining the cycle.

According to this aspect, in the case where the welding result is an intermediate state between good and bad, the adjustment neural network is made to learn so as to maintain the cycle of the welding steps. The intermediate state is a state in which the welding result is relatively good, but if the cycle is to be shortened, the welding result may deteriorate. By this learning, the cycle of the welding step can be minimized, and the welding result can be stabilized in a good state.

In the welding condition adjustment device of the present embodiment, the good or bad judgment part has a good or bad judgment neural network, which outputs whether the welding result is good or not related to the welding step when the welding data is obtained when the welding data has been input. And the way the bad data is caused by the short period makes the neural network learn.

According to this aspect, the good or bad judgment neural network is, for example, a deep neural network that has completed learning, and can appropriately judge whether the welding result is good or not. By using the good or bad judgment result of the good or bad judgment neural network, the adjusted neural network can learn more efficiently.

In the welding condition adjustment device of this aspect, the adjustment neural network includes substantially the same network configuration as all or a part of the good/fail judgment neural network.

According to this aspect, the adjustment neural network includes substantially the same neuron configuration as all or a part of the good-failure determination neural network. For example, a part of the adjustment neural network has the same or substantially the same intermediate layers and weighting coefficients as all or part of the good-fail decision neural network. The judgment of the quality of the welding results and the adjustment of the arc welding conditions have some common features, so the neural network for judgment of good or not can be used in the adjustment of the neural network. That is, the initial value of the weighting coefficient for adjusting the neural network can be set to a more appropriate value. Therefore, even if the learning data for learning arc welding conditions is insufficient, as long as the learning data for welding data and data indicating whether the welding result is good or not can be sufficiently prepared, and the initial value of the weighting coefficient for adjusting the neural network is appropriately set, it is possible to Make the tuned neural network learn more efficiently. In addition, of course, the network structures of the adjustment neural network and the good/fail judgment neural network may be configured in the same manner.

The welding condition adjustment device of this form has a state data acquisition part for acquiring state data, the state data at least including image data obtained by photographing the welding part after welding; the adjustment part includes: an evaluation part based on the state data The state data acquired by the acquisition section and the action data indicating the action related to the arc welding conditions calculate the evaluation value of the action in the state indicated by the state data; and the action selection section is selected by the evaluation section The calculated evaluation value is the action with the largest value.

According to this aspect, based on the state data including the image data obtained by photographing the welding part after welding, the evaluation unit that has undergone reinforcement learning is used to select actions related to the optimum arc welding conditions.

The welding condition adjustment device of this form includes: a good or bad judgment part, which judges whether the welding result is good or not and the bad caused by the short cycle based on the welding data obtained by the acquisition part; a reward calculation part is based on the adjustment of the After the arc welding condition is obtained, the judgment result of the quality judgment section and the time required for the welding step are used to calculate the reward for the arc welding condition; and the reinforcement learning section is based on the status data obtained by the status data acquisition section, Action data showing actions related to the arc welding conditions and the reward calculated by the reward calculation unit are for the evaluation unit to learn.

According to this aspect, reinforcement learning of arc welding conditions for shortening the cycle of welding steps can be performed.

In the welding condition adjustment device of the present aspect, the evaluation unit has an evaluation neural network that has been inputted with the state data acquired by the state data acquisition unit and the action data indicating the action related to the arc welding condition. case, output the evaluation value of the action in the state indicated by the state data.

According to this aspect, deep reinforcement learning can be performed on arc welding conditions for shortening the cycle of welding steps.

In the welding condition adjusting device of the present embodiment, the welding data indicating the welding state related to the welding step includes the welding current and welding voltage detected in the welding step, the wire feeding speed, the short-circuit condition, At least one of the welding sound collected in the welding and the image of the welding part taken after the welding is finished.

According to this aspect, the welding current and welding voltage detected during the welding step, the wire feeding speed, the short-circuit condition, the welding sound collected during the welding step, and the welding sound after the welding end are used to indicate the welding step. Data of at least one of the images of the welded part taken. Arc welding conditions can be adjusted.

The welding condition adjusting device of the present form has a receiving unit that receives the adjustment strength for adjusting the arc welding conditions by the adjusting unit; the adjusting unit is based on the adjusting strength received by the receiving unit. degree, adjust the arc welding conditions associated with the cycle.

According to this aspect, the user can arbitrarily set the degree of automatic adjustment of the arc welding conditions by the adjustment section.

The welding system of this aspect includes the welding condition adjustment device of any one of the above-mentioned aspects, a welding robot holding a welding torch, and a welding power source that supplies a welding current to the welding torch.

According to this aspect, the welding system including the welding robot and the welding power source can shorten the cycle of the welding steps. In addition, the welding condition adjustment device may be installed inside the welding robot and the welding power source, inside the control device for controlling the actions of the welding robot and the welding power source, or outside the welding robot, the welding power source and the control device have separately. Furthermore, the welding condition adjustment device may be a server, and the control device or the welding power source may be configured to communicate with the server, thereby shortening the cycle of the welding step.

The welding condition adjustment method of this form is a welding condition adjustment method for adjusting arc welding conditions, which acquires welding data indicating the welding state related to the welding step, and then adjusts and adjusts the arc starting step and the arc ending according to the obtained welding data. Arc welding conditions associated with the cycle of at least one welding interval between steps.

According to this form, the cycle of the welding step can be shortened. The welding condition adjustment method may be implemented automatically by a welding power source, a control device, and the like constituting the welding system, or an operator may connect the welding condition adjustment device to the welding system to implement the welding condition adjustment method.

The computer program of this form is a computer program for making a computer adjust arc welding conditions. It causes the computer to execute the following processing to obtain welding data indicating the welding state related to the welding step, and then adjust the welding data according to the obtained welding data. and arc welding conditions associated with a period of at least one welding interval between the arc start step and the arc end step.

According to this aspect, a computer can be used as the welding condition adjustment device.

According to the present invention, the arc welding conditions in at least one welding section between the arc starting step and the arc ending step are automatically adjusted, so that the cycle of the welding step can be shortened.

1: Welding robot

2: Welding power source

3: Control device

4: Camera device

5,205,305: Welding condition adjustment device

11: Welding gun

12: Welding wire feeding device

21: Power Department

22: Wire feed control section

23: Shielding gas supply part

24: Detection Department

50: Control Department

50a: Input section

50b: Output section

50c: Memory Department

50d: Computer programming

51a: Welding monitoring data acquisition department

51b: Image data acquisition department

51c: Welding condition data acquisition department

51d: Status Data Acquisition Department

52a: 1st Good or Bad Judgment Section

52b: 2nd Good or Bad Judgment Section

53a: Good or bad judgment RNN

53b: Good or bad judgment CNN

54,254: Comprehensive Judgment Department

55,255,355: Adjustment Department

56: Welding Control Department

57: Shortest welding condition memory section

258: Tuning NN

258a: Welding Status Recognition Network Department

258b: Appearance Status Recognition Network Department

258c: Adjusting the Network Department

259: Learning Processing Department

355a: Evaluation Department

355b: Action Selection Division

355c: Compensation Calculation Department

355d: Reinforcement Learning Division

355e: Evaluating NNs

406: Adjustment Method Acceptance Department

407: Adjust the screen

471: Adjustment method selection section

472: Priority display part

473: Priority adjustment slider

A: Base material

W: welding wire

FIG. 1 is a schematic diagram showing an arc welding system according to Embodiment 1. FIG.

Fig. 2 is a block diagram showing the welding condition adjusting device according to the first embodiment.

Fig. 3 is a functional block diagram showing the welding condition adjusting device according to the first embodiment.

FIG. 4 is a flowchart showing a welding condition adjustment method according to Embodiment 1. FIG.

FIG. 5 is a flowchart showing a welding condition adjustment method according to Embodiment 1. FIG.

6] It is a functional block diagram which shows the welding condition adjustment apparatus of Embodiment 2. [FIG.

Fig. 7 is a schematic diagram showing the network configuration of the adjustment unit.

Fig. 8 is a functional block diagram showing the welding condition adjusting device according to the third embodiment.

FIG. 9 is a schematic diagram showing an arc welding system according to Embodiment 4. FIG.

FIG. 10 is a schematic diagram showing an adjustment screen.

Hereinafter, the present invention will be described in detail based on the drawings showing the embodiments. Moreover, at least a part of embodiment described below may be combined arbitrarily.

(Embodiment 1)

FIG. 1 is a schematic diagram showing an arc welding system according to the first embodiment. The arc welding system of the present embodiment is a consumable electrode type gas shielded arc welding machine, and includes a welding robot 1 , a welding power source 2 , a control device 3 , a camera device 4 , and a welding condition adjustment device 5 . The welding condition adjustment device 5 is provided in the control device 3 . In addition, for the convenience of drawing and description, it is assumed that the control device 3 includes a welding condition adjustment device. The unit 5 is used for description, but the control device 3 and the welding condition adjustment device 5 may be integrally constituted, and the hardware and software of the control device 3 may realize the function of the welding condition adjustment device 5 in appearance. In addition, the welding condition adjustment device 5 may be installed in the welding power source 2, or may be installed in another device. Furthermore, the functions of the welding condition adjustment device 5 may be distributed and processed by a plurality of devices and servers.

The welding robot 1 automatically performs arc welding of the base material A. The welding robot 1 has a base that is fixed in place on the floor. At the base, a plurality of arms are rotatably connected via a shaft, and the torch 11 is fastened to the tip end of the arms. Also, a wire feeding device 12 is provided at an appropriate position of the arm. A motor is provided at the connecting portion of each arm, and each arm rotates around the shaft portion by the rotational driving force of the motor. The rotation of the motor is controlled by the control device 3 . The control device 3 can move the welding torch 11 with respect to the base material A up, down, front, back, right and left by rotating the arms. In addition, an encoder that outputs a signal indicating the rotational position of the arm to the control device 3 is provided at the connecting portion of each arm, and the control device 3 recognizes the position of the welding torch 11 based on the signal output from the encoder.

The welding torch 11 has a cylindrical contact piece made of a conductive material such as copper alloy, and guides the welding wire W to the base metal A to be welded, and supplies a welding current necessary for generating an arc. The welding current is supplied by the welding power source 2 . The welding wire W is supplied to the welding torch 11 from a welding wire supply source (not shown) by the wire feeding device 12 . The bonding wire W is, for example, a single wire, and functions as a consumable electrode.

The contact piece is in contact with the bonding wire W inserted thereinto, and supplies a welding current to the bonding wire W. As shown in FIG. Further, the welding torch 11 has a nozzle formed in a hollow cylindrical shape surrounding the contact piece, and sprays shielding gas toward the base material A from the opening of the tip end. The shielding gas system is used to prevent oxidation of the base metal A and the welding wire W melted by the arc. The shielding gas system is, for example, carbon dioxide, a mixed gas of carbon dioxide and argon, an inert gas such as argon, or the like. The shielding gas system is supplied from the welding power source 2 .

The welding power source 2 includes a power source unit 21 , a wire feed control unit 22 , a shielding gas supply unit 23 , and a detection unit 24 . The power supply unit 21 is connected to the contact piece of the welding torch 11 and the base material A via a power supply cable, and supplies welding current. The wire feeding control unit 22 controls the feeding of the wire W of the wire feeding device 12. Feed rate. The shielding gas supply unit 23 supplies shielding gas to the welding torch 11 . The detection unit 24 includes a current detection unit that detects the welding current flowing in the arc during the welding step, and a voltage detection unit that detects the voltage applied to the welding torch 11 and the base material A. The power supply unit 21 includes a power supply circuit, a signal processing circuit, and the like that output a DC current subjected to PWM control based on the welding current and the welding voltage detected by the detection unit 24 . In addition, the welding power source 2 outputs to the control device 3 welding monitoring data indicating the state of the welding state in the welding step. The welding monitoring data is, for example, welding current data or welding voltage data indicating the welding current or welding voltage detected in the welding step. In addition, as welding monitoring data, feed speed data indicating the feeding speed of the welding wire W, short-circuit condition data indicating a short-circuit condition, and welding sound data collected by a microphone (not shown) may also be output to the control device 3 . .

In addition, the welding monitoring data is an example of welding data showing the welding state in the next welding step.

After the welding step, the imaging device 4 photographs the welded portion of the base metal A, and outputs the image data obtained by the photographing to the control device 3 . In addition, the imaging device 4 may be an infrared camera that detects the temperature of the welding site.

In addition, this image data is an example of welding data which shows the welding state concerning the next welding step.

The control device 3 controls the operation of the welding robot 1 , and outputs welding conditions such as welding current, welding voltage, feeding speed of the welding wire W, and supply amount of shielding gas to the welding power source 2 , and controls the operation of the welding power source 2 . The control device 3 memorizes various initial welding conditions such as the material of the base metal A and the type of groove. In addition, the control device 3 outputs arc welding conditions and executes the welding process.

The welding step includes an arc starting step, a formal welding step performed in one or more welding sections, and an arc ending step. The actual welding is the welding performed between the arc start step and the arc end step. The control device 3 outputs arc welding conditions for controlling the arc starting step at the time of arc starting. Arc start is a process of generating an arc to start welding. In addition, the control device 3 is connected to the electric At the end of the arc, the arc welding conditions used to control the processing of the arc end step are output. The processing of the arc end step is the processing to prepare for the next welding step, such as anti-stick processing and welding release processing. Further, the control device 3 outputs arc welding conditions for controlling the actual welding performed in one or a plurality of welding sections. In a welding section, welding is performed according to an arc welding condition. When there are a plurality of welding sections, different arc welding conditions may be set in each welding section. Hereinafter, in the first embodiment, adjustment of the arc welding conditions in the welding step, especially in one welding section, will be described. In addition, the main welding is composed of a plurality of welding sections, and when the arc welding conditions are different in each welding section, the arc welding conditions in each welding section may be adjusted according to the method described in the first embodiment. In addition, of course, in the case where the main welding is constituted by one welding section, it is only necessary to configure the arc welding conditions in the welding section.

The arc welding conditions memorized by the control device 3 are not necessarily optimal, and the arc welding conditions are adjusted by the welding condition adjusting device 5 so that the cycle of the welding steps becomes the shortest within the range where the welding result does not deteriorate.

1 shows only one set of the welding power source 2 and the welding torch 11, but when a plurality of welding power sources 2 are installed in the production line, one welding condition adjustment device 5 may be configured so that each of the plurality of welding power sources 2 can be divided into To adjust the arc welding conditions in the power source, the welding condition adjustment device 5 may be provided individually for each of a plurality of welding power sources 2, and the arc welding conditions in each power source may be adjusted.

FIG. 2 is a block diagram showing the welding condition adjusting device 5 according to the first embodiment. The welding condition adjusting device 5 includes a control unit 50 that controls the operation of each component of the welding condition adjusting device 5 . The control unit 50 is connected to an input unit 50a, an output unit 50b, and a memory unit 50c.

The memory unit 50c is a nonvolatile memory such as EEPROM (Electrically Erasable Programmable ROM) and flash memory. The memory unit 50c stores a computer program 50d, and the computer program 50d is used to optimize the welding step and minimize the cycle of the welding step within the range where the welding result is not deteriorated.

The control unit 50 is a computer having a processor such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit) or a multi-core CPU, a ROM (Read Only Memory), a RAM (Random Access Memory), an I/O interface, and the like. , which connects the input unit 50a, the output unit 50b, and the memory unit 50c. The control unit 50 executes the welding condition adjustment method by executing the computer program 50d stored in the memory unit 50c, and makes the computer function as the welding condition adjustment device 5. The welding condition adjustment method is continuous production in which welding is repeated in the production line. The cycle of the welding step is minimized. In addition, the repeated welding step refers to the repeated welding performed by a welding power source 2 or a welding torch 11 provided in the production line.

The input unit 50 a is connected to the welding power source 2 and the imaging device 4 . In the input part 50a, the welding monitoring data output from the welding power source 2 and the image data output from the camera 4 are input. The welding monitoring data is, for example, time-series data indicating welding current and welding voltage in the welding step, feeding speed of the welding wire W, short-circuit condition, welding sound, and the like. The image data are data showing the appearance of the beads after welding.

The output unit 50b is connected to the welding robot 1 and the welding power source 2 . The control unit 50 controls the welding steps and arc welding conditions, and outputs control data for changing the arc welding conditions to the welding robot 1 and the welding power source 2 . The control data for changing the arc welding conditions may also indicate the change of the arc welding conditions, or may indicate the arc welding conditions after the change.

FIG. 3 is a functional block diagram showing the welding condition adjusting device 5 according to the first embodiment. The welding condition adjustment device 5 is a functional block, and includes a welding monitoring data acquisition unit 51a (acquisition unit), an image data acquisition unit 51b (acquisition unit), a first quality judgment unit 52a (quality judgment unit), and a second quality judgment unit 52b (Quality/Determination Unit), a Quality/Determination Comprehensive Determination Unit 54 , an Adjustment Unit 55 , a Welding Control Unit 56 , and a Shortest Welding Condition Memory Unit 57 .

The welding monitoring data acquisition unit 51a acquires the welding monitoring data outputted from the welding power source 2, and outputs the acquired welding monitoring data to the first quality determination unit 52a.

The image data acquisition unit 51b acquires the image data output from the imaging device 4, and outputs the acquired image data to the second quality determination unit 52b.

The first good or bad judgment part 52a has a good or bad judgment RNN (Recurrent Neutral Network) 53a, and the good or bad judgment RNN 53a outputs the good or bad of the welding result related to the welding step when the welding monitoring data is obtained when the welding monitoring data is input. data of. The good or bad judgment RNN53a is, for example, a recurrent neural network that completes learning.

Good or bad judgment RNN53a includes, for example, in the output layer: the first neuron, which outputs the data indicating the probability that the welding result is good; the second neuron, which outputs the data indicating the probability that the welding result is bad; the third neuron, which outputs the data indicating the probability that the welding result is bad The output indicates data indicating the probability of poor welding caused by a short cycle; and the fourth neuron outputs data indicating the probability of poor welding caused by a long cycle. In this case, the data indicating good or bad is the data output from the 1st to 4th neurons. In addition, generally, since the welding monitoring data when the welding failure caused by the short cycle occurs and the welding monitoring data when the welding failure caused by the long cycle occurs have different characteristics, the first quality judgment unit 52a can determine Poor welding is caused by a short cycle or a long cycle.

In addition, it is also possible to judge whether the RNN 53a has a neuron in the output layer that outputs the good or bad of the welding result as a binary value. In this case, the data indicating good or bad is the binary data output from the neuron.

Furthermore, it is also possible to judge whether the RNN 53a has a neuron in the output layer that outputs a ratio indicating the degree of good or bad of the welding result.

Good or bad judgment RNN53a only needs to use the welding monitoring data (input data) and the data (teacher data) corresponding to the welding monitoring data indicating whether the welding result is good or not as learning data, and supply it to the recurrent deep neural network before learning , so that you can learn. In the data indicating good or bad, it includes the data indicating that the welding result is good, the data indicating that the welding failure is caused by a short cycle, the data indicating that the welding failure is caused by a long cycle, and the data indicating that the welding is caused by other reasons. Bad welding information. When the cycle is too short, the molten metal at the beginning of the welding or the end of the welding is insufficient, and there is a problem of welding. Tendency to connect badly. In addition, when the cycle is too long, the molten metal at the welding start portion or the welding end portion becomes excessive, and there is a tendency for poor welding.

In addition, the structure of the intermediate layer of the RNN 53a, the number of neurons in each layer, and the like are not particularly limited. In addition, the RNN 53a is not necessarily a recurrent neural network, and may be constituted by other types of neural networks.

The second quality judgment unit 52b has a quality judgment CNN (Convolutional Neutral Network) 53b, and the quality judgment CNN 53b outputs, when the image data is input, the data indicating whether the welding result is good or not related to the welding step when the image data is obtained. . Good or bad judgment CNN53b is a convolutional neural network that has completed learning.

Good or bad judgment CNN53b includes, for example, in the output layer: the fifth neuron, which outputs the data indicating the probability that the welding result is good; the sixth neuron, which outputs the data indicating the probability that the welding result is bad; the seventh neuron, which outputs the data indicating the probability that the welding result is bad; The output indicates data indicating the probability of poor welding caused by a short cycle; and the eighth neuron outputs data indicating the probability of poor welding caused by a long cycle. In this case, the data indicating good or bad is the data output from the 5th to 8th neurons. In addition, in general, the image data showing the appearance of the bead after welding when the welding defect caused by the short cycle occurs, and the image data showing the appearance of the bead after welding when the welding defect caused by the long cycle occurs are related. Since they have different characteristics, the second quality determination unit 52b can determine whether the welding failure is caused by a short cycle or a long cycle.

In addition, it is also possible to judge whether the CNN53b has a neuron in the output layer that outputs the good or bad of the welding result as a binary value. In this case, the data indicating good or bad is the binary data output from the neuron.

Furthermore, it is also possible to judge whether the CNN 53b is good or not. The output layer has a neuron that outputs a ratio indicating the degree of good or bad of the welding result.

Good or bad judgment CNN53b only needs to use the image data (input data) and the data (teacher data) corresponding to the image data indicating whether the welding result is good or not as the learning data, and to the convolutional neural network before learning. Supply, so that you can learn. In the data indicating good or bad, it includes the data indicating that the welding result is good, the data indicating that the welding failure is caused by a short cycle, the data indicating that the welding failure is caused by a long cycle, and the data indicating that the welding is caused by other reasons. Bad welding information.

In addition, the number of layers in the middle layer, the number of neurons in each layer, and the like of the good-failure determination CNN53b are not particularly limited. In addition, the good-failure determination CNN53b is not necessarily a convolutional neural network, and may be composed of other types of neural networks.

The quality comprehensive judgment unit 54 judges the quality of the welding result based on the data output from the first quality judgment unit 52 a and the second quality judgment unit 52 b , and outputs the judgment result to the adjustment unit 55 .

For example, the quality judgment unit 54 comprehensively judges the data output from the first to fourth neurons of the quality judgment RNN53a and the data output from the fifth to eighth neurons of the quality judgment CNN53b. Specifically, by comparing the sum of the values of the data output from the first neuron and the fifth neuron, and the sum of the values of the data output from the second neuron and the sixth neuron, the third neuron The sum of the values of the data output from the seventh neuron, and the sum of the values of the data output from the fourth neuron and the eighth neuron can be used to determine whether the next welding result is good or not. Alternatively, the values of data output from each neuron may be weighted and added for comparison.

In addition, in the case where binary data is output from the good/bad judgment RNN 53a and the good/bad judgment CNN 53b, the good/bad comprehensive judgment section 54 outputs the data indicating that it is good at both the first good/bad judgment section 52a and the second good/bad judgment section 52b. , when the welding result of the next welding step is judged to be good, and one of the first good or bad judgment part 52a and the second good or bad judgment part 52b outputs data indicating that it is bad, the welding result of the next welding step is judged to be bad . In addition, the method of comprehensive determination is an example, and it is also possible to configure the welding result of the next welding step to be determined to be good when either the first quality determination unit 52a or the second quality determination unit 52b outputs data indicating good quality.

The adjustment part 55 adjusts the arc welding conditions so that the cycle of the welding step is shortened when the judgment result of the quality comprehensive judgment part 54 is good, or the welding is defective due to a long cycle. is short, and in the case of poor welding caused by a short cycle, or in the case of poor welding caused by other reasons, the arc welding conditions are adjusted so that the cycle of the welding step is extended, and the output is output to the welding control unit 56 Adjust the result. The adjustment result is, for example, data representing various parameters of arc welding conditions, that is, welding current, welding voltage, wire feed speed, and welding torch moving speed. Information on at least one increase or decrease of various parameters of welding conditions.

The adjustment unit 55 performs adjustment processing at one time, and may change the values of a plurality of parameters, or may change the value of one parameter. In addition, when the arc welding conditions are adjusted so that the cycle becomes the shortest by repeatedly executing steps S11 to S21 described later, it is also possible to configure different parameters to be adjusted by repeating each adjustment process.

In addition, the increase and decrease amount of each parameter may be correlated, and the variable may be reduced. In addition, it may be configured to limit the amount of change from the standard parameter value to a range of a predetermined ratio.

In addition, the adjustment unit 55 is determined to shorten the cycle of the welding step. When the welding result changes from a good state to a poor state, the shortest welding condition memory unit 57 memorizes the arc welding conditions before the cycle shortening.

Furthermore, in the case where the determination result is poor welding due to other causes, the above description has been made to adjust the arc welding conditions so that the cycle of the welding step is extended, but it is also possible to adjust the parameters so that the determination result is performed in the maintenance cycle. into good.

The welding control unit 56 performs welding control by outputting control data for changing arc welding conditions to the welding power source 2 based on the adjustment result of the adjustment unit 55 . However, when the cycle of welding steps memorized by the shortest welding condition memory unit 57 becomes the shortest arc welding condition, the welding control unit 56 performs welding control based on the arc welding conditions memorized by the shortest welding condition memory unit 57 .

Next, the processing routine of the control unit 50 related to the adjustment of arc welding conditions will be described.

4 and 5 are flowcharts showing a welding condition adjustment method according to the first embodiment. The control unit 50 is, for example, The following processes are repeatedly performed for each welding step in continuous production. The control part 50 judges whether the arc welding condition with the shortest period is memorized in the memory part 50c (step S11). When it is determined that the arc welding condition with the shortest period is memorized (step S11: YES), the control unit 50 controls welding based on the arc welding condition with the shortest period memorized in the memory unit 50c (step S12). For example, the control part 50 performs welding control by outputting the control data which shows the arc welding condition with the shortest cycle to the welding power source 2. Of course, the control unit 50 may perform welding control by outputting to the welding power source 2 control data indicating a change amount for setting the arc welding conditions with the shortest cycle.

When it is determined that the arc welding condition with the shortest period is not stored in the memory unit 50c (step S11: NO), the control unit 50 acquires welding monitoring data (step S13), and then acquires image data (step S14). Then, the control unit 50 determines whether the welding result is good or not based on the obtained welding monitoring data and image data (step S15). For example, the control part 50 judges the quality of the welding result using the quality judgment RNN53a and the quality judgment CNN53b which completed the learning.

Next, when it is determined that the welding result is good (step S15: YES), the control unit 50 shortens the cycle of the welding step (step S16). When it is determined that the welding result is defective (step S15: NO), it is determined whether the welding result has changed from a good state to a defective state as a result of shortening the welding step last time (step S17). Of course, the change to the defective state is not limited to a configuration that is judged based on only one welding result, but also includes a configuration that uses two or more welding results to judge. For example, it may be determined that the welding result has changed to the defective state when the welding result is in a defective state by a fixed ratio or more in 10 times. When it is determined that the welding result has not changed from a good state to a defective state (step S17: NO), the control unit 50 determines whether or not the welding is defective due to a short cycle (step S18). When it is determined that it is not (step S18: NO), the control unit 50 determines whether or not the welding is defective due to a long cycle (step S19). When the control unit 50 judges that the welding is defective due to a short cycle (step S18: YES), or when it is judged not to be at step S19, the control unit 50 prolongs the cycle of the welding step (step S20). When the control unit 50 determines that the welding defect is caused by a long cycle (step S19: YES), the control unit 50 shortens the welding step cycle of the steps (step S16). The control part 50 which completed the process of step S16 or step S20 performs welding control based on the adjusted arc welding conditions (step S21). Specifically, the control unit 50 performs welding control by outputting control data indicating arc welding conditions after adjustment processing to the welding power source 2 . Of course, the control unit 50 may perform welding control by outputting control data indicating the amount of change in arc welding conditions to the welding power source 2 .

When it is determined that the cycle of the welding step is shortened and the welding result is changed from a good state to a poor state (step S17: YES), the control unit 50 returns the cycle of the welding step to the arc welding condition before the cycle shortening (step S22). , and the arc welding condition before the cycle shortening is stored in the memory unit 50c as the arc welding condition with the shortest cycle (step S23), and the process returns to step S12.

According to the welding condition adjusting device 5, the welding system, the welding condition adjusting method and the computer program 50d constructed in this way, at least one welding interval between the arc starting step and the arc ending step can be made without deteriorating the welding result. The welding control becomes the best, and the cycle of welding steps can be effectively shortened.

In addition, since the memory section 50c stores the arc welding condition with the shortest period, the welding condition adjusting device 5 can control welding so that the period of the welding step can be quickly shortened.

In addition, in the first embodiment, the welding condition adjusting device 5 has been described as including the quality judgment RNN 53a and the quality judgment CNN 53b for which the learning has been completed, but it may be configured to download the predetermined first quality judgment unit 52a and the second quality judgment unit 52a from an external server. Various parameters of the neural network of the good or bad judgment unit 52b are updated later. The parameter is information including, for example, the number of layers in the middle layer, the number of neurons in each layer, the weighting coefficient of each neuron, the type of activation function, and the like. In addition, the welding condition adjusting device 5 may be configured as a memory flag indicating whether or not to reflect the downloaded parameters to the first quality judgment unit 52a and the second quality judgment unit 52b, and the flag indicates the allowable parameters. In this case, use the downloaded parameters to update the neural networks of the good or bad judgment RNN53a and the good or bad judgment CNN53b.

In addition, in the case where a plurality of welding systems having the welding condition adjusting device 5 are installed in a factory, the parameter can be replaced by the welding condition adjusting device 5 of each welding system as needed.

Furthermore, the welding condition adjustment device 5 may be constituted by a cloud server. Alternatively, the welding power source 2 or the control device 3 may request the server to adjust the arc welding conditions, and adjust the arc welding conditions by receiving the adjustment amount of the arc welding conditions transmitted from the server in response to the request.

Furthermore, the welding condition adjustment device 5 may be included in the welding power source 2 . In addition, the welding condition adjustment device 5 may be implemented as a dedicated device for arc welding condition adjustment. The operator connects this special device to the welding system, and the arc welding conditions can be adjusted automatically.

Furthermore, an example in which the first quality determination unit 52a and the second quality determination unit 52b include the quality determination RNN53a and the quality determination CNN53b has been described, but both or one of the determination units may be configured to determine the welding result without using a neural network. Good or not. For example, the first quality determination unit 52a may perform a simple determination process of comparing the welding current value and a predetermined threshold value to determine whether the quality is satisfactory. Alternatively, the second good or bad judgment unit 52b may perform a simple judgment process of extracting a predetermined feature value from the video data, and comparing the presence or absence of the feature value, the number of feature values, and the like with a threshold value to judge whether the feature value is good or not. Furthermore, it is not necessary to include both the first quality determination unit 52a and the second quality determination unit 52b, and either one may be included. In this case, the quality comprehensive judgment unit 54 is unnecessary. Furthermore, an example in which the first quality determination unit 52a and the second quality determination unit 52b are configured by a neural network has been described, but it may be configured using another machine learning device such as a support vector machine.

(Embodiment 2)

The welding condition adjustment device 205, the welding system, the welding condition adjustment method, and the computer program 50d of the second embodiment are similar to those of the first embodiment in that the adjustment part 55 and the shortest welding condition memory part 57 of the first embodiment are constituted by a deep neural network. Therefore, the difference is mainly explained in the following section. Since other structures and effects are the same as those of the first embodiment, the same reference numerals are attached to the corresponding parts, and the detailed description is omitted.

FIG. 6 is a functional block diagram showing the welding condition adjusting device 205 according to the second embodiment. Reality The welding condition adjustment device 205 of the second embodiment is the same as that of the first embodiment, and includes a welding monitoring data acquisition unit 51a (acquisition unit), an image data acquisition unit 51b (acquisition unit), a first quality judgment unit 52a (quality judgment unit), The second quality determination unit 52 b (quality determination unit), the quality comprehensive determination unit 254 , the adjustment unit 255 , the learning processing unit 259 , and the welding control unit 56 .

The welding monitoring data acquisition unit 51 a acquires the welding monitoring data output from the welding power source 2 , and outputs the acquired welding monitoring data to the first quality determination unit 52 a and the adjustment unit 255 .

The image data acquisition unit 51 b acquires the image data output from the imaging device 4 , and outputs the acquired image data to the second quality determination unit 52 b and the adjustment unit 255 .

The quality comprehensive judgment unit 254 of the second embodiment outputs to the learning processing unit 259 the data indicating the probability that the welding result is good, the data indicating the probability that the welding result is bad, the data indicating the probability that the welding result is poor due to the short cycle, And data indicating the probability of poor welding caused by a long cycle. For example, the probability that the welding result is good may be calculated by using the value of the data output from the first neuron of the good or bad judgment RNN53a and the value of the data output from the fifth neuron of the good or bad judgment CNN53b. Similarly, the probability that the welding result is bad can be calculated by using the value of the data output from the second neuron of the good-fail judgment RNN53a and the data output from the sixth neuron of the good-fail judgment CNN53b. Also, as long as the values of the data output from the 3rd and 4th neurons of the RNN53a and the data output from the 7th and 8th neurons of the CNN53b are used, the reason for the poor welding is calculated due to the length of the cycle. the probability can be.

The adjustment unit 255 includes an adjustment NN (Neutral Network) 258 that outputs data indicating the amount of change in arc welding conditions in at least one welding section that can shorten the cycle when welding monitoring data and image data have been input. Tuning NN258 is a deep neural network that completes learning.

The adjustment NN258 system has a plurality of neurons in the output layer, and the plurality of neurons output a plurality of adjustment parameters such as welding current, welding voltage, welding wire feeding speed and welding gun moving speed. Each adjustment amount represents the data of the probability that the adjustment amount is better.

In addition, the configuration of the NN 258 may be adjusted to include neurons in the output layer that output data representing the adjustment amount. Furthermore, the NN 258 can also be adjusted to include a neuron in the output layer that outputs the adjustment amount with binary data. Hereinafter, in the second embodiment, it is assumed that the adjustment NN 258 does not output binary data, but outputs data indicating the probability that the change amount is appropriate for the change amount of each parameter.

In addition, the adjustment part 255 may change the value of a plurality of parameters in one adjustment process, and may change the value of one parameter. In addition, it is also possible to configure different parameters to be adjusted by repeating each adjustment process.

FIG. 7 is a schematic diagram showing the network configuration of the adjustment unit 255 . The adjustment NN258 of the adjustment part 255 has a welding state recognition network part 258a, an appearance state recognition network part 258b, and an adjustment network part 258c.

The welding state identification network part 258a is a neural network for inputting welding monitoring data, and after recognizing the welding state in the welding step, and outputting the data corresponding to the state. When the welding monitoring data is the welding current, the welding state identification network unit 258a can identify the changing state of the welding current. The welding state recognition network unit 258a can have the same neural network structure as the first quality determination unit 52a except for the output layer, for example. The output layer has a plurality of neurons, preferably three or more neurons. In addition, as the initial value of the weighting coefficient before learning, the weighting coefficient of each neuron constituting the first quality determination unit 52a can be set. The adjustment unit 255 can learn more efficiently.

The appearance state recognition network part 258b is a neural network for inputting image data, and after recognizing the state of the welding part after welding, and outputting the data corresponding to the state. The appearance state recognition network unit 258b can have the same neural network structure as the second quality determination unit 52b except for the output layer, for example. The output layer has a plurality of neurons, preferably three or more neurons. In addition, as the initial value of the weighting coefficient before learning, the weighting coefficient of each neuron constituting the second quality determination unit 52b can be set. Allows tuning NN258 to learn more efficiently.

The adjustment network part 258c is input from the welding state recognition network part 258a and the appearance state recognition network part 258b separates the output data, and outputs the data representing the change amount of the arc welding conditions which can shorten the cycle time of the neural network which completes the learning. The neural network is preferably formed by a deep neural network with a plurality of intermediate layers.

In addition, the neural network configuration of the adjustment unit 255 is an example, and it may be configured by a single neural network, or a plurality of neural networks may be combined.

The learning processing unit 259 uses the welding monitoring data and the image data inputted by the adjusting unit 255 as input data, and uses the data indicating whether the welding result is good or not when the arc welding conditions are changed according to the input data as the learning data, so that the adjustment NN258 learns processing department.

Specifically, the learning processing unit 259 causes the adjustment NN 258 to learn as follows, and from the judgment result of the comprehensive judgment unit 254, when the welding result is good, or when the welding is defective due to a long cycle, the cycle is shortened ; In the case of poor welding caused by short cycle or other reasons, prolong the cycle; in the case of the intermediate state between good and bad welding results, maintain the cycle.

The welding result is good is, for example, a state in which the probability of the welding result being good is 50% or more and the probability of the welding result being poor is less than 50%. The welding result is bad is, for example, a state in which the probability of the welding result being good is less than 50% and the probability of the welding result being bad is 50% or more. 50% of the threshold value is an example, and may be a value larger than 50%.

The welding result is an intermediate state, for example, when both the probability that the welding result is good and the probability that the welding result is bad are 50% or more, or both are less than 50%. In addition, when the threshold value is larger than 50%, for example, 60%, such as the case where the welding result is good and the probability of failure is between 40% and 60%, it is also an intermediate state. In addition, this intermediate state is an example. The intermediate state is a state in which the cycle time is to be shortened, and there is a possibility that the welding result may deteriorate.

As described above, by learning the adjustment NN258, the time required for welding can be automatically adjusted without deteriorating the welding result, and the cycle of the welding step can be minimized.

In addition, it is also possible to adjust the initial stage of NN258 learning, in the case of an intermediate state, without dimension The arc welding conditions are appropriately changed according to the maintenance cycle.

In addition, the learning system for adjusting the NN258 can be performed at an appropriate timing, such as when the welding system has been installed, when the external environment changes, when the welding conditions are changed, and when other layout changes are made.

In addition, in this Embodiment 2, although the example which made the adjustment NN258 learn was demonstrated, it is good also as a structure which has the adjustment NN258 which completed the learning, and does not perform further learning.

According to the welding condition adjustment device 205, the welding system, the welding condition adjustment method, and the computer program 50d according to the second embodiment configured in this way, the adjustment unit 255 constituted by the deep neural network determines the amount of change in the arc welding conditions Therefore, the welding control can be performed more appropriately and the cycle time can be shortened without deteriorating the welding result.

Moreover, the adjustment part 255 can output the change amount of the arc welding condition which shortens the cycle of a welding step. The adjustment unit 255 can output a large amount of change when the welding result is stable, for example, and output a small amount of change when the welding result is good but not stable. Therefore, the cycle of the welding step can be shortened more quickly.

Furthermore, the adjustment part 255 can be adjusted to the adjustment part 255 suitable for the environment in which a welding system is installed by learning using the judgment result of whether the welding result is good or not. Therefore, the cycle of the welding step can be minimized according to the welding conditions and the external environment.

Furthermore, the learning processing unit 259 can make the adjustment neural network learn in the following manner, and after the adjustment unit 255 learns in the direction of shortening the cycle of the welding step, and can output the data that can make the cycle of the welding step shortest, Good welding results are obtained stably. Therefore, the welding result can be stabilized in a good state, and the cycle of the welding step can be minimized.

In addition, although the welding condition adjustment apparatus 205 has demonstrated the example which has the adjustment NN258 which completed the learning in this Embodiment 2, it can also be comprised so that various parameters which define the neural network of the adjustment part 255 may be downloaded from an external server and updated. The parameter is information including, for example, the number of layers in the middle layer, the number of neurons in each layer, the weighting coefficient of each neuron, the type of activation function, and the like. Also, welding conditions The adjustment device 205 is configured as a memory flag, and the flag indicates whether the downloaded parameters are allowed to be reflected to the adjustment unit 255. If the flag indicates permission, the downloaded parameters are used to update and adjust the neural network of the NN 258.

In addition, when a plurality of welding systems having the welding condition adjusting device 205 are installed in a factory, the parameter can be replaced by the welding condition adjusting device 205 of each welding system as needed.

In addition, the welding condition adjustment device 205 may be configured to upload various parameters for the adjustment of the NN 258 after the predetermined learning to an external server. Other welding condition adjustment devices 205 can use the parameter uploaded to the server to update and adjust the NN 258 .

In addition, in the second embodiment, an example in which the adjustment NN 258 and the first quality judgment unit 52a and the second quality judgment unit 52b have neural networks have been described, but both or one of the quality judgment RNN 53a and the quality judgment CNN 53b may be configured. In order to judge whether the welding result is good or not without using a neural network.

(Embodiment 3)

FIG. 8 is a functional block diagram showing the welding condition adjusting device 305 according to the third embodiment. The welding condition adjusting device 305, the welding system, the welding condition adjusting method, and the computer program 50d according to the third embodiment are configured to form the adjusting unit 355 and the shortest welding condition memory unit 57 of the first embodiment so that arc welding conditions are learned by deep reinforcement learning. The above is different from Embodiment 1, so the following mainly describes the difference. Since other structures and effects are the same as those of the first embodiment, the same reference numerals are attached to the corresponding parts, and the detailed description is omitted.

The welding condition adjustment device 305 of the third embodiment includes a welding monitoring data acquisition unit 51a (acquisition unit), an image data acquisition unit 51b (acquisition unit), a state data acquisition unit 51d (acquisition unit), and a first quality judgment unit 52a (quality judging unit), a second quality judgment unit 52 b (quality judgment unit), a quality comprehensive judgment unit 54 , an adjustment unit 355 , and a welding control unit 56 .

The state data acquisition unit 51d acquires state data indicating the state s of the welding system. The state data includes, for example, image data obtained by photographing the welding portion after welding. Welded parts can also be borrowed The image capturing device 4 can also use other moving image capturing devices to capture the image.

In addition, in the third embodiment, image data is mainly described as an example of the state data, but current data representing the current flowing to the bonding wire W and voltage data representing the voltage may also be used.

The adjustment unit 355 learns arc welding conditions for at least one welding section in which the cycle of the welding step is the shortest by deep reinforcement learning, and includes an evaluation unit 355a, an action selection unit 355b, a reward calculation unit 355c, and a reinforcement learning unit 355d.

The evaluation unit 355a is an arithmetic function unit that calculates a pair of states indicated by the state data based on the state data acquired by the state data acquisition unit 51d and the action data indicating the action a related to the arc welding conditions. The evaluation value Q of the action a. The state is, for example, an image showing the appearance of the bead. Action a related to arc welding conditions is determined according to welding current, welding voltage, wire feed speed, torch moving speed, and the like. The evaluation value Q is a higher value when a certain action a is taken according to the appearance of the ball joint, the shorter the cycle of the welding step can be suitably, and the better the welding result is.

The evaluation unit 355a has an evaluation NN (Neutral Network) 355e that, for example, has input state data representing the state s of the welding system acquired by the state data acquiring unit 51d, and represents actions related to arc welding conditions. In the case of the action data of a, the evaluation value Q(s, a) of the action a in the state s is output.

In addition, the evaluation NN355e can have a convolutional neural network in the front stage, which is used to identify the status data that represents the status of the welding system in images.

In addition, instead of the neural network, the evaluation unit 355a may have a table in which state data, action data, and evaluation values are related, and may be configured to output evaluation values using this table.

The action selection unit 355b selects the action a in which the evaluation value Q calculated by the evaluation unit 355a is the largest in a certain state s. The adjustment part 355 adjusts the arc welding conditions according to the action a selected by the action selection part 355b, and the welding control part 56 performs welding control according to the adjusted arc welding conditions.

The reward calculation unit 355c calculates the reward for the arc welding conditions based on the judgment result output from the good/fail comprehensive judgment unit 54 and the time from when the welding torch 11 reaches the welding position until the arc is generated. The reward is calculated in such a way that the better the welding result, the shorter the time until the arc is generated, and the higher the value. The mathematical formula for calculating the reward is not particularly limited.

The reinforcement learning unit 355d causes the evaluation NN 355e to learn based on the state data and action data input by the evaluation NN 355e, the evaluation value Q output when each data has been input, and the reward calculated by the reward calculation unit 355c. Specifically, the weighting coefficient of the neural network can be learned based on the evaluation value Q represented by the following mathematical formula (1).

Q(s,a)←Q(s,a)+α(r+γmaxQ(s_next,a_next)-Q(s,a))...(1)

in,

s: status

a: the action selected in state s

α: Learning coefficient

r: the reward for the result of the action

γ: discount rate

maxQ(s_next, a_next): the maximum value of the evaluation value Q of the actions that can be taken in the next state

The learning coefficient α is a positive value equal to or less than 1, for example, a value of about 0.1. The discount rate γ is a positive value of 1 or less, for example, a value of about 0.9.

By using the machine learning of the above-mentioned mathematical formula (1), the evaluation NN355e can be made to learn by assigning a higher evaluation value Q to the action a that can obtain a higher reward. In addition, when performing reinforcement learning, an ε-greedy algorithm or the like can be used which randomly moves with a fixed probability and learns the Q value for various actions.

According to the welding condition adjustment device 305 constructed in this way, the action selection unit 355b can select a more suitable action a according to the state s of the welding system, that is, the welding current, the welding voltage, and the wire feed speed As well as the moving speed of the welding torch, the cycle of the welding step can be minimized.

In addition, the reinforcement learning of the welding condition adjustment device 305 may be performed using data such as status data, action data, and evaluation values obtained from a plurality of welding devices or welding systems, data obtained from a simulation device, and the like.

According to the welding condition adjusting device 305, the welding system, the welding condition adjusting method, and the computer program 50d according to the third embodiment, deep reinforcement learning can be performed on arc welding conditions that shorten the cycle of the welding steps.

In addition, in the above-mentioned Embodiment 3, deep reinforcement learning has been described, but instead of the neural network, a matrix may be provided for the evaluation value Q corresponding to the action and the state, and the arc welding conditions may be adjusted.

(Embodiment 4)

FIG. 9 is a schematic diagram showing an arc welding system according to Embodiment 4. FIG. The welding system of Embodiment 4 is different from Embodiments 1 to 3 in that it has an adjustment method accepting unit 406 that accepts an adjustment method of arc welding conditions. The difference will be mainly described below. Since other structures and effects are the same as those of Embodiments 1 to 3, the same reference numerals are attached to corresponding parts, and detailed descriptions are omitted.

FIG. 10 is a schematic diagram showing the adjustment screen 407 . The adjustment method accepting unit 406 displays an adjustment screen 407 on a terminal, for example, and accepts an adjustment method of arc welding conditions. The adjustment screen 407 includes, for example, an adjustment method selection unit 471 that accepts selection of automatic adjustment or manual adjustment of arc welding conditions. The adjustment method selection unit 471 is, for example, a wireless button, and the user can select automatic adjustment or manual adjustment of arc welding conditions by checking the wireless button.

In addition, the adjustment screen 407 includes: a priority display unit 472, which indicates that, when automatic adjustment of arc welding conditions is selected, an adjustment that prioritizes the shortest cycle or an adjustment that emphasizes welding quality should be performed; and a priority Adjustment slider 473 is used to specify the priority. By sliding the priority adjustment slider 473, the user can specify how much welding quality should be emphasized to adjust the arc. The arc welding conditions are adjusted according to the welding conditions, or to what extent the cycle minimization should be emphasized.

The adjustment method accepting section 406 that accepts the priority related to the welding quality by the priority adjustment slider 473 corresponds to the accepting section that accepts the adjustment strength of the arc welding conditions adjusted by the adjustment section 55 .

For example, the control device 3 controls the arc welding conditions in such a way that the higher the importance of the welding quality, the smaller the adjustment amount of the arc welding conditions, and the higher the importance of the shortest cycle, the smaller the arc welding conditions. Welding control is performed in such a way that the adjustment amount becomes larger.

According to the welding system of the fourth embodiment, the importance or priority of minimizing the welding quality or cycle time can be received, and the arc welding conditions can be adjusted according to the method desired by the user.

It should be construed that the embodiments disclosed this time are illustrative in all matters and are not intended to be limiting. The scope of the present invention is not the meaning described above, but is indicated by the scope of the patent claim, and is intended to include the meaning equivalent to the scope of the patent claim and all changes within the scope.

1: Welding robot

2: Welding power source

3: Control device

4: Camera device

5: Welding condition adjustment device

11: Welding gun

12: Welding wire feeding device

21: Power Department

22: Wire feed control section

23: Shielding gas supply part

24: Detection Department

A: Base material

W: welding wire

Claims (9)

A welding condition adjustment device, which is a welding condition adjustment device for adjusting arc welding conditions, comprising: an acquisition unit that acquires welding data representing welding states related to a welding step; and an adjustment unit based on the acquisition unit the welding data, adjust the arc welding conditions associated with the cycle of at least one welding interval between the arc starting step and the arc ending step, wherein the arc welding conditions include the welding current and welding voltage of the actual welding performed in the welding interval , the feeding speed of the welding wire, the supply amount of the shielding gas; and the quality judgment part, which judges whether the welding result is good or not and the defect caused by the short cycle according to the welding data obtained by the acquisition part; the adjustment part The content of the arc welding conditions to be changed in relation to the cycle is determined in such a manner that the cycle is shortened if the cycle is determined to be good, and the cycle is longer if the cycle is determined to be defective due to the shorter cycle. the cycle. The welding condition adjusting device of claim 1, wherein the adjusting portion is a result of shortening the cycle of the welding step, and in the case where the welding result changes from a good state to a poor state, according to the arc welding associated with the cycle before the cycle shortening The conditions are determined and adjusted, and the arc welding conditions determined are memorized in the memory unit. The welding condition adjustment device according to claim 1, wherein the adjustment section has an adjustment neural network that outputs data indicating the content of changes in arc welding conditions associated with the cycle when the welding data has been input. Make the neural network learn. The welding condition adjustment device of claim 3, wherein the adjustment neural network outputs data representing the amount of change in the arc welding conditions. The welding condition adjusting device according to claim 3 or 4, comprising: a learning processing unit based on the judgment of the good or bad judgment unit obtained after adjusting the arc welding conditions As a result, the tuning neural network is made to learn. According to the welding condition adjustment device of claim 1, there is a state data acquisition part for acquiring state data, and the state data at least includes image data obtained by photographing the welding part after welding; the adjustment part includes: an evaluation part, which is based on borrowed The status data acquired by the status data acquisition unit and the action data indicating the action related to the arc welding condition calculates the evaluation value of the action in the state indicated by the status data; and the action selection section is selected by borrowing The action with the largest evaluation value calculated by the evaluation department. The welding condition adjustment device according to claim 6, further comprising: a reward calculation unit that calculates the arc welding conditions based on the determination result of the good/failure determination unit obtained after adjusting the arc welding conditions and the time required for the welding step and the Reinforcement Learning Department, based on the state data acquired by the state data acquisition department, the action data indicating the action related to the arc welding condition, and the reward calculated by the reward calculation department, the evaluation department study. The welding condition adjusting device according to any one of claim 1 to claim 4, claim 6 or claim 7, wherein there is an accepting part for accepting the adjustment strength of arc welding conditions adjusted by the adjusting part; the adjusting part is based on the accepting part. Adjust the strength and adjust the arc welding conditions accepted by the department. The welding condition adjusting device according to claim 5, further comprising an accepting unit for accepting the adjustment strength for adjusting the arc welding conditions by the adjusting unit; the adjusting unit adjusts the arc welding conditions according to the adjustment strength accepted by the accepting unit.

Images