Method for estimating nugget diameter in spot welding

Abstract

To provide a method for estimating a nugget diameter in spot welding, which does not need man-hours for preparing a large quantity of data and which, when a deficiency occurs in inner function in a machine learning model, enables a person to easily understand a portion to be corrected and a correction method.SOLUTION: The method for estimating a nugget diameter in spot welding comprises: an obtaining step of obtaining a data set including teacher data in which a nugget diameter at the time when the spot welding is executed while changing disturbance conditions and a waveform of welding are associated with each other, for each disturbance condition; a feature amount generating step; a selecting step; a regression equation generating step of generating, by machine learning, for each of a plurality of regression models, a regression equation in which a feature amount related to a selected feature parameter is set as an explanatory variable and the nugget diameter is set as a target variable; and a regression equation selecting step of selecting a regression equation having high correlation with the nugget diameter, out of a plurality of regression equations subjected to the machine learning.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present disclosure relates to a method for estimating a nugget diameter in spot welding. [Background technology]

[0002] Conventionally, a technique for diagnosing the welding quality of spot welding is known using a machine learning model that has learned the judgment results regarding the quality of welding based on actual inspection such as a chisel test (Patent Document 1). In this technique, time-series data showing the time-series transition during welding such as the amount of displacement between electrodes is input as an internal function, and the judgment results based on the actual inspection are directly input to a machine learning model configured to output the quality of welding. In this way, the internal function is learned without adjusting thresholds or parameters. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2016-203246 A Summary of the Invention [Problem to be solved by the invention]

[0004] In the conventional technology, the system learns the internal function for judging the welding quality from scratch. In other words, the selection of extraction items for extracting elements related to the quality of the welding quality from various features included in the time-series data must also be performed during the learning process. Therefore, in order to build a more accurate internal function, it requires a lot of work to prepare a large amount of data. In addition, in the conventional technology, since the system builds the internal function, the internal function may contain complex features. Therefore, when a defect occurs in the internal function, it may be difficult for a human to understand the parts that need to be corrected and how to correct them. [Means for solving the problem]

[0005] The present disclosure can be realized in the following forms.

[0006] According to one embodiment of the present disclosure, a method for estimating a nugget diameter in spot welding is provided. A method for estimating a nugget diameter in spot welding includes: an acquisition step of acquiring a dataset including teacher data in which the nugget diameter when the spot welding is performed while changing a disturbance condition that can affect the size of the nugget diameter and a welding waveform, which is time-series data of welding parameters in the spot welding, are associated with each of the disturbance conditions; a feature amount generation step of generating values ​​for each of a plurality of predetermined feature parameters as feature amounts of the welding waveform by inputting the dataset into a trained first learning model; a selection step of selecting the feature parameters having a high correlation with the nugget diameter by inputting the feature parameters and the feature amounts corresponding to the feature parameters into a trained second learning model; a regression equation creation step of inputting the feature amounts related to the selected feature parameters into each of a plurality of regression models that are a plurality of machine learning models, thereby creating a regression equation by machine learning for each of the plurality of regression models in which the feature amounts related to the selected feature parameters are used as explanatory variables and the nugget diameter is used as a target variable; and a regression equation selection step of selecting a regression equation having a high correlation with the nugget diameter from the plurality of regression equations trained by machine learning. According to this embodiment, the feature parameters and feature quantities used as explanatory variables of the regression equation can be selected by a learning model in which extraction items for extracting elements with a high correlation with the nugget diameter from various features included in the welding waveform are preset. Furthermore, according to this embodiment, the regression equation is created using a data set including teacher data in which the nugget diameter and the welding waveform when spot welding is performed while changing the disturbance condition are linked to each disturbance condition. This makes it possible to suppress a decrease in accuracy related to the selection of the feature parameters even when the number of teacher data included in the data set is small. Therefore, it is possible to select a more accurate regression equation and estimate the nugget diameter without requiring man-hours for preparing a large amount of data. Furthermore, according to this embodiment, the type of the feature parameters and the selection method of the feature parameters are preset, and the feature quantities related to the selected feature parameters are used as explanatory variables of the regression equation.Therefore, humans can understand the basis for selecting feature parameters and feature quantities, and the elements that make up the explanatory variables. In other words, humans can recognize the parameters that are the basis for deriving the regression equation during the learning process. Therefore, when a defect occurs in the regression equation, humans can easily understand the parts that need to be corrected and how to correct them. The present disclosure may be realized in various forms other than the above-described method for estimating a nugget diameter in spot welding, for example, in the form of a manufacturing method for an estimation device for a nugget diameter in spot welding, a control method for the estimation device, a computer program for implementing the control method, a non-transitory recording medium on which the computer program is recorded, etc. [Brief description of the drawings]

[0007] [Figure 1] 1 is a schematic diagram showing a schematic configuration of a spot welding device according to an embodiment of the present invention; [Diagram 2] FIG. 2 is a diagram showing a schematic configuration of a control device. [Diagram 3] 4 is a flowchart showing a method for estimating a nugget diameter in the present embodiment. [Figure 4] 11 is a flowchart showing details of an acquisition process. [Diagram 5] Graph showing welding waveforms of resistance values ​​for each disturbance condition. [Figure 6] Graph showing welding waveforms of pressure for each disturbance condition. [Figure 7] FIG. 13 is a diagram showing an example of nugget diameters when welding is performed under different disturbance conditions. [Figure 8] 11 is a flowchart showing details of a feature generating step. [Figure 9] A figure showing a portion of pre-selection feature data output from the first learning model. [Figure 10] 11 is a flowchart showing details of a selection process. [Figure 11] A figure showing a portion of the selected feature data output from the second learning model. [Figure 12] FIG. 4 is a diagram showing a portion of input feature amount data. [Figure 13] 11 is a flowchart showing details of a regression equation creating process. [Figure 14] 11 is a flowchart showing details of a regression equation selection step. [Figure 15] A graph showing the contribution rate for each regression equation. [Figure 16] 10 is a flowchart showing details of a selection process in a second embodiment. [Figure 17] 13 is a graph showing the relationship between the feature amounts of selected feature parameters and the nugget diameter. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] A. First embodiment: 1 is a schematic diagram showing a schematic configuration of a spot welding device 10 in this embodiment. The spot welding device 10 is a device that melts and joins a workpiece W, which is a plurality of overlapping metal plates W1, W2. The spot welding device 10 includes a spot welding gun G, a robot arm RA, a control device 100, and a measurement mechanism 9.

[0009] The spot welding gun G includes a gun body 1, a pair of electrodes, an upper electrode 2 and a lower electrode 3, an electrode lifting device 4, and a current adjusting device 5. The gun body 1 is held by a robot arm RA. The upper electrode 2 is attached to the upper part 1a of the gun body 1 via the electrode lifting device 4. The lower electrode 3 is disposed in the lower part 1b of the gun body 1. When welding the workpiece W, the workpiece W is sandwiched and pressurized between the upper electrode 2 and the lower electrode 3, and a current is passed between the upper electrode 2 and the lower electrode 3. As a result, the workpiece W melts due to resistance heating and then solidifies, thereby joining the multiple metal plates W1, W2. A nugget is formed at the joint interface where the multiple metal plates W1, W2 are joined. The nugget is a joint portion where the metal plates W1, W2 melt and then solidify. The shape of the nugget is generally roughly a go stone shape with the joint portion as the center plane.

[0010] The electrode lifting device 4 is an electric device that holds and lifts the upper electrode 2. The electrode lifting device 4 is attached to the tip of the upper part 1a of the gun body 1. The electrode lifting device 4 includes a servo motor 41 and a lifting member 42 that is connected to the drive shaft of the servo motor 41. The electrode lifting device 4 lifts and lowers the lifting member 42 by operating the servo motor 41 in accordance with a command signal from the control device 100. As a result, the material to be welded W is clamped between the upper electrode 2 and the lower electrode 3.

[0011] The current adjusting device 5 adjusts the value of the welding current (hereinafter, referred to as the current value) flowing between the upper electrode 2 and the lower electrode 3 in response to a current command transmitted from the control device 100. The current adjusting device 5 is, for example, a device equipped with a variable resistor or a device equipped with a converter.

[0012] The control device 100 controls the operation of the spot welding device 10, and estimates the size of the nugget in spot welding (hereinafter, nugget diameter D1) as a means for determining the welding quality. That is, in this embodiment, the control device 100 also functions as an estimation device for estimating the nugget diameter D1 in spot welding. Note that the estimation device may be configured separately from the control device 100 and may be configured to communicate data by wire or wirelessly.

[0013] 2 is a diagram showing a schematic configuration of the control device 100. The control device 100 includes a communication unit 30, a display 40, an input operation unit 50, a storage unit 60, and a CPU 20. The control device 100 is, for example, a computer including the components 20 to 60. The communication unit 30 connects the other components in the spot welding device 10 to the control device 100 so that they can communicate with each other. The display 40 is, for example, a liquid crystal display, and displays information in response to a command from the CPU 20. The input operation unit 50 includes, for example, a keyboard and a mouse, and receives instructions from a user.

[0014] The storage unit 60 stores various information including various programs for controlling the operation of the spot welding device 10, a data set D described later, and a plurality of machine learning models M described later. In this embodiment, the machine learning model M is composed of a first learning model M1, a second learning model M2, a first regression model MR1, a second regression model MR2, and a third regression model MR3. The data set D and the machine learning model M will be described in detail later. The storage unit 60 includes a RAM, a ROM, a rewritable non-volatile memory, and the like.

[0015] The CPU 20 functions as an operation control unit 210, an acquisition unit 220, a feature generation unit 230, a feature selection unit 240, a regression equation creation unit 250, and a regression equation selection unit 260 by deploying various programs stored in the memory unit 60.

[0016] The operation control unit 210 controls the operation of the spot welding apparatus 10. The operation of the spot welding apparatus 10 is controlled according to parameters (hereinafter, welding parameters) that are set in advance for welding. That is, the welding parameters are welding conditions that are set in advance when spot welding is performed. The welding parameters are, for example, a current value, a voltage value, a current flow time, a current flow timing, a pressure applied to the electrodes 2 and 3, a pressure application timing, and a displacement amount of the upper electrode 2 (hereinafter, electrode displacement). The operation control unit 210 controls these operations in an integrated manner. The operation of the spot welding apparatus 10 is controlled by the operation control unit 210, for example, when the welding parameters are set by a user via the input operation unit 50.

[0017] The acquisition unit 220 acquires a data set D including a nugget diameter D1 and various data D2 measured by the measurement mechanism 9. The feature amount generation unit 230 inputs the data set D to a trained first learning model M1 to generate a feature amount for each feature parameter described later. The feature amount selection unit 240 inputs data such as the generated feature amount to a trained second learning model M2 to select the feature parameter described later. The regression equation creation unit 250 creates multiple regression equations F1 to F3 for estimating the nugget diameter D1 in spot welding by machine learning the first regression model MR1 to the third regression model MR3. The regression equation selection unit 260 selects the regression equations F1 to F3 having a high correlation with the nugget diameter D1. At least a part of the functions of the CPU 20 may be realized by a hardware circuit.

[0018] As shown in Fig. 1, measurement mechanism 9 is a mechanism that measures actual values ​​of welding parameters as physical quantities required for welding using spot welding device 10. The actual values ​​of welding parameters referred to here are values ​​of welding parameters that are actually observed when spot welding is performed according to commands related to welding conditions transmitted from operation control unit 210, i.e., preset welding parameters. In this embodiment, measurement mechanism 9 includes a current measurement unit 91, a voltage measurement unit 92, a resistance value calculation unit 93, a welding force measurement unit 94, and an electrode displacement measurement unit 95, as shown in Fig. 2.

[0019] The current measuring unit 91 measures the actual value of the current flowing between the upper electrode 2 and the lower electrode 3. The current measuring unit 91 is, for example, a current sensor. The current measuring unit 91 measures the current value at predetermined time intervals while spot welding is being performed (hereinafter, referred to as during welding).

[0020] The voltage measuring unit 92 measures the voltage (potential difference) between the upper electrode 2 and the lower electrode 3. The voltage measuring unit 92 is, for example, a voltage sensor. The voltage measuring unit 92 measures the voltage value at predetermined time intervals during welding.

[0021] The resistance value calculation unit 93 calculates the value of electrical resistance (hereinafter, resistance value) using the current value and voltage value measured when electricity is applied. Specifically, the resistance value calculation unit 93 calculates the resistance value by dividing the voltage value by the current value. At this time, the current value and voltage value measured at the same timing during welding are used to calculate the resistance value. Note that the resistance value calculation unit 93 may be realized as one function of the CPU 20.

[0022] The pressing force measuring unit 94 measures the pressing force applied by the electrodes 2, 3 to the workpiece W. The pressing force measuring unit 94 is, for example, a load cell housed inside the electrode lifting device 4. The pressing force measuring unit 94 measures the pressing force at predetermined time intervals during welding.

[0023] The electrode displacement measuring unit 95 measures the elevation position of the upper electrode 2. The electrode displacement measuring unit 95 is, for example, an encoder that is housed inside the electrode lifting device 4 and detects the rotation angle position of the output shaft of the servo motor 41 to measure the elevation position of the upper electrode 2. The electrode displacement measuring unit 95 measures the elevation position of the upper electrode 2 at predetermined time intervals during welding.

[0024] The actual values ​​of the welding parameters measured by the measurement mechanism 9 are transmitted to the CPU 20 via the communication unit 30. The measurement mechanism 9 may further include a mechanism for measuring other welding parameters and a mechanism for measuring the nugget diameter D1. For example, when the nugget diameter D1 is measured by non-destructive testing, a nugget diameter measuring unit for measuring the nugget diameter D1 may be provided. In this case, the nugget diameter measuring unit is, for example, a device for measuring the nugget diameter D1 by an ultrasonic method or an electromagnetic method.

[0025] FIG. 3 is a flowchart showing a method for estimating the nugget diameter D1 in this embodiment. In this embodiment, the control device 100 (FIG. 2) executes steps S1 to S9 shown in FIG. 3, and the regression equations F1 to F3 for estimating the nugget diameter D1 are acquired by machine learning. In this way, when the welding conditions (for example, at least one value of a plurality of welding parameters) are changed, the changed welding parameters, etc. are input to the regression equations F1 to F3, and the nugget diameter D1 when spot welding is performed under the changed welding conditions can be estimated. The method for estimating the nugget diameter D1 in spot welding is executed in the order of an acquisition step (step S1), a feature amount generation step (step S3), a selection step (step S5), a regression equation creation step (step S7), and a regression equation selection step (step S9).

[0026] 4 is a flowchart showing details of the acquisition step (step S1). The acquisition step (step S1) is a step of acquiring a data set D including teacher data in which a nugget diameter D1 and a welding waveform D2, which will be described later, when spot welding is performed while changing the disturbance condition are linked to each disturbance condition. In the acquisition step (steps S11 to S18), a welding step (steps S11 to S14) and a data set generating step (steps S15 to S18) are executed in this order. Note that the teacher data also includes conditions when the disturbance conditions are normal, which will be described later (hereinafter, referred to as reference disturbance conditions).

[0027] The disturbance condition is a condition that can affect the size of the nugget diameter D1. Specifically, the disturbance condition is a condition for intentionally creating a non-standard state different from the normal state when spot welding in a normal state is used as a reference. By changing the disturbance condition, for example, spot welding is performed in a state where a gap is formed between the multiple metal plates W1, W2, or in a state where at least one of the contact angle and contact position of each electrode 2, 3 with respect to the metal plates W1, W2 is different from the normal state. Here, in this embodiment, as shown in FIG. 1, the normal state is a state in which each electrode 2, 3 contacts the metal plates W1, W2 placed on a horizontal surface vertically. In other words, the case where welding is performed in a state where the contact angle of each electrode 2, 3 with respect to the metal plates W1, W2 is different from the normal state refers to a case where at least one of the upper electrode 2 and the lower electrode 3 contacts the metal plates W1, W2 placed on a horizontal surface in a non-vertical state. The non-vertical state is, for example, a state inclined by 30° with respect to the horizontal plane. In addition, the case where the electrodes 2, 3 are welded in a state where the contact positions of the metal plates W1, W2 are different from the normal state refers to, for example, the case where the electrodes 2, 3 are in contact with the ends W1a, W2a of the metal plates W1, W2 and the ends W1a, W2a are welded. Note that the disturbance conditions are not limited to these, and other conditions may be used. In the following, a sample welded in a normal state is also referred to as a reference sample, and a sample welded in a non-reference state where the disturbance conditions are changed from the reference disturbance conditions is also referred to as a non-reference sample.

[0028] 4, first, a welding process (steps S11 to S14) is performed. The welding process (steps S11 to S14) is a process of welding each sample and measuring various data related to welding in order to obtain teacher data constituting a data set D.

[0029] In step S11, disturbance conditions are set to obtain data when spot welding is performed while changing disturbance conditions. The disturbance conditions are set, for example, by a user via the input operation unit 50. When data of a reference sample is obtained, the disturbance conditions are set to be "reference disturbance conditions". When data of a non-reference sample is obtained, desired disturbance conditions changed from the reference disturbance conditions are set. At this time, when welding one sample, one type of disturbance condition may be changed with respect to the reference disturbance conditions, or two or more types of disturbance conditions may be changed simultaneously. In this embodiment, a reference sample and a plurality of non-reference samples each having one type of disturbance condition changed from the reference disturbance conditions are prepared. Specifically, four types of samples, namely, a first non-reference sample, a second non-reference sample, and a third non-reference sample as the non-reference samples, and a reference sample of the reference disturbance conditions, are formed by spot welding. For example, 25 pieces of each of the reference sample and the first to third non-reference samples are prepared. In this case, a total of 100 pieces of data are obtained as teacher data.

[0030] After step S11, step S12 is executed. In step S12, the operation control unit 210 starts spot welding. In this embodiment, four types of samples are welded in sequence. Note that the samples have the same setting values ​​for the welding parameters, but only different disturbance conditions.

[0031] After step S12, step S13 is executed. In step S13, acquisition unit 220 acquires the actual measured values ​​of the welding parameters. Specifically, measurement mechanism 9 measures each value of the welding parameters at predetermined time intervals. In this embodiment, four welding parameters, namely, current value, voltage value, welding force, and electrode displacement, are measured. Then, resistance value calculation unit 93 calculates the resistance value using the actual measured values ​​of the current value and voltage value. Thereafter, acquisition unit 220 acquires the actual measured values ​​of the welding parameters and the resistance value.

[0032] After step S13, step S14 is executed. Operation control unit 210 determines that welding is completed (step S14: Yes) when a preset current time has elapsed from the time when welding started (hereinafter, the start time), for example. If "Yes" is determined in step S14, steps S15 and S16 are executed.

[0033] In response to this, operation control unit 210 determines that welding is not complete if, for example, a preset current application time has not elapsed since the start of welding (step S14: No). If the determination in step S14 is "No," operation control unit 210 continues the welding operation of spot welding device 10, for example, by continuing current application to electrodes 2 and 3. Then, step S13 is repeatedly executed until the determination in step S14 is "Yes." That is, the actual measured values ​​of the welding parameters are continuously measured from the start of welding to the completion of welding. It should be noted that the determination as to whether welding is complete may be performed by other methods.

[0034] After the welding process (steps S11 to S14), a data set generating process (steps S15 to S18) is executed. The data set generating process (steps S15 to S18) is a process for generating a data set D by acquiring training data from data of the sample produced in the welding process (steps S11 to S14).

[0035] After step S14, steps S15 and S16 are executed. In step S15, the acquisition unit 220 generates a welding waveform D2. The welding waveform D2 is time-series data in which the actual measured values ​​of the welding parameters are arranged in chronological order. In other words, the welding waveform D2 is data representing the transition of the actual measured values ​​for each welding parameter during welding.

[0036] FIG. 5 is a graph showing the welding waveform D2 of the resistance value for each disturbance condition. FIG. 6 is a graph showing the welding waveform D2 of the pressing force for each disturbance condition. FIGS. 5 and 6 show an example of the welding waveform D2 for each resistance value and pressing force among the welding parameters measured or calculated by the measurement mechanism 9. The vertical axis of FIG. 5 represents the actual measured value of the resistance value. The horizontal axis of FIG. 5 represents the elapsed time from the start of welding. The vertical axis of FIG. 6 represents the actual measured value of the pressing force. The horizontal axis of FIG. 6 represents the elapsed time from the start of welding. In FIGS. 5 and 6, the welding waveforms D2 of the reference sample and the first to third non-reference samples are represented by different line types in order to make it easier to understand the differences due to the contents of the disturbance conditions.

[0037] As shown in Figures 5 and 6, the shape of the welding waveform D2 varies depending on the contents of the disturbance conditions. In other words, the welding waveform D2 shows different tendencies depending on the disturbance conditions. The welding waveform D2 may be generated as a graph as shown in Figures 5 and 6, or may be generated as a collection of numerical data such as a table. The welding waveform D2 is displayed on the display 40 as appropriate.

[0038] The welding waveform D2 to which pre-processing for adjusting the data area and the like has been appropriately performed may be used as the teacher data. The data area is, for example, the width of the data on the horizontal axis of FIG. 5, and the time width. As shown in FIG. 5, some welding parameters continuously show a certain level of fluctuation from the start of welding to the end of welding, while as shown in FIG. 6, some welding parameters show extremely small fluctuations for a while from the start of welding regardless of disturbance conditions. Therefore, in order to easily compare welding waveforms D2 related to different welding parameters when extracting feature amounts from multiple welding waveforms D2 in later learning, pre-processing for unifying the data area may be performed. Note that the mode of pre-processing is not limited to this. For example, when a feature portion such as a characteristic peak is observed in the welding waveform D2, pre-processing may be performed to generate a welding waveform D2 by cutting out a data area including the feature portion.

[0039] As shown in Fig. 4, in step S16, the nugget diameter D1 is measured for each sample with varying disturbance conditions. In this embodiment, the reference sample and the first to third non-reference samples are experimental samples for obtaining training data. Therefore, these nugget diameters D1 are measured by destructive testing such as a chisel test. Note that when a product produced in an actual manufacturing process is used as training data, the nugget diameter D1 may be measured by non-destructive testing or the like.

[0040] FIG. 7 is a diagram showing an example of a nugget diameter D1 when spot welding is performed while changing the disturbance condition. FIG. 7 shows data relating to some samples among the samples used as teacher data. As shown in FIG. 7, the nugget diameter D1 varies depending on the content of the disturbance condition. Specifically, a case where a nugget diameter D1 larger than the nugget diameter D1 in the reference sample is formed as in the third non-reference sample, and a case where a nugget diameter D1 smaller than the nugget diameter D1 in the reference sample is formed as in the first and second non-reference samples can be observed. Note that, between step S14 and step S18, one of steps S15 and S16 may be performed first, or they may be performed simultaneously in parallel.

[0041] As shown in Fig. 4, step S18 is executed after step S15 and step S16. In step S18, the acquisition unit 220 generates teacher data for each sample, in which the welding waveform D2 generated in step S15 and the nugget diameter D1 acquired in step S16 are linked for each disturbance condition. Then, the acquisition unit 220 generates a data set D including a plurality of teacher data. The generated data set D is stored in the storage unit 60. The acquisition step (step S1) is ended by executing each step (step S11 to step S18) up to step S18 shown in Fig. 4.

[0042] The number and types of teacher data included in the data set D are not limited to those described above. For example, the data set D may include data related to samples under disturbance conditions different from the first to third non-reference samples. Also, a plurality of samples under the same disturbance conditions may be prepared and each of them may be used as teacher data.

[0043] As shown in Fig. 3, after the acquisition step (step S1), a feature generation step (step S3) is executed. Fig. 8 is a flowchart showing the details of the feature generation step (step S3). The feature generation step (step S3) is a step of inputting the data set D to a trained first learning model M1 (Fig. 2) to generate values ​​for each of a plurality of predetermined feature parameters as feature amounts of the welding waveform D2.

[0044] In step S31, the feature generating unit 230 reads the data set D stored in the storage unit 60.

[0045] After step S31, step S33 is executed. In step S33, the feature generation unit 230 inputs the read data set D to the trained first learning model M1. Then, in step S35, the feature generation unit 230 acquires pre-selection feature data DF1 output from the first learning model M1. The pre-selection feature data DF1 is displayed on the display 40 as appropriate.

[0046] Fig. 9 is a diagram showing a portion of pre-sorting feature data DF1 output from first learning model M1. In Fig. 9, feature parameters A to Q are shown as an example of feature parameters related to resistance value, and feature parameters a to f are shown as an example of feature parameters related to welding force. In Fig. 9, feature amounts corresponding to the feature parameters are shown for each sample and for each welding parameter.

[0047] The first learning model M1 is a trained machine learning model M. The first learning model M1 stores a program for extracting components (feature quantities) that indicate characteristics of each welding waveform D2 from a plurality of types of welding waveforms D2 that exist for each welding parameter for one sample when a data set D is input. The first learning model M1 is configured, for example, by an existing program created using python, which is a type of programming language.

[0048] In the first learning model M1, a plurality of feature parameters are set in advance as extraction items for extracting features included in the welding waveform D2 as feature quantities. Therefore, in step S33 (FIG. 8), when the data set D is input to the first learning model M1, the feature quantity generation unit 230 calculates the value of the feature parameter as a feature quantity for each welding waveform D2, as shown in FIG. 9. Then, in step S35 (FIG. 8), the feature quantity generation unit 230 generates pre-sorting feature quantity data DF1 as shown in FIG. 9 as data summarizing the correspondence between the feature parameters and the feature quantities for each sample and welding parameter. In other words, steps S33 and S35 shown in FIG. 8 are processes of automatically calculating feature quantities corresponding to the feature parameters by inputting the data set D into the algorithm of the existing learning model M1.

[0049] The characteristic parameters are, for example, the maximum value, the minimum value, the variance value, and the Fourier coefficient of the welding waveform D2. In this embodiment, about 1000 types of characteristic parameters are preset in the first learning model M1. That is, in the pre-sorting characteristic data DF1, values ​​related to about 1000 types of characteristic parameters are generated as characteristic amounts for each sample of the welding waveform D2.

[0050] Furthermore, in the pre-sorting characteristic amount data DF1, data such as characteristic amounts related to each sample is associated with the nugget diameter D1 of each sample.

[0051] The program constituting the first learning model M1 may be a program other than an existing program created by python. The first learning model M1 may have any feature parameters predefined and may calculate the value of each feature parameter as a feature when the data set D is input. The number and types of feature parameters and the value of the feature are not limited to these. The type of feature parameter may be different for each welding parameter or may be the same. The feature generation process (step S3) is completed by executing each process (steps S31 to S35) up to step S35 shown in FIG. 8.

[0052] As shown in Fig. 3, after the feature generating step (step S3), a selection step (step S5) is executed. Fig. 10 is a flowchart showing the details of the selection step (step S5). The selection step (step S5) is a step of selecting feature parameters having a high correlation with the nugget diameter D1 from among the multiple feature parameters included in the pre-selection feature data DF1 by inputting the pre-selection feature data DF1 to the trained second learning model M2 (Fig. 2).

[0053] In step S51, the feature amount selection unit 240 reads the pre-selection feature amount data DF1 stored in the storage unit 60.

[0054] After step S51, step S53 is executed. In step S53, the feature selection unit 240 inputs the read pre-selection feature data DF1 to the trained second learning model M2. Then, in step S55, the feature selection unit 240 acquires the post-selection feature data DF2 output from the second learning model M2. The post-selection feature data DF2 is displayed on the display 40 as appropriate.

[0055] 11 is a diagram showing a part of the sorted feature data DF2 output from the second learning model M2. In FIG. 11, the data added by executing the sorting step (step S5) shown in FIG. 10 is added to the data related to the resistance value in FIG.

[0056] The second learning model M2 is a trained machine learning model M. The second learning model M2 stores a program for selecting feature parameters having a high correlation with the nugget diameter D1 when the pre-selection feature amount data DF1 is input. The second learning model M2 is configured by an existing program created by, for example, python, which is a type of programming language. The second learning model M2 includes, for example, a first program for calculating a correlation coefficient in order to select feature parameters having a high correlation with the nugget diameter D1, and a second program for extracting feature parameters having a high correlation with the nugget diameter D1 using the correlation coefficient.

[0057] In step S53 (FIG. 10), when the pre-sorting feature data DF1 is input to the second learning model M2, the feature selection unit 240 uses the first program to calculate the correlation coefficient between the nugget diameter D1 and each feature parameter, as shown in FIG. 11. The correlation coefficient here is a value calculated based on each feature included in the pre-sorting feature data DF1, and is a value indicating the degree of correlation between the nugget diameter D1 and the feature parameter. In other words, the correlation coefficient is an index indicating whether or not a difference in feature between samples showing large and small nugget diameters D1 appears significantly when the feature values ​​included in the pre-sorting feature data DF1 for an arbitrary welding parameter (e.g., resistance value) are compared for each feature parameter. Therefore, a feature parameter with a relatively large correlation coefficient has a high correlation with the nugget diameter D1. In contrast, a feature parameter with a relatively small correlation coefficient has a low correlation with the nugget diameter D1.

[0058] In the example shown in Fig. 11, the correlation coefficient of feature parameter B is larger than the correlation coefficient of feature parameter D. In other words, when the pre-sorting feature amount data DF1 shown in Fig. 9 is used, feature parameter B has a higher correlation with the size of the nugget diameter D1 than feature parameter D. Therefore, by using feature parameter B rather than feature parameter D to create regression equations F1 to F3 (Fig. 2) in a later process, it is possible to create regression equations F1 to F3 with higher accuracy.

[0059] Furthermore, the feature amount selection unit 240 uses a second program to extract feature parameters with large correlation coefficients. For example, the feature amount selection unit 240 compares the correlation coefficient with a preset threshold and extracts feature parameters with correlation coefficients equal to or greater than the threshold, or extracts the top M feature parameters with large correlation coefficients so that a predetermined number of feature parameters are extracted.

[0060] Then, the feature amount selection unit 240 assigns a selection flag to the extracted feature parameter. The selection flag is an identification flag for identifying the feature parameter with a large correlation coefficient extracted by the feature amount selection unit 240. The feature amount selection unit 240 assigns a selection flag of, for example, "TRUE" to the extracted feature parameter with a large correlation coefficient. In contrast, the feature amount selection unit 240 assigns a selection flag of, for example, "FALSE" to the feature parameter whose selection flag is other than "TRUE". Note that the selection flag is not limited to this. The selection flag may be, for example, a numerical value or a symbol. When the selection flag is a numerical value, the feature amount selection unit 240 may assign, for example, "1" to the feature parameter with a large correlation coefficient and "0" to the feature parameter whose selection flag is other than "1".

[0061] In this way, in step S55 (FIG. 10), the feature selection unit 240 generates the post-selection feature data DF2 as shown in FIG 11, which is data obtained by selecting feature parameters according to predetermined conditions. That is, steps S53 and S55 shown in FIG 10 are processes for automatically selecting feature parameters having a high correlation with the nugget diameter D1 by inputting the pre-selection feature data DF1 into the algorithm of the existing learning model M2.

[0062] In this embodiment, the feature selection unit 240 selects feature parameters having a high correlation with the nugget diameter D1 from data related to multiple types of feature parameters present in each of multiple welding parameters included in the pre-selection feature data DF1 (FIG. 9). In other words, when the post-selection feature data DF2 is input in the subsequent regression equation creation process (step S7), one regression equation F1-F3 is created for each regression model MR1-MR3, not for each welding parameter. Furthermore, the subsequent process is executed using the top three feature parameters B, F, and N having the largest correlation coefficients from the calculated correlation coefficients.

[0063] For example, Fig. 11 shows a case where, among a plurality of welding parameters and feature parameters, three feature parameters B, F, and N related to resistance values ​​are determined to have the highest correlation with the nugget diameter D1. Therefore, the feature amount selection unit 240 extracts the feature parameters B, F, and N as the top three feature parameters having the highest correlation coefficients, and assigns a selection flag of "TRUE" to these. Furthermore, the feature amount selection unit 240 assigns a selection flag of "FALSE" to feature parameters other than the feature parameters B, F, and N.

[0064] In addition, when the regression equations F1 to F3 are created for each welding parameter, the extraction of the characteristic parameters and the assignment of the selection flag using the second program are executed for each welding parameter. That is, the characteristic parameters that are highly correlated with the nugget diameter D1 among the resistance values ​​and the characteristic parameters that are highly correlated with the nugget diameter D1 among the pressing forces are extracted. In this case, the regression equations F1 to F3 are created for each welding parameter.

[0065] The program constituting the second learning model M2 may be a program other than an existing program created by Python. The second learning model M2 only needs to be able to select feature parameters that are highly correlated with the nugget diameter D1 when the pre-sorting feature data DF1 (FIG. 9) is input. The method of selecting feature parameters is not limited to this.

[0066] As shown in FIG. 10, step S57 is executed after step S55. FIG. 12 is a diagram showing a part of the input feature amount data DF3. In step S57 (FIG. 10), the feature amount selection unit 240 processes the selected feature amount data DF2 to generate the input feature amount data DF3. The input feature amount data DF3 shown in FIG. 12 is data obtained by linking the feature amount related to the feature parameter selected in step S55 (FIG. 10) and the nugget diameter D1 for each sample out of the selected feature amount data DF2 shown in FIG. 11. That is, in step S57, the feature amount selection unit 240 generates the input feature amount data DF3 by extracting only data related to the feature parameter to which the selection flag of "TRUE" shown in FIG. 11 is attached. The selection process (step S5) is completed by executing each process (steps S51 to S57) up to step S57 shown in FIG. 10.

[0067] As shown in FIG. 3, after the selection step (step S5), a regression equation creation step (step S7) is executed. FIG. 13 is a flowchart showing the details of the regression equation creation step (step S7). The regression equation creation step (step S7) is a step of creating regression equations F1 to F3 by inputting the feature amounts related to the feature parameters selected in the selection step (step S5) and the corresponding nugget diameter D1 for each of the multiple regression models MR1 ​​to MR3. The regression equations F1 to F3 referred to here are relational expressions for estimating the nugget diameter D1, in which the feature amounts related to the feature parameters selected in the selection step (step S5) are used as explanatory variables and the nugget diameter D1 is used as a response variable, and are regression equations derived by multiple regression analysis, for example. The multiple regression equations F1 to F3 are created by machine learning for each of the multiple regression models MR1 ​​to MR3.

[0068] In step S71, the regression equation creation unit 250 generates a training data set and a test data set using the input feature amount data DF3 (FIG. 12). The training data set is a data set input to each of the multiple regression models MR1 ​​to MR3 in order to create multiple regression equations F1 to F3. The test data set is a data set used to evaluate the accuracy of the regression equations F1 to F3 created using the training data set after the multiple regression equations F1 to F3 are created. For example, when the input feature amount data DF3 includes data related to 100 samples, the data is divided as follows. Of the data related to the 100 samples, data related to 75 samples is set as the training data set, and data related to the remaining 25 samples is set as the test data set, thereby generating the training data set and the test data set. Note that the number and type of data included in the training data set and the test data set are not limited to this.

[0069] The regression equation creation unit 250 generates a test dataset by, for example, randomly selecting samples from all samples included in the input feature data DF3. In this case, the regression equation creation unit 250 generates, as a training dataset, data related to samples in the input feature data DF3 that were not used in the test dataset. In addition, the regression equation creation unit 250 may create multiple combinations of training datasets and test datasets by K-fold cross-validation.

[0070] After step S71, step S73 is executed. In step S73, the regression equation creation unit 250 inputs a learning data set to each of the multiple regression models MR1 ​​to MR3. Then, in steps S751 to S753, the regression equation creation unit 250 creates and outputs multiple regression equations F1 to F3 using each of the multiple regression models MR1 ​​to MR3. In this embodiment, the regression equation creation unit 250 inputs a learning data set to each of the first regression model MR1, the second regression model MR2, and the third regression model MR3.

[0071] The first regression model MR1, the second regression model MR2, and the third regression model MR3 are all machine learning models M, and programs for creating regression equations F1 to F3 from a learning data set by different methods are stored. Specifically, the first regression model MR1, the second regression model MR2, and the third regression model MR3 have different algorithms for regression analysis performed on the learning data set. The algorithms of the multiple regression models MR1 ​​to MR3 are, for example, Ridge regression, Lasso regression, and Elastic Net. As a result, in steps S751 to S753, the regression equation creation unit 250 can use the same learning data set to express the relationship between the feature amount related to the selected feature parameter and the nugget diameter D1 by multiple different regression equations F1 to F3. Note that the algorithms of the regression models MR1 ​​to MR3 are not limited to this.

[0072] In this embodiment, in step S751, the regression equation creation unit 250 creates the first regression equation F1 by determining the coefficient of the first regression equation F1 through machine learning using the first regression model MR1. In step S752, the regression equation creation unit 250 creates the second regression equation F2 by determining the coefficient of the second regression equation F2 through machine learning using the second regression model MR2. In step S753, the regression equation creation unit 250 creates the third regression equation F3 by determining the coefficient of the third regression equation F3 through machine learning using the third regression model MR3. Note that, after step S73, any of step S751, step S752, and step S753 may be executed first, or may be executed simultaneously in parallel. In addition, the number of regression equations F1 to F3 to be created is not limited to this. The regression equation creation process (step S7) is completed by executing each process (step S71 to step S753) up to step S753.

[0073] As shown in Fig. 3, after the regression equation creation step (step S7), a regression equation selection step (step S9) is executed. Fig. 14 is a flowchart showing the details of the regression equation selection step (step S9). The regression equation selection step (step S9) is a step of selecting regression equations F1 to F3 having a high correlation with the nugget diameter D1 from among multiple regression equations F1 to F3 that have been machine-learned.

[0074] In step S91, the regression equation selection unit 260 calculates a contribution rate for each of the multiple regression equations F1 to F3 created in the regression equation creation step (step S7). The contribution rate is a value calculated to evaluate the accuracy of the created regression equations F1 to F3. The accuracy of the regression equations F1 to F3 referred to here indicates the magnitude of correlation with the nugget diameter D1, and the higher the correlation with the nugget diameter D1, the higher the accuracy of the regression equations F1 to F3. The contribution rate indicates the proportion of the total variation of the objective variable that can be explained by the explanatory variables. In other words, the contribution rate is an index that indicates the degree of deviation between the actual measurement value of the nugget diameter D1 and the estimated value of the nugget diameter D1 calculated using the created regression equations F1 to F3. In other words, the regression equations F1 to F3 with a large contribution rate have a high accuracy in estimating the nugget diameter D1. The regression equations F1 to F3 with a small contribution rate have a low accuracy in estimating the nugget diameter D1.

[0075] Fig. 15 is a diagram showing the contribution rate for each of the regression equations F1 to F3. In this embodiment, the regression equation selection unit 260 calculates two types of contribution rates, that is, a contribution rate related to the learning dataset and a contribution rate related to the test dataset, for each of the first regression equation F1, the second regression equation F2, and the third regression equation F3. In Fig. 15, for each of the regression equations F1 to F3, the contribution rate related to the learning dataset is shown in the upper row, and the contribution rate related to the test dataset is shown in the lower row.

[0076] The contribution rate of the test data set is calculated by inputting the test data set to each of the multiple regression equations F1 to F3 that have been created. The contribution rate of the test data set is calculated to evaluate the accuracy of the regression equations F1 to F3. The contribution rate of the test data set expresses the accuracy of the regression equations F1 to F3 using data that has not been used to create the regression equations F1 to F3 and in which the nugget diameter D1 as the objective variable is known by actual measurement. In other words, the larger the contribution rate of the test data set, the more likely it is that the nugget diameter D1 can be estimated with high accuracy even from data other than the learning data set (hereinafter, unknown data).

[0077] The contribution rate of the training dataset is calculated by inputting the training dataset used in machine learning again for each of the multiple regression equations F1 to F3 created. The contribution rate of the training dataset is calculated to confirm the credibility of the calculated contribution rate of the test dataset. Specifically, the contribution rate of the training dataset is a value calculated by inputting the same data as the training dataset used in machine learning into the regression equations F1 to F3. Therefore, if the regression equations F1 to F3 are created without any problem in the regression equation creation step (step S7), it is unlikely that the contribution rate of the training dataset will be low. A case where the contribution rate of the training dataset is low refers to a case where the contribution rate is less than 0.900, for example.

[0078] Even if the contribution rate of the test data set is large, if the contribution rate of the learning data set is low, the following possibility may occur. Specifically, since the data included in the test data set coincidentally indicates values ​​close to the tendency represented by the regression equations F1 to F3, the regression equations F1 to F3 may be erroneously determined to be highly accurate, whereas the regression equations F1 to F3 are actually less accurate. Therefore, in this embodiment, the contribution rate of the test data set and the contribution rate of the learning data set are compared to select one regression equation F1 having the highest correlation with the nugget diameter D1 from among the multiple regression equations F1 to F3.

[0079] 14, after step S91, step S93 is executed. In step S93, the regression equation selection unit 260 compares the contribution rate calculated for each of the regression equations F1 to F3 with the contribution rate calculated for the test data set and the contribution rate calculated for the learning data set.

[0080] In the example shown in FIG. 15, the contribution ratio of the learning data set is approximately 0.950 in all of the regression equations F1 to F3, and the credibility of the contribution ratio of the test data set is considered to be satisfactory.

[0081] On the other hand, the first regression equation F1 has the largest contribution rate for the test data set among the three regression equations F1 to F3. At this time, a significant difference is observed between the contribution rate in the first regression equation F1 and the contribution rates in the second regression equation F2 and the third regression equation F3. Therefore, in step S95 (FIG. 14), the regression equation selection unit 260 selects the first regression equation F1 as the regression equation F1 having a high correlation with the nugget diameter D1 from among the multiple regression equations F1 to F3. Note that the regression equation selection unit 260 may select the regression equations F1 to F3 having a high correlation with the nugget diameter D1 by a method other than the contribution rate. The regression equation selection process (step S9) is completed by executing each process (steps S91 to S95) up to step S95.

[0082] According to the first embodiment, as shown in FIG. 9, a feature value for each of a plurality of welding parameters is calculated by a trained first learning model M1 in which a plurality of welding parameters, which are extraction items for expressing the features of the welding waveform D2, are set in advance. Then, as shown in FIG. 11, a feature parameter having a high correlation with the nugget diameter D1 is selected by a trained second learning model M2 in which a program for evaluating the correlation between the feature value and the nugget diameter D1 for each feature parameter is set in advance. That is, the learning models M1 and M2 in which extraction items for extracting elements having a high correlation with the nugget diameter D1 from various features included in the welding waveform D2 are set in advance can select the feature parameters and feature values ​​to be used as explanatory variables of the regression equations F1 to F3. Therefore, the feature parameters and feature values ​​having a high correlation with the nugget diameter D1 can be selected without using a learning method such as deep learning. Therefore, even if the number of teacher data included in the data set D is small, the accuracy of selecting the feature parameters can be suppressed from decreasing. As a result, the regression equations F1 to F3 with high accuracy can be created and selected without requiring the man-hours of preparing a large amount of data.

[0083] According to the first embodiment, the regression equations F1 to D3 are created using a data set D including training data in which the nugget diameter D1 (FIG. 7) and the welding waveform D2 (FIGS. 5 and 6) when spot welding is performed while changing the disturbance conditions are linked to each disturbance condition. In other words, the data set D includes data with different tendencies due to differences in the disturbance conditions, and the training data can be given a range. This makes it possible to prevent a decrease in accuracy in selecting the characteristic parameters even when the amount of training data included in the data set D is small. Therefore, it is possible to select the regression equations F1 to F3 with higher accuracy without requiring the effort of preparing a large amount of data.

[0084] Moreover, according to the first embodiment, as shown in Fig. 7, data relating to samples in which the nugget diameter D1 is varied by changing the disturbance conditions is used as teacher data. This reduces the possibility that regression equations F1 to F3 that are suitable only for a specific nugget diameter D1 are created. In other words, overlearning can be prevented, and regression equations F1 to F3 that are highly versatile in estimating the nugget diameter D1 can be created. Therefore, it is possible to increase the possibility that the nugget diameter D1 can be accurately estimated even for unknown data.

[0085] According to the first embodiment, as shown in FIG. 11, the feature parameters and the feature quantities corresponding to the feature parameters are input to the trained second learning model M2, and a correlation coefficient is calculated for each of the multiple welding parameters. Then, the feature parameters having a correlation coefficient equal to or greater than a predetermined standard are selected as feature parameters having a high correlation with the nugget diameter D1. That is, according to the first embodiment, the correlation coefficients are compared to select feature parameters having a high correlation with the nugget diameter D1. This allows a person to understand the reasons for selecting the feature parameters and feature quantities and the elements that constitute the explanatory variables.

[0086] According to the first embodiment, as shown in FIG. 11, a selection flag is attached to a feature parameter whose correlation coefficient is equal to or greater than a predetermined standard. This allows a human to confirm the selected feature parameters and feature quantities. In other words, a human can understand the feature parameters and feature quantities used as explanatory variables of the regression equations F1 to F3. Therefore, when a defect occurs in the regression equations F1 to F3, a human can easily understand the parts to be corrected and the correction method.

[0087] According to the first embodiment, the teaching data uses the welding waveform D2 that has been pre-processed to adjust the data area, etc. This makes it possible to calculate the feature amount using the welding waveform D2 that has been processed into a state suitable for analysis.

[0088] According to the first embodiment, as shown in Fig. 9, the first learning model M1 is a trained model configured by an existing program, and it is possible to generate feature quantities related to a large number of feature parameters at once. This makes it possible to reduce the time and man-hours required for selecting feature parameters and feature quantities.

[0089] According to the first embodiment, as shown in Fig. 11, the second learning model M2 configured by an existing program can easily select characteristic parameters having a high correlation with the nugget diameter D1 from a large number of characteristic parameters. This can further reduce the time and labor required to select the characteristic parameters and characteristic amounts.

[0090] According to the first embodiment, as shown in Fig. 13, the input feature data DF3 is divided into a learning data set used to create the regression equations F1 to F3 and a test data set used to evaluate the created regression equations F1 to F3. The learning data set is then input to a plurality of regression models MR1 ​​to MR3. This allows a plurality of different regression equations F1 to F3 for estimating the nugget diameter D1 to be created from the same input feature data DF3.

[0091] According to the first embodiment, as shown in Fig. 14, the contribution ratios of the test data set are calculated for the created regression equations F1 to F3. This makes it possible to select an optimal regression equation F1 having a high correlation with the nugget diameter D1 by comparing the contribution ratios of the regression equations F1 to F3. Therefore, even when the same input feature data DF3 is used, it is possible to select a regression equation F1 to F3 that is more suitable for estimating the nugget diameter D1 from among the multiple regression equations F1 to F3.

[0092] According to the first embodiment, as shown in Fig. 15, both the contribution rate related to the test data set and the contribution rate related to the learning data set are calculated for the created regression equations F1 to F3. This makes it possible to confirm the reliability of the calculated contribution rate related to the test data set. Therefore, it is possible to select the regression equations F1 to F3 that are more suitable for estimating the nugget diameter D1.

[0093] According to the first embodiment, characteristic parameters having a high correlation with the nugget diameter D1 are selected from among all characteristic parameters existing in each of the plurality of welding parameters, thereby making it possible to create regression equations F1 to F3 that uniquely represent the welding waveform D2 existing in each of the plurality of welding parameters.

[0094] B. Second embodiment: FIG. 16 is a flowchart showing details of the selection step (step S5) in the second embodiment. The spot welding device 10 that executes the estimation method of this embodiment is the same as that of the first embodiment (FIGS. 1 and 2). In addition, the control device 100 creates regression equations F1 to F3 for estimating the nugget diameter D1 in spot welding, and selects the optimal regression equation F1, as in the first embodiment. In this embodiment, a part of the process in the selection step (step S5) is different from that of the first embodiment (FIG. 10). As a result, a part of the data input to the multiple regression models MR1 ​​to MR3 in the regression equation creation step (step S7) is different from that of the first embodiment (FIG. 12). Therefore, the acquisition step (step S1), feature amount generation step (step S3), and regression equation selection step (step S9) shown in FIG. 3 are the same as those of the first embodiment. Before the selection step (step S5) of this embodiment is performed, the acquisition step (step S1) and feature amount generation step (step S3) shown in FIG. 3 are performed in this order. In addition, after the selection step (step S5) and the regression equation creation step (step S7) of this embodiment are performed, a regression equation selection step (step S9) is performed. Note that the same steps as those in the first embodiment are denoted by the same reference numerals and the description thereof is omitted.

[0095] 16, step S51, step S53, step S55, and step S57 are executed in this order. Each process from step S51 to step S57 is the same as that in the first embodiment.

[0096] After step S57, step S59 is executed. In step S59, the feature selection unit 240 processes the input feature data DF3 (FIG. 12) to generate processed input feature data. The processed input feature data is data obtained by removing, from the features included in the input feature data DF3, features that indicate abnormal values.

[0097] FIG. 17 is a graph showing the relationship between the feature amount related to the selected feature parameter and the nugget diameter D1. In this embodiment, the top three feature parameters B, F, and N selected in the first embodiment as feature parameters having a high correlation with the nugget diameter D1 will be described as an example. A first graph G1 shown on the left side of FIG. 17 is a graph showing the relationship between the feature amount related to the feature parameter B and the nugget diameter D1. A second graph G2 shown in the center of FIG. 17 is a graph showing the relationship between the feature amount related to the feature parameter F and the nugget diameter D1. A third graph G3 shown on the right side of FIG. 17 is a graph showing the relationship between the feature amount related to the feature parameter N and the nugget diameter D1. That is, each of the graphs G1 to G3 is a graph showing the feature for the same welding waveform D2 for each feature parameter.

[0098] Graphs G1 to G3 for three lower electrodes shown in Fig. 17 are graphs in which, for example, input feature amount data DF3, which has 13 samples with varying disturbance conditions, is plotted for each of the feature parameters B, F, and N, with the nugget diameter D1 on the vertical axis and the feature amount on the horizontal axis. Note that, although 13 plots are shown in each of the graphs G1 to G3 shown in Fig. 17, Fig. 17 merely shows a schematic diagram of the relationship between the feature amounts related to the feature parameters B, N, and F and the nugget diameter D1, and the number and values ​​of data are not limited to this.

[0099] With reference to the graphs G1-G3, the characteristic parameters B, F, and N can be roughly divided into large diameter sample groups Da1-Da3 and small diameter sample groups Db1-Db3 according to the size of the nugget diameter D1. The large diameter sample groups Da1-Da3 are, for example, plot groups of samples with large nugget diameters D1 that are equal to or larger than a predetermined first threshold value T1. The small diameter sample groups Db1-Db3 are, for example, plot groups of samples with small nugget diameters D1 that are less than the first threshold value T1.

[0100] Referring to the first graph G1, for the characteristic parameter B, a certain correlation is seen between the nugget diameter D1 and the characteristic amount in the large diameter sample group Da1. Also, referring to the second graph G2, for the characteristic parameter F, a certain correlation is seen between the nugget diameter D1 and the characteristic amount in the large diameter sample group Da2. In particular, for these two characteristic parameters B and F, the plots for the large diameter sample groups Da1 and Da2 show a nugget diameter D1 equal to or greater than the second threshold value T2, which is greater than the first threshold value T1. Therefore, these two characteristic parameters B and F are particularly effective for selecting characteristic parameters when the nugget diameter D1 is equal to or greater than the second threshold value T2.

[0101] At this time, when the nugget diameter D1 is extremely small, as in the small diameter sample groups Db1 to Db3, there is a high possibility that the nugget has not grown and the metal sheets W1, W2 (FIG. 1) have not been joined. Therefore, when data relating to samples with extremely small nugget diameters D1 are included in the explanatory variables of the regression equations F1 to F3, the accuracy of the regression equations F1 to F3 may decrease. Therefore, it is preferable to remove data relating to the small diameter sample groups Db1 to Db3 with extremely small nugget diameters D1 as abnormal values ​​before inputting them into the regression models MR1 ​​to MR3.

[0102] Referring to the third graph G3, the large diameter sample group Da3 and the small diameter sample group Db3 show different tendencies for the characteristic parameter N depending on the size of the nugget diameter D1. Specifically, all plots for the large diameter sample group Da3 show extremely small values ​​for the characteristic parameter N compared to the small diameter sample group Db3. In other words, by focusing on the feature amount related to the characteristic parameter N, data related to samples with an extremely small nugget diameter D1 can be extracted.

[0103] When the small diameter sample group Db3 is extracted as an abnormal value using the feature parameter N, for example, as shown in the third graph G3, a third threshold T3 is set. The third threshold T3 is set between the feature amount related to the large diameter sample group Da3 and the feature amount related to the small diameter sample group Db3. As a result, in step S59, for example, the feature amount selection unit 240 extracts a feature amount that is equal to or greater than the third threshold T3 for the feature parameter N as an abnormal value. At this time, for example, the feature amount selection unit 240 assigns "1" to sample data whose feature amount is equal to or greater than the third threshold T3, and assigns "0" to sample data whose feature amount is less than the third threshold T3. Then, the feature amount selection unit 240 determines that the data assigned with "1" is an abnormal value. As a result, the feature amount selection unit 240 generates processed input feature amount data by removing data determined to be an abnormal value from the input feature amount data DF3 shown in FIG. 12.

[0104] The selection step (step S5) in this embodiment is completed by executing each step (step S51 to step S59) up to step S59 shown in FIG. 16. As shown in FIG. 3, after the selection step (step S5), a regression equation creation step (step S7) is executed. At this time, as shown in FIG. 13, in step S71, the regression equation creation unit 250 inputs the processed input feature data generated in step S59 to each of the multiple regression models MR1 ​​to MR3. That is, in this embodiment, regression equations F1 to F3 are created using data from which data indicating abnormal values ​​has been removed. The other steps (steps S73 to S753) are the same as those in the first embodiment, and therefore will not be described.

[0105] According to the second embodiment, as shown in Fig. 16, in step S59, data relating to a sample having an extremely small nugget diameter D1 is removed as an abnormal value using a feature parameter N indicating a different tendency of the feature depending on the size of the nugget diameter D1. This generates processed input feature data. Then, the regression equations F1 to F3 are created using the processed input feature data. This reduces the possibility that the accuracy of the created regression equations F1 to F3 will decrease.

[0106] C. Other embodiments: C-1. Alternative embodiment 1: The teacher data may be a waveform obtained by adding together welding waveforms D2 associated with a plurality of welding parameters (hereinafter, a composite waveform) or a waveform obtained by multiplying welding waveform D2 by a predetermined coefficient (hereinafter, a processed waveform). The teacher data may also include data related to samples created by changing welding conditions such as current values. When generating a composite waveform, for example, a pre-processing is performed to make the data areas of the welding waveforms D2 to be added the same, and then a welding waveform D2 associated with a first welding parameter and a welding waveform D2 associated with a second welding parameter are added together. The composite waveform may be a waveform obtained by adding together welding waveforms D2 associated with three or more welding parameters. The composite waveform or the processed waveform is used when the characteristics of the welding waveform D2 are more prominently expressed by using them as teacher data. In such a form, by using teacher data in which the welding waveform D2 is pre-processed, it is possible to further easily extract characteristic parameters and characteristic quantities that are highly correlated with the nugget diameter D1.

[0107] C-2. Alternative embodiment 2: In the above embodiment, as shown in FIG. 14, the contribution ratios calculated for the multiple regression equations F1 to F3 are compared to select the regression equation F1 that is optimal for estimating the nugget diameter D1. However, the method of selecting the optimal regression equation F1 is not limited to this. The optimal regression equation F1 may be selected, for example, by calculating a residual for the multiple regression equations F1 to F3 and comparing the calculated residual. The residual here refers to the numerical difference between the actual value of the nugget diameter D1 and the estimated value of the nugget diameter D1 calculated using the multiple regression equations F1 to F3 that have been created. Even in this form, it is possible to select the regression equation F1 with high estimation accuracy of the nugget diameter D1 from the multiple regression equations F1 to F3. The optimal regression equation F1 may be selected using both the contribution ratio and the residual, or may be selected by another method.

[0108] The present disclosure is not limited to the above-mentioned embodiment, and can be realized in various configurations without departing from the spirit of the present disclosure. For example, the technical features of the embodiments corresponding to the technical features in each aspect described in the Summary of the Invention can be appropriately replaced or combined to solve some or all of the above-mentioned problems or to achieve some or all of the above-mentioned effects. Furthermore, if a technical feature is not described as essential in this specification, it can be appropriately deleted. [Explanation of symbols]

[0109] 1...gun body, 1a...upper part, 1b...lower part, 2...upper electrode, 3...lower electrode, 4...electrode lifting device, 5...current adjustment device, 9...measurement mechanism, 10...spot welding device, 20...CPU, 30...communication unit, 40...display, 41...servo motor, 42...lifting member, 50...input operation unit, 60...storage unit, 91...current measurement unit, 92...voltage measurement unit, 93...resistance value calculation unit, 94...pressure measurement unit, 95...electrode displacement measurement unit, 100...control device, 210...operation control unit, 220...acquisition unit, 230...feature generation unit, 240...feature selection unit, 250...regression equation creation unit, 260...regression equation selection unit, A to Q, a to f...feature parameters, D...data set, D1...nugget diameter, D 2...Welding waveform, DF1...Feature amount data before sorting, DF2...Feature amount data after sorting, DF3...Feature amount data for input, Da1, Da2, Da3...Large diameter sample group, Db1, Db2, Db3...Small diameter sample group, F1...First regression equation, F2...Second regression equation, F3...Third regression equation, G...Spot welding gun, G1...First graph, G2...Second graph, G3...Third graph, M...Machine learning model, M1...First learning model, M2...Second learning model, MR1...First regression model, MR2...Second regression model, MR3...Third regression model, RA...Robot arm, T1...First threshold, T2...Second threshold, T3...Third threshold, W...Welded material, W1, W2...Metal plate, W1a, W2a...End

Claims

[Claim 1] A method for estimating a nugget diameter in spot welding, comprising: an acquisition step of acquiring a data set including teacher data in which the nugget diameter when the spot welding is performed while changing a disturbance condition that may affect the size of the nugget diameter and a welding waveform that is time-series data of welding parameters in the spot welding are associated with each disturbance condition; a feature generating step of generating values ​​of a plurality of predetermined feature parameters as feature quantities of the welding waveform by inputting the data set into a trained first learning model; a selection process for selecting the feature parameters having a high correlation with the nugget diameter by inputting the feature parameters and the feature amounts corresponding to the feature parameters into a trained second learning model; a regression equation creation process for creating, by machine learning, a regression equation for each of a plurality of regression models, which are a plurality of machine learning models, by inputting the feature amounts related to the selected feature parameters into the respective regression models, the feature amounts related to the selected feature parameters being explanatory variables and the nugget diameter being an objective variable; and selecting a regression equation having a high correlation with the nugget diameter from the plurality of regression equations trained by machine learning.

Images