JP2020163453A - Generation method of learned model, computer program, welding power supply device, and learned model - Google Patents

Abstract

To provide a generation method for generating a learned model which efficiently generates and stores learning data, and accurately determines the quality of a weld state from arc sound.SOLUTION: A generation method of a learned model includes the steps of: acquiring arc sound data obtained by collecting arc sound produced during welding by using a microphone; imaging a welded site after welding or scanning the welded site using detection light to acquire detection data obtained by detection of a state of the welded site; determining the quality of a weld state on the basis of the acquired detection data; storing the acquired arc sound data and learning data associated with quality data indicating the quality of the weld state, in a storage part; and generating, upon input of arc sound data, a learned model that outputs quality data corresponding to the arc sound data, on the basis of the learning data stored in the storage part, by using machine learning.SELECTED DRAWING: Figure 4

Description

The present invention relates to a method of generating a trained model, a computer program, a welding power supply, and a trained model.

The consumable electrode type welding power supply device includes a welding power supply, a wire feeding device, a welding torch, and a gas supply unit that injects shield gas around an arc. Patent Document 1 discloses a technique in which an arc sound is collected by a microphone provided on a welding torch and the quality of the welding state is determined from the sound pressure level of a specific frequency of the arc sound. Further, Patent Document 1 discloses a technique of collecting the flowing sound of the shield gas and determining the injection state of the shield gas based on the sound pressure level of a specific frequency.

Japanese Unexamined Patent Publication No. 2014-113637

However, in the welding power supply device according to Patent Document 1, the data of the arc sound, which is a reference for judging whether or not the welding state is good with respect to the collected arc sound, is not sufficient, and the judgment of good or bad occurs. There was a technical problem that it would end up. Further, when the welding conditions such as the material, thickness, groove shape, joint type, work W attitude, welding wire material, wire diameter, welding current, and welding voltage of the work change, the problem of judgment variation becomes more remarkable. become.
In order to solve such a problem, it is conceivable to analyze a large amount of data indicating the quality of the welding state with respect to the arc sound, but it is generally difficult to efficiently prepare such data.

An object of the present invention is a generation method, a computer program, a welding power supply device, and learning capable of efficiently generating and accumulating learning data and generating a trained model capable of accurately determining the quality of a welded state from an arc sound. It is to provide a finished model.

The trained model generation method according to this aspect is a trained model generation method that outputs quality data indicating the quality of the welding state when arc sound data is input, and is an arc sound generated during welding. The detection data obtained by detecting the state of the welded part is obtained by acquiring the arc sound data obtained by collecting the data with a microphone, imaging the welded part after welding, or scanning the welded part with detection light. Based on the acquired detection data, the quality of the welding state is determined, and the learning data associated with the acquired arc sound data and the quality data indicating the quality of the welding state is stored in the storage unit. Based on the learning data stored in the storage unit, when the arc sound data is input, a trained model that outputs the good / bad data corresponding to the arc sound data is generated by machine learning.

According to this aspect, it is possible to efficiently prepare the learning data necessary for generating the trained model in order to judge the quality of the welding state from the arc sound. The learning data includes arc sound data acquired during welding and quality data indicating the quality of the welding state when the arc sound is generated.
The training data is obtained by collecting the arc sound generated during welding with a microphone. However, the quality of the welded state when the arc sound is generated is not immediately known.
Therefore, by imaging the welded portion after welding, the detection data obtained by detecting the state of the welded portion is acquired. Further, the detection data obtained by detecting the state of the welded portion may be acquired by scanning the welded portion with the detection light. Then, the quality data is obtained by determining the quality of the welding state based on the detection data. It is relatively easy to determine the quality of the welded state from the detection data related to the welded part, and it is possible to automatically obtain the quality data.
The arc sound data thus obtained is associated with the quality data and stored in the storage unit as learning data. Next, using the learning data accumulated in the storage unit, a trained model for determining the quality of the welding state from the arc sound is generated. The trained model can be generated by machine learning the pre-trained neural network. When the arc sound data is input to the generated trained model, the quality data indicating the quality of the welding state is output, so that it is possible to judge the quality of the welding state from the arc sound.

The method of generating the trained model according to this aspect is to input the acquired detection data into the trained model for quality determination that outputs data indicating the quality of the welding state when the detection data is input. Judge the quality of the welded condition.

According to this aspect, learning data is stored in a storage unit for each of a plurality of different welding conditions, and a trained model is generated for each of a plurality of welding conditions. Therefore, a trained model suitable for each welding condition can be generated. By machine learning for each welding condition, a trained model can be efficiently generated.

The method of generating the learned model according to this aspect is based on the learning data of the plurality of welding conditions accumulated in the storage unit by accumulating the learning data for each of a plurality of different welding conditions. Therefore, a plurality of the trained models that are different for each of the plurality of welding conditions are generated.

According to this aspect, the quality of the welding state can be accurately determined from the detection data by using the trained model for quality determination, which is a trained neural network.

The computer program according to this aspect is a computer program for causing a computer to generate a trained model that outputs quality data indicating the quality of the welded state when arc sound data is input. The state of the welded part is detected by acquiring the arc sound data obtained by collecting the arc sound generated during welding with a microphone, imaging the welded part after welding, or scanning the welded part with detection light. The detection data obtained is obtained, the quality of the welding state is determined based on the acquired detection data, and the acquired arc sound data and the quality data indicating the quality of the welding state are associated with the learning data. Is stored in the storage unit, and when the arc sound data is input based on the learning data stored in the storage unit, a trained model that outputs the quality data corresponding to the arc sound data is provided. Execute the process generated by machine learning.

According to this aspect, by causing the computer to execute the computer program, it is possible to efficiently accumulate learning data and generate a trained model for accurately determining the quality of the welding state from the arc sound.

The welding power supply device according to this embodiment has a microphone that collects arc sound, a learned model that outputs quality data indicating the quality of the welding state when arc sound data is input, and a microphone that collects data. It is provided with a notification unit that inputs the obtained arc sound data to the trained model and notifies the quality of the welding state indicated by the quality data output from the trained model.

According to this aspect, the welding power supply device can determine the quality of the welding state from the arc sound and notify the user. Specifically, the welding power supply device acquires arc sound data during welding, causes the acquired arc sound data to be input to the trained model, and determines the quality of the welding state indicated by the quality data output from the trained model. Notify by the notification unit.

In the welding power supply device according to this embodiment, the trained models are plural, and each trained model is machine-learned for each of a plurality of different welding conditions, and corresponds to the welding conditions based on the welding conditions. A selection unit for selecting the trained model is provided, and the notification unit notifies the quality of the welding state indicated by the quality data output from the trained model selected by the selection unit.

According to this aspect, the welding power supply device selects a trained model corresponding to the welding conditions, determines the quality of the welding state using the selected trained model, and notifies the user. By using a learned model suitable for the welding conditions, it is possible to determine the welding state more accurately and notify the quality of the welding condition.

The trained model according to this aspect is a trained model in which a computer functions so as to output quality data indicating the quality of the welding state when arc sound data is input, and an arc generated during welding. An input layer into which arc sound data obtained by collecting sound with a microphone is input, an intermediate layer in which calculations based on learned weighting coefficients are performed on the arc sound data input to the input layer, and the arc. The arc sound data obtained by collecting the arc sound generated during welding with a microphone and having an output layer for outputting the quality data indicating the quality of the welding state when the sound related to the sound data is emitted, and the arc sound data. The arc sound input to the input layer based on the quality data obtained by detecting the state of the welded portion by imaging the welded portion or scanning the welded portion with detection light after welding. The computer is provided with a neural network in which the weighting coefficient is trained so that the data and the quality data output from the output layer correspond to the arc sound data input to the input layer and the arc sound data. The calculation based on the weight coefficient is performed, and the quality data is output.

According to this aspect, when the arc sound data is input to the input layer of the trained model, the quality data indicating the quality of the welding state when the arc sound is generated is output from the output layer. By using the trained model, it is possible to judge the quality of the welding state from the arc sound.

According to this aspect, it is possible to efficiently generate and accumulate learning data and generate a trained model for accurately determining the quality of the welding state from the arc sound.

It is a schematic diagram which shows the welding power-source device which concerns on Embodiment 1. FIG. It is a block diagram which shows the structural example of the welding torch which concerns on Embodiment 1. FIG. It is a block diagram which shows the structural example of the welding power source which concerns on Embodiment 1. FIG. It is a block diagram which shows the structural example of the learning part which concerns on Embodiment 1. FIG. It is a conceptual diagram which shows an example of the learning data. It is a flowchart which shows the generation method of the trained model. It is a flowchart which shows the quality determination and notification method of a welding state. It is a schematic diagram which shows the learning system which concerns on Embodiment 2.

Hereinafter, the present invention will be described in detail with reference to the drawings showing the embodiments thereof. In addition, at least a part of the embodiments described below may be arbitrarily combined.

(Embodiment 1)
FIG. 1 is a schematic view showing a welding power supply device according to the first embodiment. The consumable electrode type welding power supply device includes a welding power supply 1, a wire feeding device 2, and a welding torch 3. The welding power supply is semi-automatic.

The welding power supply 1 includes a first output terminal and a second output terminal for supplying electric power to the welding torch 3, and a signal terminal for transmitting and receiving signals. One end of the first power cable 41 is connected to the first output terminal of the welding power source 1, and the other end of the first power cable 41 is connected to the welding torch 3 via the wire feeding device 2. The second output terminal of the welding power source 1 is connected to the work W by the second power cable 42. The work W is grounded. The welding power supply 1 converts the three-phase AC of the power system P into the required welding current and welding voltage, and supplies the power required for arc welding to the welding torch 3 through the first and second power cables 41 and 42.

The welding power supply 1 and the wire feeding device 2 are connected by a power transmission line 5 for drive control. Further, the wire feeding device 2 and the welding torch 3 are also connected by the power transmission line 5. The welding power supply 1 transmits electric power for driving the feeding motor of the wire feeding device 2, the control unit 36 of the welding torch 3 (see FIG. 2), and the like through the power transmission line 5, the wire feeding device 2 and the welding torch 3. Supply to.

The welding power supply device includes a gas cylinder 6 that supplies a shield gas for preventing oxidation of the molten metal during arc welding. One end of the gas pipe 7 is connected to the gas cylinder 6, and the other end of the gas pipe 7 is connected to the welding torch 3 via the welding power supply 1 and the wire feeding device 2. The shield gas of the gas cylinder 6 is supplied to the welding torch 3 through the gas pipe 7.

One end of the signal line 8 is connected to the signal terminal of the welding power source 1, and the other end of the signal line 8 is connected to the wire feeding device 2. The wire feeding device 2 and the welding torch 3 are also connected by a signal line 8. The welding power supply 1 outputs a control signal to the wire feeding device 2 through the signal line 8 to control the feeding speed of the welding wire and the like. Further, the welding power supply 1 and the welding torch 3 transmit and receive various information necessary for the generation process of the learned model 154 for determining the quality of the welding state and the quality determination process of the welding state using the learned model 154.

The wire feeding device 2 is a device that feeds a welding wire that functions as a consumable electrode to the welding torch 3. The welding wire is, for example, a solid wire. The wire feeding device 2 and the welding torch 3 are connected by a torch cable 39, and the welding wire passes through the inside of the liner provided inside the torch cable 39 and the welding torch 3 and the tip of the welding torch 3. Guided to. The wire feeding device 2 includes a feeding roller for feeding the welding wire, a feeding motor, and the like, and is driven by the electric power supplied from the welding power source 1.
Further, inside the torch cable 39, a first power cable 41, a gas pipe 7, a liner, a power transmission line 5, and a signal line 8 are arranged. The drive control electric power supplied from the welding power source 1 to the wire feeding device 2 is also supplied to the welding torch 3 through the power transmission line 5 arranged inside the torch cable 39. The wire feeding device 2 can communicate with the welding power source 1 and the welding torch 3 through the signal line 8. Similarly, the welding power source 1 and the welding torch 3 can communicate with each other via the wire feeding device 2.

The welding torch 3 is made of a conductive material such as a copper alloy, and has a cylindrical contact tip that guides the welding wire to the work W to be welded and supplies the welding current required for generating an arc. The welded wire fed from the wire feeding device 2 is fed so as to protrude from the tip of the contact tip. The first power cable 41 is electrically connected to the contact chip. The contact tip comes into contact with the welding wire penetrating the inside thereof, and a welding current is supplied to the welding wire.
Further, the welding torch 3 has a hollow cylindrical shape surrounding the contact tip, and has a nozzle for injecting shield gas from the opening at the tip to the work W. The shield gas is for preventing oxidation of the work W and the welding wire melted by the arc. The shield gas is, for example, carbon dioxide gas, a mixed gas of carbon dioxide gas and argon gas, an inert gas such as argon, or the like.

<Welding torch>
FIG. 2 is a block diagram showing a configuration example of the welding torch 3 according to the first embodiment. The welding torch 3 includes a communication unit 31, a display unit 32, an operation unit 33, a storage unit 34, a sensor 35, a control unit 36, a microphone 37, and a notification unit 38.

The communication unit 31 is a circuit that communicates with the wire feeding device 2 or the welding power source 1. The communication unit 31 modulates the data given by the control unit 36 according to a predetermined communication protocol, and transmits the data to the wire feeding device 2 or the welding power source 1 through the signal line 8. The communication protocol is, for example, CAN (Controller Area Network). Further, the data transmitted from the welding power supply 1 and the wire feeding device 2 is received, demodulated, and the demodulated signal is given to the control unit 36. The communication between the welding torch 3 and the welding power source 1 is not limited to wired communication, and may be wireless communication.

The display unit 32 has a display for displaying various information related to welding. The display is, for example, a liquid crystal display panel, and displays welding conditions such as the material, thickness, groove shape, joint type, work W posture, welding wire material, wire diameter, welding current, and welding voltage of the work W. To do.

The operation unit 33 is various switches and buttons for accepting operations by the welder. By operating the operation unit 33, the welder can select a welding sound learning mode for generating the learned model 154 according to the present embodiment and a welding mode for performing welding using the learned model 154. .. When the operation unit 33 is operated, an operation signal is input to the control unit 36, and the control unit 36 can recognize the operation state of the operation unit 33.

The storage unit 34 is a non-volatile memory such as an EEPROM (Electrically Erasable Programmable ROM) or a flash memory. The storage unit 34 stores information such as welding conditions set by the operation of the operation unit 33. Welding conditions include, for example, the welding current and welding voltage recommended for the material, thickness, groove shape, joint type, work W posture, movement speed, posture and angle of a predetermined welding torch 3, and the like. This is information on set values such as the feeding speed of the welding wire. The welding conditions may be set on the welding power supply 1 side.

The sensor 35 is, for example, an imaging unit, and is provided on the welding torch 3 in a posture capable of imaging a welded portion such as a bead or a welding mark after welding. The imaging unit outputs image data obtained by imaging the welded portion to the control unit 36 as detection data.
Further, the sensor 35 may be an inspection device that detects the state of the welded portion by scanning the welded portion with a laser (detection light). The inspection device detects the state of the welded portion by laser scanning, and outputs detection data indicating the detection result to the control unit 36.

The microphone 37 is provided on the welding torch 3, collects the arc sound generated during welding, and outputs the arc sound data to the control unit 36.

The notification unit 38 is, for example, a speaker, and outputs a sound indicating the quality of the welding state. The audio output is controlled by the control unit 36.

The control unit 36 is a computer having a CPU (Central Processing Unit), a processor such as a multi-core CPU, a ROM (Read Only Memory), a RAM (Random Access Memory), an input / output interface, and the like. The control unit 36 executes a predetermined process such as setting welding conditions according to the operation of the operation unit 33. The control unit 36 executes a process of transmitting welding condition data indicating the set welding conditions to the welding power source 1 by the communication unit 31. Further, the control unit 36 executes a process of transmitting the detection data detected by the sensor 35 and the arc sound data collected by the microphone 37 to the welding power source 1 by the communication unit 31. The arc sound data and the detection data may be configured to be output directly from the sensor 35 and the microphone 37 to the welding power source 1.
Further, the control unit 36 can also control the display unit 32 to appropriately display the welding conditions, the information read from the storage unit 34, and the like. Further, the welding power supply 1 determines the quality of the welding state based on the arc sound data as described later, and the control unit 36 transmits the quality data indicating the quality of the welding state transmitted from the welding power supply 1 to the communication unit 31. It is controlled to notify the notification unit 38 of the quality of the welding state by receiving the data via the above.

<Welding power supply>
FIG. 3 is a block diagram showing a configuration example of the welding power supply 1 according to the first embodiment. The welding power supply 1 includes a main control unit 11, a power supply unit 12, a communication unit 13, a storage unit 14, and a learning unit 15.

The main control unit 11 is a microcomputer having a CPU, and controls the operation of each component that constitutes the welding power supply 1.

The power supply unit 12 is a circuit that supplies electric power for performing arc welding to the welding torch 3. The power supply unit 12 converts the three-phase AC power input from the power system P into power suitable for arc welding and outputs it. Further, the power supply unit 12 converts the three-phase AC power into electric power suitable for driving control of the wire feeding device 2 and the welding torch 3, and supplies the three-phase AC power to the wire feeding device 2 and the welding torch 3.

The communication unit 13 is a circuit that communicates with the wire feeding device 2 and the welding torch 3. For example, it receives arc sound data, detection data, welding condition data, etc. transmitted from the welding torch 3 and the wire feeding device 2, and outputs various received data to the main control unit 11, the learning unit 15, and the like. Further, as will be described later, the communication unit 13 transmits quality data indicating the quality of the welding state determined from the arc sound to the welding torch 3.

The storage unit 14 is a non-volatile memory such as an EEPROM or a flash memory. The storage unit 14 stores a computer program 14a for generating a learned model 154 that determines the quality of the welding state from the arc sound. In addition, the storage unit 14 stores learning data for generating the trained model 154. Further, the storage unit 14 stores the learned model 154 generated for each of a plurality of welding conditions.

The learning unit 15 is a computer having a CPU, a processor such as a GPU (Graphics Processing Unit) or a multi-core CPU, a ROM, a RAM, an input / output interface, and the like. The learning unit 15 executes a process of accumulating learning data in which arc sound data and detection data are associated with each welding condition in the storage unit 14.
When sufficient learning data is accumulated, the learning unit 15 performs a process of generating a learned model 154 for determining the quality of the welding state from the arc sound based on the learning data accumulated in the storage unit 14. Execute.
After the trained model 154 is generated, the learning unit 15 uses the trained model 154 to execute a process of determining the quality of the welded state. The main control unit 11 transmits the quality data indicating the quality of the welded state determined using the trained model 154 to the welding torch 3 via the communication unit 13, and notifies the welder of the quality of the welded state.

<Learning Department>
FIG. 4 is a block diagram showing a configuration example of the learning unit 15 according to the first embodiment, and FIG. 5 is a conceptual diagram showing an example of learning data. The learning unit 15 includes a determination device 151, an arc sound conversion unit 152, a learning processing unit 153, and a learned model 154.

The determination device 151 determines the quality of the welding state based on the detection data acquired from the sensor 35, and outputs the quality data indicating the quality of the welding state to the learning processing unit 153. When the welding condition is good, the welding current, welding voltage, welding wire feeding speed, and short-circuit condition during welding are normal, and an expert can check the quality of the welding condition from the appearance of the welded part after welding. I can judge.
The determination device 151 includes, for example, a learned model 151a for pass / fail determination. The trained model 151a for pass / fail determination is a trained deep neural network, and has an input layer, an intermediate layer, and an output layer. When the detection data is input to the input layer, the quality data indicating the quality of the welding state is output from the output layer. The configuration and type of the neural network are not particularly limited, and are deep neural networks such as a convolutional neural network, a recurrent neural network, and an LSTM (Long Short-Term Memory). , Or a deep neural network in which these are combined. When the detected data is image data, a convolutional neural network may be used, and when the detected data is time series data, a recursive neural network or the like may be used.
Since a relatively large amount of image data of the welded portion associated with the quality of the welded state is accumulated, it is relatively easy to prepare the training data for generating the trained model 151a for quality determination. is there.
The determination device 151 may acquire monitor data indicating the welding current or welding voltage during welding, and determine the quality of the welding state based on the monitor data. Further, although the example of generating the determination device 151 by supervised learning of the neural network has been described, the determination device 151 that has been machine-learned by other known methods such as logistic regression and support vector machine may be provided.

The arc sound conversion unit 152 converts the arc sound data acquired from the microphone 37 into digital data for generating a trained model. The arc sound conversion unit 152 converts, for example, into digital data of a waveform image showing the sound pressure of the arc sound. Needless to say, it may be converted into time-series digital data. In the welding sound learning mode, the arc sound conversion unit 152 outputs arc sound data to the learning processing unit 153. In the welding mode in which the quality of the welding state is determined using the trained model 154, the arc sound conversion unit 152 outputs the arc sound data to the trained model 154.

The learning processing unit 153 executes a learning data collection and storage process and a learning model 154 generation process. In the welding sound learning mode, particularly in the stage of collecting and accumulating learning data, the learning processing unit 153 collects the arc sound generated during welding with the microphone 37 each time welding is performed, and the arc sound data obtained. The quality data indicating the quality of the welding state and the welding condition data are acquired. Specifically, the learning processing unit 153 acquires the arc sound data output from the arc sound conversion unit 152. In addition, the learning processing unit 153 acquires quality data from the determination device 151. Further, the learning processing unit 153 acquires welding condition data from the main control unit 11. Then, as shown in FIG. 5, the learning data associated with the acquired welding condition data, arc sound data, and quality data is stored in the storage unit 14.
When sufficient learning data is accumulated in the welding sound learning mode, the learning processing unit 153 reads the accumulated learning data from the storage unit 14, and based on the read learning data, from the arc sound. A trained model 154 for determining the quality of the welded state is generated.
Further, in the welding mode in which the quality of the welding state is determined using the trained model 154, the learning processing unit 153 is a trained model corresponding to the welding condition from among the plurality of trained models 154 stored in the storage unit 14. A selection unit 153a for selecting 154 and reading from the storage unit 14 is provided.

When the welding sound learning mode is selected by the operation unit 33, the learning unit 15 executes the following processing. If the trained model 154 has not been generated yet, the welding power supply 1 may be configured to automatically select the welding sound learning mode.

<How to generate a trained model>
FIG. 6 is a flowchart showing a method of generating the trained model 154. The welding power supply device monitors the operation by the welder by the operation unit 33 and accepts the welding conditions (step S11). The welding condition data input to the welding torch 3 is transmitted from the welding torch 3 to the welding power source 1. The welding power source 1 may accept welding conditions.

Next, the learning unit 15 of the welding power source 1 acquires arc sound data from the microphone 37 via the arc sound conversion unit 152 (step S12). Further, the learning unit 15 acquires detection data from the sensor 35 (step S13) and determines whether the welding state is good or bad based on the detection data (step S14). Then, the learning unit 15 associates the welding condition data, the arc sound data, and the quality data, and stores the data as learning data in the storage unit 14 (step S15).

Next, the learning unit 15 determines whether or not a predetermined amount of learning data has been accumulated in the storage unit 14 (step S16). When it is determined that a predetermined amount of learning data is not accumulated (step S16: NO), the learning unit 15 returns the process to step S11.

When it is determined that a predetermined amount of learning data is accumulated (step S16: YES), the learning unit 15 reads out the learning data accumulated in the storage unit 14, and based on the read out learning data. When the arc sound data is input, a trained model 154 that outputs the good / bad data corresponding to the arc sound data is generated by machine learning (step S17). For example, the weighting coefficient of the intermediate layer can be obtained by minimizing the error function by the gradient descent method.
In step S17, the learning unit 15 generates a different learned model 154 for each of a plurality of welding conditions. For example, the learning unit 15 uses the learning data of the welding condition α to generate the learning data for the welding condition α, and uses the learning data of the welding condition β to generate the learning data for the welding condition β. Generate. The generated trained model 154 is stored in the storage unit 14. The trained model 154 may be generated by determining whether or not a sufficient trained model 154 has been accumulated for each welding condition, and generating the trained model 154 from the data for training.

The trained model 154 includes a neural network 154a as shown in FIG. The neural network 154a includes an input layer into which arc sound data is input, an intermediate layer in which operations are performed on the arc sound data input to the input layer based on learned weight coefficients, and quality data indicating the quality of the welded state. It has an output layer that outputs.
Arc sound data representing the arc sound is input to the input layer. The arc sound data may be image data representing the sound pressure waveform of the arc sound, or may be time series data. The image information is input to the input layer via a convolution layer and a convolution layer (not shown).
The output layer has one or more neurons. The output layer includes, for example, a neuron showing the probability that the welded state is good and a neuron showing the probability that the welded state is bad.

The configuration and type of the neural network 154a are not particularly limited, and may be a convolutional neural network, a recursive neural network, a deep neural network such as LSTM, or a deep neural network in which these are combined.
In addition, an example of creating a trained model 154 by supervised learning of a neural network has been described. However, a logistic regression and a support vector machine can be used as a pass / fail judge for determining the quality of a welded state based on arc sound data and quality data. It may be generated by machine learning by other known methods such as.

When the welding mode is selected by the operation unit 33, the learning unit 15 executes the following processing. When the storage unit 14 stores the learned model 154, the welding power supply 1 may be configured to automatically select the welding mode.

<Method of determining the quality of the welded state using the trained model 154>
FIG. 7 is a flowchart showing a method of determining whether the welding state is good or bad and a notification method. The welding power supply device monitors the operation by the welder by the operation unit 33 and accepts the welding conditions (step S31). The welding condition data input to the welding torch 3 is transmitted from the welding torch 3 to the welding power source 1.

Next, the learning unit 15 of the welding power source 1 selects one learned model 154 corresponding to the welding condition from the plurality of learned models 154 (step S32). Then, the learning unit 15 acquires the arc sound data (step S33) and determines the quality of the welding state using the learned model 154 selected in step S32 (step S34). Specifically, the learning unit 15 inputs the arc sound data to the input layer of the trained model 154, so that the quality data is output from the output layer.

Next, the main control unit 11 transmits the quality data output from the learned model 154 to the welding torch 3 via the communication unit 13, and the notification unit 38 notifies the welder of the quality of the welding state (step S35). ).

According to the first embodiment, it is possible to efficiently generate and accumulate learning data and generate a learned model 154 that accurately determines the quality of the welding state from the arc sound.

Further, the trained model 154 can be efficiently generated by machine learning for each welding condition.

Further, the welding power supply device can determine the quality of the welding state from the arc sound using the trained model 154 and notify the welder. Specifically, the welding power supply device acquires arc sound data during welding, causes the acquired arc sound data to be input to the trained model 154, and indicates the welding state indicated by the quality data output from the trained model 154. The notification unit 38 can notify the quality.

Furthermore, the welding power supply device selects the trained model 154 corresponding to the welding conditions, and uses the selected trained model 154 to determine and notify the quality of the welding state. By using the trained model 154 suitable for the welding conditions, the welding state can be determined and notified more accurately.

Furthermore, by using the learned model 151a for quality determination, which is a trained neural network, the quality of the welding state can be accurately determined from the detection data.

(Embodiment 2)
FIG. 8 is a schematic diagram showing a learning system according to the second embodiment. The welding power supply device A according to the second embodiment is different from the first embodiment in that the trained model 154 is generated by the server. Therefore, the above differences will be mainly described below. Since other configurations and actions and effects are the same as those in the embodiment, the corresponding parts are designated by the same reference numerals and detailed description thereof will be omitted.

The learning system according to the second embodiment includes a learning processing server 9 and a plurality of welding power supply devices A connected to the learning processing server 9 by wire or wirelessly.

The welding power supply 1 of each welding power supply device A transmits learning data to the learning processing server 9. The learning processing server 9 accumulates learning data transmitted from each welding power supply device A, and generates a learned model 154 for each welding condition in the same manner as in the first embodiment. Then, the learning processing server 9 distributes the generated learned model 154 to each welding power supply device A. The welding power supply device A receives and stores the trained model 154 delivered from the learning processing server 9. The method of notifying the quality of the welded state using the trained model 154 is the same as that of the first embodiment.

According to the second embodiment, the learning data can be collected and accumulated more efficiently by receiving the arc sound data from the plurality of welding power supply devices A in the learning processing server 9 of the cloud. Then, by performing machine learning with a large amount of learning data, it is possible to generate a learned model 154 that can more accurately determine the quality of the welding state from the welding sound.

Further, by distributing the learned model 154 from the learning processing server 9, the learned model 154 can be shared by a plurality of welding power supply devices A.

Although the semi-automatic welding power supply device has been mainly described in the present embodiment, the present invention can also be applied to a welding system using a welding robot.

Further, although an example in which the process of determining the quality of the welding state based on the detection data is performed by the welding power supply 1, the welding torch 3 may be configured to determine the quality of the welding state.

Further, although the example in which the memory of the trained model 154 and the quality determination using the trained model 154 are performed on the welding power supply 1 side has been described, the processing may be executed on the welding torch 3 side.

Further, in the present embodiment, an example of selecting one trained model 154 corresponding to the selected welding condition has been described, but when there is no trained model 154 completely corresponding to the selected welding condition, the selected welding is performed. It may be configured to select a plurality of trained models 154 that are close to the conditions. The learning unit 15 may input arc sound data into the plurality of trained models 154 and determine the quality of the welding state based on the quality data output from the plurality of trained models 154.

Further, in the present embodiment, the trained model 154 has described an example of outputting quality data indicating the quality of the welded state, but the quality data may be subdivided and output. For example, the welding power supply device may be provided with a learned model 154 machine-learned so as to classify and output data indicating that the welding state is defective according to the cause of the defect.
For example, the trained model 154 has a node that outputs data indicating the probability of being in a defective state due to the first defect cause and a node that outputs data indicating the probability of being in a defective state due to the second defect cause. It is preferable that the output layer is provided with a node indicating the probability of being in a state. Needless to say, as the learning data, the bad state due to the first bad cause, the bad state due to the second bad cause, and the good / bad data indicating the good state are included as the teacher data, and the neural network is machine-learned. Just do it.
Further, when the welding condition is defective, the welding power supply device may be configured to output advice according to the cause of the defect to the notification unit 38. The advice is how to correct the welding method, the cause of the defect, and other information necessary to improve the welding condition.

The embodiments disclosed this time should be considered as exemplary in all respects and not restrictive. The scope of the present invention is indicated by the scope of claims, not the above-mentioned meaning, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims.

1 welding power supply, 2 wire feeder, 3 welding torch, 11 main control unit, 14a computer program, 15 learning unit, 151 judgment device, 151a learned model for quality judgment, 152 arc sound conversion unit, 153 learning processing unit, 153a selection unit, 154 trained model, 154a neural network, 34 storage unit, 35 sensor, 36 control unit, 37 microphone, 38 notification unit

Claims (7)

It is a method of generating a trained model that outputs quality data indicating the quality of the welded state when arc sound data is input.
Acquire the arc sound data obtained by collecting the arc sound generated during welding with a microphone and
By imaging the welded part after welding or scanning the welded part with detection light, the state of the welded part is detected and the detection data obtained is acquired.
Based on the acquired detection data, the quality of the welding condition is determined.
The acquired learning data associated with the arc sound data and the quality data indicating the quality of the welding state is stored in the storage unit.
Based on the learning data stored in the storage unit, when the arc sound data is input, learning to generate a learned model that outputs the good / bad data corresponding to the arc sound data by machine learning. How to generate a completed model. The first aspect of claim 1, wherein the quality of the welding state is determined by inputting the acquired detection data into a learning model for quality determination that outputs data indicating the quality of the welding state when the detection data is input. How to generate a trained model of. The learning data is stored in the storage unit for each of a plurality of different welding conditions.
The trained according to claim 1 or 2, wherein a plurality of trained models different for each of the plurality of welding conditions are generated based on the learning data of the plurality of welding conditions stored in the storage unit. How to generate a model. It is a computer program for generating a trained model that outputs quality data indicating the quality of the welded state when arc sound data is input to the computer.
On the computer
Acquire the arc sound data obtained by collecting the arc sound generated during welding with a microphone and
By imaging the welded part after welding or scanning the welded part with detection light, the state of the welded part is detected and the detection data obtained is acquired.
Based on the acquired detection data, the quality of the welding condition is determined.
The acquired learning data in which the arc sound data and the quality data indicating the quality of the welding state are associated with each other are stored in the storage unit.
A process of generating a trained model that outputs the quality data corresponding to the arc sound data by machine learning when the arc sound data is input based on the learning data stored in the storage unit. A computer program to run. With a microphone that collects arc sounds,
A trained model that outputs quality data indicating the quality of the welded state when arc sound data is input, and a trained model that outputs quality data.
A welding power supply device including a notification unit that inputs arc sound data collected by the microphone to the trained model and notifies the quality of the welding state indicated by the quality data output from the trained model. There are a plurality of the trained models, and each trained model is machine-learned for each of a plurality of different welding conditions.
A selection unit for selecting the trained model corresponding to the welding condition based on the welding condition is provided.
The welding power supply device according to claim 5, wherein the notification unit notifies the quality of the welding state indicated by the quality data output from the learned model selected by the selection unit. It is a trained model that makes the computer function so that when the arc sound data is input, the quality data indicating the quality of the welded state is output.
An input layer into which arc sound data obtained by collecting the arc sound generated during welding with a microphone is input,
An intermediate layer that performs an operation based on a learned weighting coefficient for the arc sound data input to the input layer, and an intermediate layer.
It has an output layer that outputs the quality data indicating the quality of the welding state when the sound related to the arc sound data is emitted.
The state of the welded portion is detected by capturing the arc sound data obtained by collecting the arc sound generated during welding with a microphone and imaging the welded portion after welding or scanning the welded portion with detection light. A neural network that trains the weighting coefficient so that the arc sound data input to the input layer and the quality data output from the output layer correspond to each other based on the quality data obtained. Equipped with a network
On the computer
A trained model for performing an operation based on the arc sound data and the weighting coefficient input to the input layer and outputting the quality data.

Images