Learned model generation method, welding system, welding assistance method, and program

Abstract

To provide a learned model generation method capable of improving welding quality determination accuracy.SOLUTION: This learned model generation method comprises: detecting an event accompanying welding; acquiring learning data containing a plurality of sub-data; acquiring a label indicating whether the welding is normal regarding the learning data; inputting each of the sub-data of the learning data to which labels indicating abnormality are added to the model for calculating an abnormality score; calculating a plurality of abnormality scores; and performing model learning so as to reduce a difference between the label indicating the abnormality and an abnormality score equal to or higher than a prescribed score among the abnormality scores.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a method for generating a learned model, a welding system, a welding support method, and a program.

Background Art

[0002] Patent Document 1 discloses a technique in which a camera captures images of a welding pool and images of the ripple shape and the geometric shape of the fillet portion, a processor receives the images, and communicates with a database that stores potential defects in the associated welding together with images of the molten pool of a mock weld portion and images of the ripple shape and the geometric shape of the fillet portion of the mock weld portion, and calculates an aggregated probability that a defect is included at a welding position corresponding to the image captured by the camera based on the potential defects associated in the database.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] By the way, when process data obtained by detecting events associated with welding by a sensor is used as learning data, since the period for assigning normal / anomaly labels is longer than the period during which an anomaly occurs in welding, inappropriate labels may be locally assigned. Performing supervised machine learning using such inappropriate labels may reduce the accuracy of welding quality determination.

[0005] The present invention has been made in view of the above problems, and its main object is to provide a method for generating a learned model, a welding system, a welding support method, and a program capable of improving the accuracy of welding quality determination.

Means for Solving the Problems

[0006] In order to solve the above problems, a method for generating a learned model according to one aspect of the present invention acquires learning data including a plurality of sub-data obtained by detecting events associated with welding by a sensor, and for the learning data, acquires a label indicating whether the welding is normal or abnormal, inputs a plurality of the sub-data of the learning data to which the label indicating an abnormality is assigned into a model for calculating an abnormality score respectively, calculates a plurality of the abnormality scores, and performs learning of the model so as to reduce an error between the label indicating an abnormality and a predetermined or more abnormality scores among the plurality of abnormality scores. According to this, it becomes possible to improve the welding quality determination accuracy.

[0007] In the above aspect, the plurality of sub-data may be data respectively detected in a plurality of sub-periods included in a period corresponding to the learning data. According to this, it becomes possible to perform learning of the model using the sub-data detected in a sub-period with a high probability of occurrence of an abnormality.

[0008] In the above aspect, the learning data includes a plurality of types of detection data each including a plurality of the sub-data obtained by detecting events associated with welding by a plurality of types of the sensors, and the calculation of the abnormality score may input a set of the sub-data corresponding to each other among the plurality of types of the detection data into the model and calculate the abnormality score. According to this, it becomes possible to further improve the welding quality determination accuracy using a plurality of types of sensors.

[0009] In the above aspect, the set of the sub-data may be a set of data detected in a common sub-period among a plurality of sub-periods included in a period corresponding to the learning data. According to this, it becomes possible to improve the welding quality determination accuracy using the set of the sub-data detected in the common sub-period.

[0010] In the above aspect, the sub-data may be image data obtained by imaging at least one of the arc of the welding and the welded portion. According to this, it is possible to improve the welding quality determination accuracy by using the image data.

[0011] In the above aspect, the sub-data may be spectral data obtained by measuring the arc light of the welding. According to this, it is possible to improve the welding quality determination accuracy by using the spectral data obtained by measuring the arc light of the welding.

[0012] In the above aspect, the sub-data may be waveform data obtained by measuring at least one of the voltage and current of the welding. According to this, it is possible to improve the welding quality determination accuracy by using the waveform data of the voltage or current.

[0013] In the above aspect, when a defect is found in the welded portion by the inspection after welding, the label indicating an abnormality may be assigned to the learning data corresponding to the position of the defect. According to this, it is possible to assign a label indicating an abnormality to the learning data corresponding to the position of the defect.

[0014] In the above aspect, the model includes a neural network, and the learning of the model may adjust the weights of the neural network by the error backpropagation method. According to this, it is possible to improve the welding quality determination accuracy by using the neural network.

[0015] In the above aspect, the learning of the model may be performed so as to reduce the error between the label indicating an abnormality and the highest abnormality score among the plurality of abnormality scores. According to this, it is possible to perform the learning of the model by using the sub-data with the highest probability of occurrence of an abnormality.

[0016] Further, a welding system according to another aspect of the present invention includes a welding device, a sensor that detects an event associated with welding performed by the welding device, and an estimation unit that calculates an abnormality score of the welding performed by the welding device from detection data generated by the sensor using a learned model. The learned model acquires learning data including a plurality of sub-data in which events associated with welding are detected by a sensor, acquires a label indicating whether the welding is normal or abnormal for the learning data, and inputs the plurality of sub-data of the learning data to which the label indicating abnormality is assigned into a model for calculating an abnormality score, respectively, to calculate a plurality of the abnormality scores, and performs learning of the model so as to reduce an error between the label indicating abnormality and a predetermined or more of the plurality of abnormality scores. The welding system is generated in this way. According to this, it is possible to improve the accuracy of welding quality determination.

[0017] Further, a welding support method according to another aspect of the present invention is a welding support method that detects an event associated with welding performed by a welding device by a sensor and calculates an abnormality score of the welding performed by the welding device from detection data generated by the sensor using a learned model. The learned model acquires learning data including a plurality of sub-data in which events associated with welding are detected by a sensor, acquires a label indicating whether the welding is normal or abnormal for the learning data, and inputs the plurality of sub-data of the learning data to which the label indicating abnormality is assigned into a model for calculating an abnormality score, respectively, to calculate a plurality of the abnormality scores, and performs learning of the model so as to reduce an error between the label indicating abnormality and a predetermined or more of the plurality of abnormality scores. The welding support method is generated in this way. According to this, it is possible to improve the accuracy of welding quality determination.

[0018] Moreover, a program according to another aspect of the present invention causes a computer to acquire detection data generated by a sensor that detects events associated with welding performed by a welding apparatus, and to calculate an abnormality score of the welding performed by the welding apparatus from the detection data using a learned model. The learned model acquires learning data including a plurality of sub-data obtained by detecting events associated with welding by a sensor, acquires a label indicating whether the welding is normal or abnormal for the learning data, inputs the plurality of sub-data of the learning data to which the label indicating abnormality is assigned into a model for calculating an abnormality score respectively, calculates a plurality of the abnormality scores, and performs learning of the model so as to reduce an error between the label indicating abnormality and a predetermined or more abnormality scores among the plurality of abnormality scores. According to this, it is possible to improve the accuracy of welding quality determination.

Advantages of the Invention

[0019] According to the present invention, it is possible to improve the accuracy of welding quality determination.

Brief Description of the Drawings

[0020]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Embodiments for Carrying Out the Invention

[0021] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In this specification and each drawing, elements that are the same as those described above with respect to the already presented drawings may be denoted by the same reference numerals, and detailed descriptions may be omitted as appropriate.

[0022] FIG. 1 is a diagram showing a configuration example of a learning system 100. The learning system 100 includes a learning device 1, a plurality of types of sensors 2, and a storage device 5. The sensor 2 detects events associated with welding by a welding device 3. Hereinafter, the detection data detected by the sensor 2 is also referred to as "welding process data".

[0023] The learning device 1 is a computer including a CPU, a GPU, a RAM, a ROM, a non-volatile memory, an input / output interface, and the like. The CPU of the learning device 1 executes information processing according to a program loaded from the ROM or the non-volatile memory into the RAM.

[0024] The program may be supplied via an information storage medium such as an optical disk or a memory card, or may be supplied via a communication network such as the Internet or a LAN.

[0025] The learning device 1 includes a data acquisition unit 11, a label acquisition unit 12, and a learning unit 13. These functional units are realized by the CPU of the learning device 1 executing information processing according to a program.

[0026] The data acquisition unit 11 acquires the welding process data from the sensor 2 as learning data D. The label acquisition unit 12 acquires a label L (normal label / anomaly label) indicating whether the welding is normal or abnormal for the learning data D.

[0027] The learning unit 13 performs machine learning of a model M for calculating an anomaly score from the welding process data using the learning data D and the label L. The model M is a model including, for example, a neural network.

[0028] The welding device 3 is a welding robot including a welding torch 31, an articulated arm 32 that supports the welding torch 31, and a control device 33 that controls the operations of the welding torch 31 and the arm 32.

[0029] As shown in FIG. 2, for example, the welding device 3 performs arc welding while weaving the welding torch 31 in the welding progress direction at the groove G formed between two workpieces U and L to be welded. The arc welding is, for example, TIG welding. However, it is not limited to this, and it may be MIG welding or MAG welding, etc.

[0030] The welding process data detected by a plurality of types of sensors 2 (see FIG. 1) includes a plurality of types of detection data. Specifically, one of the sensors 2 is, for example, an image sensor, and outputs image data obtained by imaging the arc during welding. The image sensor outputs time-series image data generated sequentially.

[0031] Another one of the sensors 2 is, for example, a spectroscopic sensor, and outputs spectroscopic data obtained by spectroscopically measuring the arc light during welding. The spectroscopic sensor outputs time-series spectroscopic data generated sequentially.

[0032] Still another one of the sensors 2 is, for example, a voltage sensor, and outputs waveform data obtained by measuring the voltage during welding. The voltage sensor is provided in the power supply of the welding device 3 and measures the voltage during arc welding.

[0033] Not limited to this, image data obtained by imaging a molten pool during arc welding using an image sensor may be used, or waveform data obtained by measuring the current during arc welding using a current sensor may be used.

[0034] The sensor 2 is not particularly limited to the above example as long as it can detect an event associated with arc welding. Note that only one type of sensor 2 may be used.

[0035] FIG. 3 is a flowchart showing an example of a procedure for generating a learned model realized in the learning system 100. The learning device 1 executes the information processing shown in the figure according to a program.

[0036] First, the learning device 1 acquires welding process data (for example, image data, spectroscopic data, voltage data) as learning data (processing as the data acquisition unit 11 in S11), and assigns a label to the acquired learning data (processing as the label acquisition unit 12 in S12).

[0037] In this embodiment, porosity defects caused by gusts during welding are targeted for welding quality determination. In the case of gas shielded arc welding, the welding part is protected by the shielding gas ejected from the nozzle, but when the wind disturbance exceeds a certain intensity, the shield is disrupted and the atmosphere enters. At that time, the welding process instantaneously becomes abnormal, which becomes a factor for the occurrence of porosity defects.

[0038] Since it is difficult to identify the occurrence of an abnormality due to wind disturbance from the welding process data, the assignment of an abnormal label is performed based on the defect inspection result after welding. That is, when a defect is found in the welded part by inspection after welding, an abnormal label is assigned to the learning data corresponding to the position of the defect.

[0039] Specifically, the defect occurrence part is identified from the appearance of the weld bead, the X-ray inspection image, etc., the defect occurrence time is calculated from the defect occurrence position using the welding speed of the robot in the welding direction, and an abnormal label is assigned to the learning data acquired at the defect occurrence time.

[0040] FIG. 4 is a diagram showing an example of the timing of occurrence of an abnormality in an actual process, and FIGS. 5 and 6 are diagrams showing examples of label assignment. As shown in these figures, the period during which an abnormality occurs in the actual process is shorter than the unit period to which a label is assigned. This is because the unit period to which a label is assigned is set according to the resolution of the defect inspection after welding.

[0041] Among the welding process data, the data acquired for each unit period to which a label is assigned becomes "learning data". That is, the unit period to which a label is assigned is the period corresponding to the learning data.

[0042] In the examples of FIGS. 4 to 6, although the actual process instantaneously enters an abnormal state at 5.0 seconds, 5.5 seconds, and 6.0 seconds respectively, the abnormal label is assigned over the period from 4.5 to 6.5 seconds. In this example, the label is assigned every 0.5 seconds. That is, the period of the learning data is 0.5 seconds.

[0043] FIG. 7 is a diagram showing an example of a sub-period included in the period of learning data. The period of learning data includes a plurality of sub-periods. Among the learning data, the data acquired in the sub-period is "sub-data". In a plurality of sub-periods included in the period of learning data, a plurality of sub-data are acquired respectively.

[0044] The sub-data becomes the data input to the model at one time. That is, among the period of the learning data, the period during which the data input to the model at one time is acquired is the "sub-period".

[0045] The learning data has a hierarchical structure including a plurality of sub-data. The learning data is upper-layer data (also called "Bag"), and the sub-data is lower-layer data. The learning data can also be said to be a set of sub-data.

[0046] Regarding the training data with abnormal labels, when looking at the state of the actual process in each sub-period, the normal state and the abnormal state are mixed. That is, in some sub-periods, the actual process is in the abnormal state, which is consistent with the abnormal label. However, in other sub-periods, the actual process is in the normal state, which deviates from the abnormal label.

[0047] In the example of Figure 7, in the sub-periods from 5.0 to 5.1 seconds and from 5.4 to 5.5 seconds, the actual process is in the abnormal state, which is consistent with the abnormal label. In contrast, in the sub-periods from 5.1 to 5.2 seconds, from 5.2 to 5.3 seconds, and from 5.3 to 5.4 seconds, the actual process is in the normal state, which deviates from the abnormal label. In this example, the sub-period is 0.1 second.

[0048] When performing machine learning using such sub-data with labels that deviate from the state of the actual process, there is a risk that the welding quality determination accuracy will decrease. Therefore, in this embodiment, the multiple instance learning described below is used to improve the welding quality determination accuracy.

[0049] Returning to the description of Figure 3. The learning device 1 acquires training data and assigns labels (S11, S12), and then performs model learning by multiple instance learning (S13, the process as the learning unit 13).

[0050] Specifically, the learning device 1 inputs a plurality of sub-data of the training data with abnormal labels into the model respectively to calculate a plurality of abnormal scores. Then, the learning device 1 performs model learning so as to reduce the error between the abnormal label and the highest abnormal score among the plurality of abnormal scores.

[0051] Figure 8 is a diagram showing an example of calculating the abnormal score. The learning device 1 inputs a plurality of sub-data obtained in a plurality of sub-periods (0.1 second) included in the period (0.5 second) of the training data with abnormal labels into the model respectively to calculate a plurality of abnormal scores.

[0052] At this time, for the sub-data obtained during the sub-period when the actual process is in an abnormal state, a relatively high anomaly score is calculated, and for the sub-data obtained during the sub-period when the actual process is in a normal state, a relatively low anomaly score is calculated.

[0053] The learning device 1 selects the highest anomaly score from the plurality of anomaly scores and calculates the error between the anomaly label and the selected anomaly score. In the example of FIG. 8, the error between the value 1 of the anomaly label and the value 0.8 of the highest anomaly score is calculated.

[0054] The error is, for example, the cross-entropy error. In the example of FIG. 8, the cross-entropy error is -(1×log0.8)=0.22. Note that, for example, the value 0.7 of the second-highest anomaly score may be used for calculating the error.

[0055] Then, the learning device 1 performs learning of the model so as to reduce the error between the anomaly label and the highest anomaly score. Specifically, the learning device 1 performs learning of the model by adjusting the weights of the neural network by the error backpropagation method.

[0056] FIG. 9 is a diagram showing an example of a set of sub-data. Since the welding process data includes a plurality of types of detection data (image data, spectroscopic data, voltage data), a set of sub-data of a plurality of types of detection data is input to the model at once.

[0057] That is, the learning device 1 inputs a set of corresponding sub-data among the plurality of types of detection data to the model and calculates an anomaly score. The set of sub-data is a set of sub-data detected in a common sub-period among a plurality of sub-periods included in the learning data period.

[0058] For example, as shown in FIG. 9, the image data G 21 ~G 2m 、the spectroscopic data D 21 ~D 2n 、and the voltage data V 21 ~V2p is input into the model at once as one of the sets of sub-data. Sets of sub-data are created in the same way for other sub-periods.

[0059] FIG. 10 is a diagram showing a configuration example of the model M. For example, the model M includes a plurality of neural networks m1 to m3, a multiple instance learning layer mm, and an output layer mp.

[0060] The neural network m1 is a 3D convolutional neural network, and the neural networks m2 and m3 are 1D convolutional-LSTM (Long Short-Term Memory) neural networks.

[0061] Image data is input into the neural network m1, spectral data is input into the neural network m2, and voltage data is input into the neural network m3. In addition to this, image data of the molten pool and current data may be input.

[0062] The outputs of the plurality of neural networks m1 to m3 are integrated and input into the multiple instance learning layer mm. The output of the multiple instance learning layer mm is input into the output layer mp.

[0063] The output layer mp is provided with an element that outputs normal / abnormal of the welding. The said element is composed of, for example, a sigmoid function or a softmax function, and outputs an abnormality score of 0 or more and 1 or less. The closer the abnormality score is to 0, the more it represents normal, and the closer it is to 1, the more it represents abnormal.

[0064] For example, in the training data, there are 250 image data per 0.5 seconds (500 Hz), 250 spectral data points per 0.5 seconds (500 Hz), and 10,000 voltage data points per 0.5 seconds (20 kHz).

[0065] In addition, in the set of sub-data, the image data is 50 sheets (500 Hz) per 0.1 second, the spectroscopic data is 50 points (500 Hz) per 0.1 second, and the voltage data is 2,000 points (20 kHz) per 0.1 second.

[0066] According to the embodiment described above, by performing learning of the model using the sub-data detected in the sub-period with a high probability of occurrence of an abnormality, it is possible to generate a learned model with high welding quality determination accuracy.

[0067] FIG. 11 is a diagram showing a configuration example of the welding system 200. The welding system 200 includes a welding device 3, a welding support device 6, a plurality of types of sensors 2, and a storage device 5. The learned model LM generated by the above-described learning device 1 is stored in the storage device 5.

[0068] The welding support device 6 is the same computer as the above-described learning device 1 and executes information processing according to a program. The welding support device 6 includes a data acquisition unit 61, an estimation unit 62, and a notification unit 63.

[0069] The data acquisition unit 61 acquires welding process data from the plurality of sensors 2. The estimation unit 62 calculates an abnormality score of the welding performed by the welding device 3 from the welding process data using the learned model LM.

[0070] When it is determined that the welding performed by the welding device 3 is abnormal, the notification unit 63 notifies the welding device 3 of the occurrence of the abnormality. When receiving the abnormality notification, the welding device 3 performs a predetermined operation such as stopping the welding.

[0071] FIG. 12 is a flowchart showing an example of a procedure of a welding support method realized in the welding system 200. The welding support device 6 executes the information processing shown in the figure according to a program.

[0072] First, the welding support device 6 acquires welding process data (for example, image data, spectroscopic data, voltage data) from the plurality of sensors 2 (S21, processing as the data acquisition unit 61).

[0073] Next, the welding support device 6 inputs the acquired welding process data into the learned model to calculate an abnormality score (processing as the estimation unit 62 in S22). Specifically, the welding support device 6 inputs the welding process data into the learned model for each sub-data and sequentially calculates the abnormality score.

[0074] Next, the welding support device 6 determines the welding quality based on the calculated abnormality score (processing as the estimation unit 62 in S23). Specifically, the welding support device 6 determines that the welding is abnormal when the sequentially calculated abnormality score exceeds the threshold value.

[0075] According to the present embodiment, by calculating the abnormality score from the welding process data using the learned model generated as described above, it is possible to improve the accuracy of welding quality determination.

[0076] FIG. 13 is a diagram for explaining the determination result of welding quality. The upper part of the figure is an image showing the appearance of the weld bead. Blow holes (porosity defects) are generated at the positions of the broken lines. The positions of the blow holes can be specified by palpation with a finger or X-ray inspection.

[0077] The lower part of the figure is a graph showing the abnormality score according to the welding time, and the welding time is displayed corresponding to the position of the weld bead. In this example, sampling is performed at intervals of 0.25 seconds, and one abnormality score is output every 0.25 seconds.

[0078] According to this figure, it can be seen that the abnormality score has increased at the welding time corresponding to the position of the blow hole, and the abnormality of the welding can be determined.

[0079] Although the embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above, and it goes without saying that various modifications are possible for those skilled in the art.

[0080] In the above embodiment, the training data (Bag) and the sub-data were configured in a temporal relationship, but the present invention is not limited thereto, and they may be configured in a positional relationship or a frequency relationship, for example.

[0081] For example, by using data converted into a spectrogram (an image three-dimensionally displayed in terms of time, frequency, and intensity) as one Bag and using the sub-data as divided images of the spectrogram, the Bag and the sub-data can be configured in a positional relationship.

Description of Reference Numerals

[0082] 1 Learning device, 2 Sensor, 3 Welding device, 5 Storage device, 6 Welding support device, 11 Data acquisition unit, 12 Label acquisition unit, 13 Learning unit, 31 Welding torch, 32 Arm, 33 Control device, 61 Data acquisition unit, 62 Estimation unit, 63 Notification unit, 100 Learning system, 200 Welding system, D Training data, L Label, M Model, LM Trained model

Claims

1. Obtain learning data including a plurality of sub-data detected by a sensor for events associated with welding, For the learning data, obtain a label indicating whether the welding is normal or abnormal, Input each of the plurality of sub-data of the learning data with the label indicating an abnormality into a model for calculating an abnormality score to calculate a plurality of the abnormality scores, Perform learning of the model so as to reduce an error between the label indicating an abnormality and a predetermined number or more of the abnormality scores among the plurality of abnormality scores, A method for generating a learned model.

2. The plurality of sub-data are data respectively detected in a plurality of sub-periods included in a period corresponding to the learning data, The method for generating a learned model according to Claim 1.

3. The learning data includes a plurality of types of detection data each including a plurality of sub-data detected by a plurality of types of the sensors for events associated with the welding, The calculation of the abnormality score is performed by inputting a set of the sub-data corresponding to each other among the plurality of types of detection data into the model to calculate the abnormality score, The method for generating a learned model according to Claim 1.

4. The set of sub-data is a set of data detected in a common sub-period among a plurality of sub-periods included in a period corresponding to the learning data, The method for generating a learned model according to Claim 3.

5. The sub-data is image data obtained by imaging at least one of the arc of the welding and the welded portion, The method for generating a learned model according to Claim 1.

6. The sub-data is spectroscopic data obtained by measuring the arc light of the welding, The method for generating a learned model according to Claim 1.

7. The sub-data is waveform data obtained by measuring at least one of the voltage and current of the welding, The method for generating a learned model according to Claim 1.

8. The label indicating an abnormality is assigned to the learning data corresponding to the position of the defect when a defect is found in the welded portion by inspection after the welding, The method for generating a learned model according to Claim 1.

9. The model includes a neural network, The learning of the model adjusts the weights of the neural network by the error backpropagation method, The method for generating a learned model according to Claim 1.

10. The learning of the model is performed so as to reduce the error between the label indicating an abnormality and the highest abnormality score among the plurality of the abnormality scores. The method for generating a learned model according to claim 1.

11. A welding apparatus, A sensor that detects an event associated with welding performed by the welding apparatus, An estimation unit that calculates an abnormality score of welding performed by the welding apparatus from detection data generated by the sensor using a learned model, Comprising: The learned model Obtains learning data including a plurality of sub-data in which events associated with welding are detected by a sensor, For the learning data, obtains a label indicating whether the welding is normal or abnormal, Inputs each of the plurality of sub-data of the learning data to which the label indicating an abnormality is assigned into a model for calculating an abnormality score, and calculates a plurality of the abnormality scores, Performs learning of the model so as to reduce the error between the label indicating an abnormality and a predetermined or more abnormality scores among the plurality of the abnormality scores, Is generated by Welding system.

12. Detects an event associated with welding performed by a welding apparatus by a sensor, Calculates an abnormality score of welding performed by the welding apparatus from detection data generated by the sensor using a learned model, A welding support method, comprising: The learned model Obtains learning data including a plurality of sub-data in which events associated with welding are detected by a sensor, For the learning data, obtains a label indicating whether the welding is normal or abnormal, Inputs each of the plurality of sub-data of the learning data to which the label indicating an abnormality is assigned into a model for calculating an abnormality score, and calculates a plurality of the abnormality scores, Performs learning of the model so as to reduce the error between the label indicating an abnormality and a predetermined or more abnormality scores among the plurality of the abnormality scores, Is generated by Welding support method.

13. Obtaining detection data generated by a sensor that detects an event associated with welding performed by a welding apparatus, and Calculating an abnormality score of welding performed by the welding apparatus from the detection data using a learned model, A program for causing a computer to execute, The learned model Obtains learning data including a plurality of sub-data in which events associated with welding are detected by a sensor, For the learning data, obtain a label indicating whether the welding is normal or abnormal, Input each of a plurality of the sub-data of the learning data to which the label indicating an abnormality is assigned into a model for calculating an abnormality score, and calculate a plurality of the abnormality scores, Perform learning of the model so as to reduce an error between the label indicating an abnormality and a predetermined number or more of the abnormality scores among the plurality of the abnormality scores, which is generated by a program.

Images