Welding defect prediction system, machine learning device, defect prediction method, and program

Abstract

To provide a defect predicting system, a machine learning device, a defect predicting method, and a program for further accurately performing defect prediction regarding welding.SOLUTION: A defect predicting system 100 for predicting a defect which occurs in welding includes: first prediction means 205 which predicts a defect by using a first learned model which uses a welding parameter as an input and outputs a parameter indicating a size of the defect; second prediction means 206 which predicts the defect by using a second learned model which uses the welding parameter as an input and outputs a parameter indicating the size of the defect; and third prediction means 207 which predicts the size of the defect on the basis of the parameter predicted by the first prediction means 205 and the parameter predicted by the second prediction means 206.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a defect prediction system for welding, a machine learning device, a defect prediction method, and a program. [Background technology]

[0002] Conventionally, welding systems judge the quality of welding during welding, and the judgment results are used for welding quality evaluation for quality control purposes, warning control, and control aimed at improving welding work. Naturally, the more accurate this judgment, the better. In recent years, technology has been developed that applies learning devices to predict welding results in real time during welding.

[0003] An example of a conventional technique for prediction using a learning device is Patent Document 1. Patent Document 1 discloses a configuration for targeting a welded structure, using an estimation model after machine learning with an image as input to estimate the probability of the presence or absence of a defect, and outputting an image of the structure according to the estimated value. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-060462 Summary of the Invention [Problem to be solved by the invention]

[0005] For example, in the case of welding wire additive manufacturing (AM), in which weld beads formed by melting and solidifying a filler metal are stacked to form a multilayered object, defects such as cavities may occur inside the object after fabrication, depending on the shape and arrangement of the weld beads. Furthermore, if the welding conditions or fabrication are inappropriate, narrow spaces may occur between the beads. Continuing fabrication with narrow spaces may result in incomplete fusion. From the perspective of quality control, it is important to detect or predict incomplete fusion and welding defects. However, because there are many factors that affect the occurrence of incomplete fusion and welding defects, it is difficult to detect or predict welding defects based on mechanisms. This issue is not limited to AM, but can also occur in conventional welding.

[0006] In view of the above problems, the present invention aims to provide a defect prediction system, a machine learning device, a defect prediction method, and a program that enable welding defects to be predicted with higher accuracy. [Means for solving the problem]

[0007] In order to solve the above problems, the present invention has the following configuration: That is, a defect prediction system for predicting defects that will occur in welding, a first prediction means for predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of the defect; a second prediction means for predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction means for predicting the size of the defect based on the parameters predicted by the first prediction means and the parameters predicted by the second prediction means; It has.

[0008] Another aspect of the present invention has the following configuration: A machine learning device that generates a trained model for predicting defects that occur in welding, comprising: a first learning processing unit that uses welding parameters as learning data and performs learning processing using a first learning algorithm that outputs discrete values, thereby generating a first trained model that uses the welding parameters as input data and outputs parameters that indicate the size of defects caused by the welding parameters; a second learning processing unit that uses the welding parameters as learning data and performs learning processing using a second learning algorithm that outputs continuous values, thereby generating a second trained model that uses the welding parameters as input data and outputs parameters that indicate the size of a defect caused by the welding parameters; and The parameters output from the first trained model and the parameters output from the second trained model are configured to be convertible into either parameter format.

[0009] Another aspect of the present invention has the following configuration: A defect prediction method for predicting defects that may occur in welding, comprising: a first prediction step of predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of the defect; a second prediction step of predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction step of predicting the size of the defect based on the parameters predicted in the first prediction step and the parameters predicted in the second prediction step; It has.

[0010] Another aspect of the present invention has the following configuration: On the computer, a first prediction step of predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of a defect that will occur during welding; a second prediction step of predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction step of predicting the size of the defect based on the parameters predicted in the first prediction step and the parameters predicted in the second prediction step; Execute the following. [Effects of the Invention]

[0011] According to the present invention, it is possible to predict defects in welding with higher accuracy. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a block diagram showing an example of the configuration of a defect prediction system according to an embodiment of the present invention. [Figure 2] FIG. 1 is a block diagram showing an example of the functional configuration of a defect prediction system according to an embodiment of the present invention. [Figure 3] FIG. 2 is a graph illustrating an example of a defect prediction method according to an embodiment of the present invention. [Figure 4] 1 is a schematic diagram illustrating an example of a defect prediction method according to an embodiment of the present invention. [Figure 5] 1 is a flowchart of a prediction process in a defect prediction system according to an embodiment of the present invention. [Figure 6] FIG. 4 is a graph illustrating a prediction result of the defect prediction method according to the embodiment of the present invention. [Figure 7] FIG. 4 is a graph illustrating a prediction result of the defect prediction method according to the embodiment of the present invention. [Figure 8] FIG. 4 is a graph illustrating a prediction result of the defect prediction method according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. Note that the embodiment described below is one embodiment for explaining the present invention and is not intended to be interpreted as limiting the present invention. Furthermore, not all of the configurations described in each embodiment are necessarily essential configurations for solving the problems of the present invention. In each drawing, the same components are assigned the same reference numerals to indicate corresponding relationships.

[0014] In the following description, "learning" or "machine learning" refers to generating a "trained model" by performing learning using training data and an arbitrary learning algorithm. A trained model is updated as needed as learning progresses using multiple training data, and its output changes even when the input is the same. Therefore, the state of a trained model is not limited to a specific point in time. Here, a model used in learning is referred to as a "training model," and a learning model that has undergone a certain level of learning is referred to as a "trained model." Specific examples of "training data" will be described later, but the configuration may vary depending on the learning algorithm used. Training data may include training data used for the training itself, verification data used to verify the trained model, and test data used to test the trained model. In the following description, the term "training data" is used to collectively refer to data related to training, and the term "training data" is used to refer to data used when performing the training itself. Note that this does not intend to clearly classify training data, verification data, and test data contained in training data. For example, depending on the training, verification, and testing methods, all training data may also be training data.

[0015] First Embodiment Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0016] [Configuration of defect prediction system] 1 is an explanatory diagram illustrating an example configuration of a defect prediction system 100 that can implement the welding defect prediction method according to this embodiment. The defect prediction system 100 is configured, for example, by an information processing device. Although not shown in FIG. 1, the defect prediction system 100 may be connected to a welding system that includes a welding robot that performs welding, and may be configured to predict defects that will occur during welding in conjunction with the welding system.

[0017] The defect prediction system 100 includes a main body 110, an input unit 120, and a display unit 130. The main body 110 includes a CPU (Central Processing Unit) 111, a GPU (Graphics Processing Unit) 112, a ROM 113, a RAM 114, a nonvolatile storage device 115, an input / output interface 116, a communication interface 117, and a video output interface 118. The CPU 111, the GPU 112, the ROM 113, the RAM 114, the nonvolatile storage device 115, the input / output interface 116, the communication interface 117, and the video output interface 118 are connected to each other via buses or signal lines so as to be able to communicate with each other.

[0018] The nonvolatile storage device 115 stores a learning program 115A that executes a learning process using learning data and a predetermined learning algorithm, a trained model 115B that is generated through the execution of the learning program 115A, a prediction program 115C that performs defect prediction using the trained model 115B, and training data 115D. In addition, an operating system and application programs may also be installed in the nonvolatile storage device 115.

[0019] The defect prediction system 100 realizes various functions by executing programs by the CPU 111 and the GPU 112. In this embodiment, the defect prediction system 100 realizes a function of generating a trained model by machine learning and a function of performing actual defect prediction processing using the trained model. The details of these functions will be described later.

[0020] In this embodiment, an example will be described in which an apparatus that performs a learning process to generate a trained model and an apparatus that performs prediction using the trained model are integrated, but the present invention is not limited to this configuration. That is, the apparatus that performs the learning process and the apparatus that performs the prediction may be configured as separate apparatuses. From the viewpoint of versatility, it is more preferable to separate the defect prediction system 100 according to their respective functions.

[0021] The CPU 111 and the GPU 112 are used as arithmetic units when executing the learning program 115A, the trained model 115B, and the prediction program 115C. The ROM 113 stores a BIOS (Basic Input Output System) and the like executed by the CPU 111. The RAM 114 is used as a work area for the programs read from the nonvolatile storage device 115.

[0022] The input / output interface 116 is connected to an input unit 120 that is configured to include a keyboard, a mouse, and the like. A sensor 200 is also connected to the input / output interface 116. Various data output from the sensor 200 is provided to the CPU 111 and the GPU 112 via the input / output interface 116. The sensor 200 may be configured to be moved, for example, by a welding robot (not shown) and to acquire information on the shape of the weld bead and its surroundings. The sensor 200 may be, for example, a laser sensor that acquires reflected light of an irradiated laser beam as height data, or a camera for three-dimensional shape measurement. Furthermore, multiple types of sensors may be provided as the sensor 200.

[0023] Communication interface 117 is a communication module for wired or wireless communication. The communication method used for communication interface 117 is not limited to one, and a combination of multiple communication methods may be used for connection. Communication interface 117 may be connected to, for example, a welding system to acquire various welding-related parameters by communicating with the welding system. Video output interface 118 is connected to display unit 130, which is formed of, for example, a liquid crystal display or an organic EL (Electro-Luminescence) display, and outputs a video signal to display unit 130 according to video data provided by CPU 111.

[0024] A welding system that can be linked with the defect prediction system 100 may be configured to include, for example, a welding robot and a welding power supply unit that provides power for welding. The welding robot is an articulated robot with a welding torch attached to its tip shaft. A continuously supplied filler material (welding wire) is supported at the tip of the welding torch, and welding is performed according to the desired shape. Note that when the defect prediction system 100 and the welding system are configured as an integrated unit, their control units may also be configured as an integrated unit.

[0025] [Pre-trained model] First, the generation of a trained model according to this embodiment will be described. A trained model is generated by executing a learning program. In this embodiment, two trained models are used. Therefore, in this embodiment, two trained models are generated using trained models based on two different learning algorithms. In this embodiment, the training data used as input is common to both trained models. Note that the terms "first" and "second" used in this specification are used merely for convenience to distinguish from other elements, and are not necessarily intended to be interpreted as being limited to specific elements.

[0026] For convenience, the two learning models according to this embodiment will be referred to as the "first learning model" and the "second learning model." Furthermore, the two corresponding trained models will be referred to as the "first trained model" and the "second trained model." In this embodiment, the first learning model and the first trained model use a learning algorithm that outputs discrete values. More specifically, a well-known random forest will be used as an example. Furthermore, the second learning model and the second trained model use a learning algorithm that outputs continuous values. More specifically, a well-known Gaussian process regression will be used as an example. Note that the learning algorithms described here are merely examples, and the types and combinations used may be different.

[0027] The inputs to the first learning model and the first trained model are parameters related to welding (hereinafter also referred to as "welding parameters"). More specifically, image data acquired when sensor 200 is configured as a camera may be included. The image data itself may be used, or, for example, the depth, width (short side), length (long side), and target position deviation amount of the narrow portion may be extracted from the image data and used. Point cloud data acquired when sensor 200 is configured as a laser sensor may also be included. Welding parameters such as welding current, welding voltage, bead distance, and welding speed may also be used. Other parameters such as the feed speed of the filler metal and the tilt angle of the welding torch may also be used.

[0028] The second learning model and the second trained model may also use the same data as above. Note that even if the first learning model and the first trained model and the second learning model and the second trained model originally use the same data, they may be configured to perform different preprocessing before inputting the data.

[0029] The outputs of the first learning model and the first trained model are discrete values, and may be, for example, a plurality of categories. More specifically, for example, the output may indicate which of four categories the input belongs to. Note that the number of categories is not particularly limited, and two or more categories may be used.

[0030] The output of the second learning model and the second trained model is a continuous value, for example, the size of the defect expressed in a predetermined unit.

[0031] [Function Configuration] 2 is a diagram showing an example of the functional configuration of the defect prediction system 100 according to this embodiment. Each component may be realized by the CPU 111 or the GPU 112 reading and executing various data and programs stored in the nonvolatile storage device 115. As described above, the nonvolatile storage device 115 is provided with programs in which multiple learning models based on multiple types of learning algorithms are defined, and learning is performed using these programs.

[0032] The data acquisition unit 201 acquires various data. The acquired data includes data required for defect prediction processing and learning data used in learning processing. The learning data management unit 202 holds and manages the learning data used in learning processing.

[0033] The first learning processing unit 203 generates a first trained model using the training data and the first training model. The first learning processing unit 203 also manages the generated first trained model. The first trained model may be managed by version as history information, or only the latest version may be managed.

[0034] The second learning processing unit 204 generates a second trained model using the training data and the second training model. The second learning processing unit 204 also manages the generated second trained model. The second trained model may be managed by version as history information, or only the latest version may be managed.

[0035] The first prediction unit 205 applies the first trained model generated by the first learning processing unit 203 to various input parameters and outputs the prediction result. The first prediction unit 205 may perform preprocessing according to the input format before inputting various parameters into the first trained model. The degree of learning of the first trained model used here is not particularly limited, and a first trained model that has progressed to a state where a certain level of accuracy is obtained may be used.

[0036] The second prediction unit 206 applies the second trained model generated by the second learning processing unit 204 to various input parameters and outputs the prediction result. The second prediction unit 206 may perform preprocessing according to the input format before inputting various parameters into the second trained model. The level of learning of the second trained model used here is not particularly limited, and a second trained model that has progressed to a state where a certain level of accuracy is obtained may be used.

[0037] Prediction result synthesis unit 207 performs synthesis processing based on the prediction result of first prediction unit 205 and the second prediction result. Prediction result output unit 208 outputs the result predicted by prediction result synthesis unit 207 in a predetermined format. The output format here may be configured to display the result on a screen to the user, or may be configured to provide feedback to the welding system as a control parameter for welding.

[0038] [Learning algorithm] As described above, this embodiment uses a plurality of types of learning algorithms. In this embodiment, a learning algorithm that outputs a continuous value for an input and a learning algorithm that outputs a discrete value are used, and an example of this will be described.

[0039] (Example of a learning algorithm that handles continuous values) FIG. 3 is a graph illustrating Gaussian process regression, which is one type of learning algorithm used in this embodiment. Gaussian process regression is a probabilistic model that handles continuous values with nonlinearity. In the Gaussian process regression algorithm, inputs (corresponding to six black dots in the example of FIG. 3) are fitted using a curve. Furthermore, uncertainties are used for values other than the inputs, as indicated by diagonal lines. Note that since Gaussian process regression is a well-known algorithm, a detailed description thereof will be omitted here. In this embodiment, welding parameters are used as inputs, and predicted values (continuous values) for the outputs are derived. More specifically, a parameter indicating the size of the defect is used as the predicted value for the input. In this example, the parameter indicating the size of the defect is described as two parameters (x, y) expressed in two dimensions. However, this is not limited to this, and three parameters (x, y, z) may be output when the size of the defect is expressed in three dimensions.

[0040] (Example of a learning algorithm that handles discrete values) FIG. 4 is a schematic diagram illustrating a random forest, which is a type of learning algorithm used in this embodiment. In the random forest, several pieces of data are extracted from the training data to set up multiple data groups. Then, using each data group, feature values are randomly selected to generate a tree-structured decision tree. The results of the decision trees are then integrated by majority voting to construct a trained model. In this embodiment, classification is the target, and majority voting is used for integration. Note that since the random forest is a well-known algorithm, a detailed description will be omitted here. In this embodiment, welding parameters are used as input, and a classification (discrete value) for the output is derived. More specifically, one of multiple classifications defined corresponding to the size of the defect is used as the classification for the input. In this embodiment, the explanation will be given using four classifications: "no defect (0)," "small defect (1)," "medium defect (2)," and "large defect (3)."

[0041] [Control Overview] The processing according to this embodiment is roughly divided into two phases: a learning phase and a defect prediction phase. The learning phase is a phase in which a trained model is generated, and the defect prediction phase is a phase in which actual prediction processing is performed using the trained model generated in the learning phase. The processing of each phase is realized by the CPU 111 and GPU 112 of the defect prediction system 100 reading and executing various programs stored in the nonvolatile storage device 115 or the like.

[0042] In the learning phase, the various types of learning data described above are used to repeatedly perform learning processes and verification operations, thereby generating multiple trained models with a certain level of accuracy. Multiple trained models are generated before the control phase is performed, and the latest trained model at that time is used in the control phase. Note that even after the control phase is performed, the learning phase may be repeated as needed, and there is no restriction on the time at which the trained model may be used.

[0043] As described above, in this embodiment, in the learning phase, learning is performed using a learning algorithm that handles discrete values and learning using a learning algorithm that handles continuous values. The degree of each learning is not particularly limited and may be set taking into consideration the characteristics of each learning algorithm, the processing load, etc.

[0044] In the defect prediction phase, prediction processing is performed using multiple trained models by the defect prediction system 100. In this embodiment, the defect prediction system 100 performs both learning (generation of multiple trained models) and defect prediction.

[0045] 5 is a flowchart of the defect prediction process performed by the defect prediction system 100 according to this embodiment. In this embodiment, it is assumed that two trained models have already been generated and are available for use.

[0046] In S501, the defect prediction system 100 inputs welding parameters into the first trained model, i.e., the trained model using random forest, to obtain a prediction result. As described above, the prediction result is one of four classifications for the welding parameters.

[0047] In S502, the defect prediction system 100 inputs the welding parameters to the second trained model, i.e., the trained model by Gaussian process regression, to obtain a prediction result. As described above, as a prediction result, a parameter indicating the size of the defect is obtained for the welding parameters. In this embodiment, the parameter (x a ,y a ) to get the

[0048] In S503, the defect prediction system 100 determines whether or not a defect is present based on the prediction result in S501. In the present embodiment, if any of the four classifications, "small defect (1)," "medium defect (2)," or "large defect (3)," is predicted, it is determined that a defect is present. If it is determined that a defect is present (YES in S503), the processing of the defect prediction system 100 proceeds to S505. On the other hand, if it is determined that no defect is present (NO in S503), the processing of the defect prediction system 100 proceeds to S504.

[0049] In S504, the defect prediction system 100 determines whether or not a defect exists based on the prediction result in S502. In this embodiment, a threshold value for the prediction result (continuous value) is set in advance, and if the prediction result exceeds the threshold value, it is determined that a defect exists. The threshold value here may be set by the user or may be specified depending on the shape or specifications of the object. For example, the threshold value may be set to 0.5 mm. If it is determined that a defect exists (YES in S504), the processing of the defect prediction system 100 proceeds to S505. On the other hand, if it is determined that no defect exists (NO in S504), the processing of the defect prediction system 100 proceeds to S507.

[0050] In S505, the defect prediction system 100 determines that there is a defect based on the prediction results of the two trained models. After that, the processing of the defect prediction system 100 proceeds to S506.

[0051] In S506, the defect prediction system 100 sets the defect size by combining the prediction results of the two trained models. In this embodiment, parameters indicating the defect size are defined for each of the four classifications used in the first trained model. For example, when the defect size is to be represented in two dimensions, two parameters (x, y) are defined in advance, and when the defect size is to be represented in three dimensions, three parameters (x, y, z) are defined.

[0052] For example, in the case of four categories, the following may be set: "No Defects (0)" = (0,0) "Small defect (1)" = (x b1 ,y b1 ) "Defective (2)" = (x b2 ,y b2 ) "Large defect (3)" = (x b3 ,y b3 ) In addition, x b1 <x b2 <x b3 , y b1 <y b2 <y b3 Let's say.

[0053] The set values here may be set based on the shape and specifications of the object, the relationship with a trained model that handles continuous values, etc. In other words, information for converting parameters (classification here) output as discrete values is specified assuming that they will be combined with parameters output as continuous values by other learning algorithms. The information for conversion here may be defined in table format.

[0054] In this embodiment, the size of the defect is set using parameters corresponding to the above classification and the average value of the prediction results obtained in S502. Defect size (x) = (xa +x b ) / 2 Defect size (y) = (y a +y b ) / 2 Then, this processing flow ends.

[0055] In S507, the defect prediction system 100 determines that there is no defect based on the prediction results of the two trained models. After that, the processing of the defect prediction system 100 proceeds to S508.

[0056] In S508, the defect prediction system 100 sets the size of the defect to zero based on the prediction results of the two trained models. Defect size (x) = 0 Defect size (y) = 0 Then, this processing flow ends.

[0057] [Prediction results] The prediction results obtained by the defect prediction method according to this embodiment will be described with reference to FIGS.

[0058] Figure 6 shows the relationship between importance and features when a random forest is used as the first learning model that handles discrete values. This figure shows the results of a 78% accuracy rate for classification using K-cross validation, a well-known evaluation method. In Figure 6, the horizontal axis represents importance, and the vertical axis represents features. It can be seen that features such as the depth of the narrow portion, width of the narrow portion, current, voltage, distance between beads, welding speed, and target position deviation obtained from the input affect defect detection.

[0059] Figure 7 shows the prediction results when Gaussian process regression is used as the second learning model for handling continuous values. In Figure 7, the horizontal axis shows the measured value of the defect diameter [mm], and the vertical axis shows the predicted value of the defect [mm]. The following accuracy was obtained in this prediction result. RMSE (Root Mean Squared Error) = 0.3973 R 2(variance)=0.7146

[0060] Figure 8 shows the prediction results when the defect prediction process shown in Figure 5 is performed using the first trained model with the accuracy shown in Figure 6 and the second trained model with the accuracy shown in Figure 7. In Figure 8, the horizontal axis shows the measured value of the defect diameter [mm], and the vertical axis shows the predicted value of the defect [mm]. The following accuracy was obtained in this prediction result. RMSE=0.3047 R 2 =0.83

[0061] In other words, the defect prediction method according to this embodiment, which uses the first trained model and the second trained model in combination, can achieve higher accuracy than when each model is used alone. In addition, the variance is reduced, which also makes it possible to suppress the occurrence of errors.

[0062] As described above, this embodiment makes it possible to predict welding defects with higher accuracy.

[0063] <Other embodiments> In the above embodiment, the prediction results of the first trained model and the second trained model are combined using a simple average. However, this configuration is not limited to this, and the prediction values may be combined by weighting each prediction result depending on the characteristics and accuracy of the combined learning algorithms.

[0064] Furthermore, when combining the prediction results of the first trained model and the second trained model, the maximum value may be selected from multiple prediction values and set as the prediction result. For example, when higher accuracy in welding is required, setting the prediction result on the more conservative side can improve welding quality.

[0065] Furthermore, when combining the prediction results of the first trained model and the second trained model, the smallest value may be selected from multiple prediction values and set as the prediction result. For example, when welding speed is prioritized over accuracy, setting the prediction result so that the defect size is small can prevent excessive defect detection.

[0066] Furthermore, when combining the prediction results of the first trained model and the second trained model, a configuration may be adopted in which a plurality of combining methods are combined. For example, whether to use the average value, maximum value, or minimum value to set the prediction value may be switched depending on the difference between the plurality of prediction results.

[0067] In addition, if the difference between multiple prediction results exceeds a predetermined threshold or if one prediction result shows an extreme value, one of the prediction results may be considered to be erroneous and the predicted value may be ignored.

[0068] In the above embodiment, two types of learning algorithms, one for handling discrete values and one for handling continuous values, are used, but three or more types of learning algorithms may be combined. Furthermore, defect prediction may be performed using multiple learning algorithms that handle discrete values, as long as they are different learning algorithms. Similarly, defect prediction may be performed using multiple learning algorithms that handle continuous values.

[0069] In the above embodiment, when a learning algorithm that handles discrete values and a learning algorithm that handles continuous values are combined and their outputs are synthesized, the discrete values (classifications) are converted into continuous values and then synthesized. However, the present invention is not limited to this configuration, and the defect size may be predicted by converting continuous values into discrete values.

[0070] Furthermore, the welding parameters of the welding system may be adjusted based on the defect size obtained by the defect prediction according to the present embodiment. The degree of adjustment is not particularly limited, but the welding parameters may be adjusted according to, for example, the frequency, position, and timing of appearance of defects of a certain size.

[0071] In the present invention, a program or application for realizing the functions of one or more of the above-described embodiments can be supplied to a system or device via a network or a storage medium, etc., and one or more processors in the computer of the system or device can read and execute the program.

[0072] Alternatively, the functions may be realized by a circuit that realizes one or more functions, such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0073] As described above, the present specification discloses the following: (1) A defect prediction system for predicting defects that may occur in welding, a first prediction means for predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of the defect; a second prediction means for predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction means for predicting the size of the defect based on the parameters predicted by the first prediction means and the parameters predicted by the second prediction means; A defect prediction system having This configuration makes it possible to predict welding defects with higher accuracy.

[0074] (2) the first trained model is generated using a first learning algorithm that outputs discrete values; The second trained model is generated using a second learning algorithm that outputs continuous values. (1) A defect prediction system according to (1). With this configuration, it is possible to predict welding defects using the prediction results of a trained model that outputs discrete values and a second trained model that outputs continuous values.

[0075] (3) the discrete value output by the first trained model is one of a plurality of classifications; The defect prediction system described in (2), wherein the third prediction means converts the classification, which is the parameter predicted by the first prediction means, into a value indicating a defect size predefined for that classification, and then combines it with the parameter predicted by the second prediction means. With this configuration, even when combining trained models with different output formats, it is possible to synthesize the results and predict the size of welding defects.

[0076] (4) the first trained model is generated using a first learning algorithm that outputs continuous values; The second trained model is generated using a second learning algorithm that outputs continuous values. (1) A defect prediction system according to (1). With this configuration, it is possible to predict welding defects using the prediction results of multiple types of trained models that output continuous values.

[0077] (5) the first trained model is generated using a first learning algorithm that outputs discrete values; The second trained model is generated using a second learning algorithm that outputs discrete values. (1) A defect prediction system according to (1). With this configuration, it is possible to predict welding defects using the prediction results of multiple types of trained models that output discrete values.

[0078] (6) A defect prediction system according to any one of (1) to (5), wherein the third prediction means predicts the size of the defect using the parameters predicted by the first prediction means and the average, maximum, or minimum value of the parameters predicted by the second prediction means. According to this configuration, it is possible to predict the size of a defect in a weld using the average, maximum, or minimum value of a plurality of prediction parameters.

[0079] (7) A machine learning device that generates a trained model for predicting defects that occur in welding, a first learning processing unit that uses welding parameters as learning data and performs learning processing using a first learning algorithm that outputs discrete values, thereby generating a first trained model that uses the welding parameters as input data and outputs parameters that indicate the size of defects caused by the welding parameters; a second learning processing unit that uses the welding parameters as learning data and performs learning processing using a second learning algorithm that outputs continuous values, thereby generating a second trained model that uses the welding parameters as input data and outputs parameters that indicate the size of a defect caused by the welding parameters; and A machine learning device configured to be able to convert parameters output from the first trained model and parameters output from the second trained model into either parameter format. With this configuration, multiple trained models capable of predicting the size of welding defects can be generated using multiple different learning algorithms, and the output results of each trained model can be associated and used for prediction even if they are in different formats.

[0080] (8) A defect prediction method for predicting defects that may occur in welding, comprising: a first prediction step of predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of the defect; a second prediction step of predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction step of predicting the size of the defect based on the parameters predicted in the first prediction step and the parameters predicted in the second prediction step; A defect prediction method comprising: This configuration makes it possible to predict welding defects with higher accuracy.

[0081] (9) To the computer, a first prediction step of predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of a defect that will occur during welding; a second prediction step of predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction step of predicting the size of the defect based on the parameters predicted in the first prediction step and the parameters predicted in the second prediction step; A program to execute. This configuration makes it possible to predict welding defects with higher accuracy. [Explanation of symbols]

[0082] 100...Defect prediction system 110...Main unit 111...CPU 112...GPU 113...ROM 114...RAM 115...Nonvolatile storage device 116... Input / output interface 117...Communication interface 118...Video output interface 120...Input section 130...Display section 200...sensor 201...Data acquisition section 202...Learning Data Management Department 203...first learning processing unit 204...second learning processing unit 205...First Prediction Section 206...Second Prediction Section 207...Prediction result synthesis unit 208...Prediction result output unit

Claims

1. A defect prediction system for predicting defects that will occur in welding, a first prediction means for predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of the defect; a second prediction means for predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction means for predicting the size of the defect based on the parameters predicted by the first prediction means and the parameters predicted by the second prediction means; and The first trained model is generated using a first learning algorithm that outputs discrete values; The second trained model is generated using a second learning algorithm that outputs continuous values; the discrete value output by the first trained model is one of a plurality of classifications; the third prediction means converts the classification, which is the parameter predicted by the first prediction means, into a value indicating a defect size predefined for the classification, and then combines the converted value with the parameter predicted by the second prediction means; Defect prediction system.

2. A machine learning device that generates a trained model for predicting defects that occur in welding, a first learning processing unit that uses welding parameters as learning data and performs learning processing using a first learning algorithm that outputs discrete values to generate a first trained model that uses the welding parameters as input data and outputs parameters that indicate the size of a defect caused by the welding parameters; a second learning processing unit that uses the welding parameters as learning data and performs learning processing using a second learning algorithm that outputs continuous values, thereby generating a second trained model that uses the welding parameters as input data and outputs parameters that indicate the size of a defect caused by the welding parameters; and The parameters output from the first trained model and the parameters output from the second trained model are configured to be convertible into any one of the parameter formats; the discrete value output by the first trained model is one of a plurality of classifications; The classification, which is a parameter predicted using the first trained model, is converted into a value indicating a defect size predefined for that classification, and then combined with the parameter predicted using the second trained model. Machine learning device.

3. A defect prediction method for predicting defects that will occur in welding, comprising: a first prediction step of predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of the defect; a second prediction step of predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction step of predicting the size of the defect based on the parameters predicted in the first prediction step and the parameters predicted in the second prediction step; and The first trained model is generated using a first learning algorithm that outputs discrete values; The second trained model is generated using a second learning algorithm that outputs continuous values; the discrete value output by the first trained model is one of a plurality of classifications; the third prediction step converts the classification, which is the parameter predicted in the first prediction step, into a value indicating a defect size predefined for the classification, and then combines the converted value with the parameter predicted in the second prediction step; Defect prediction methods.

4. On the computer, a first prediction step of predicting the defect using a first trained model that receives welding parameters as input and outputs parameters indicating the size of a defect that will occur during welding; a second prediction step of predicting the defect using a second trained model that receives the welding parameters as input and outputs parameters indicating the size of the defect; a third prediction step of predicting the size of the defect based on the parameters predicted in the first prediction step and the parameters predicted in the second prediction step; A program for executing The first trained model is generated using a first learning algorithm that outputs discrete values; The second trained model is generated using a second learning algorithm that outputs continuous values; the discrete value output by the first trained model is one of a plurality of classifications; the third prediction step converts the classification, which is the parameter predicted in the first prediction step, into a value indicating a defect size predefined for the classification, and then combines the converted value with the parameter predicted in the second prediction step; program.

Images