Skill analysis system and skill analysis method
The skill analysis system improves analysis accuracy by excluding low-quality data and focusing on skilled workers' data, enabling precise identification of quality-influencing feature quantities, thus enhancing training and quality control.
Patent Information
- Application Number
- PCT/JP2024/000245
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-31
- Filing Date
- 2024-01-10
- Publication Date
- 2024-12-05
AI Technical Summary
Existing skill analysis systems face limitations in analysis accuracy when extracting feature quantities that affect product quality, as they often include noise data from workers with low skills, which decreases the reliability of regression analysis results.
A skill analysis system that performs regression analysis using measurement data from workers' tasks as explanatory variables and quality data as objective variables, with a processing unit that excludes lower quality outlier data to improve analysis accuracy by focusing on high-quality data from skilled workers, and displays important feature quantities that impact product quality.
Enhances analysis accuracy by excluding noise data from low-skilled workers, allowing for more precise identification of feature quantities that influence product quality, thereby formalizing tacit knowledge and supporting quality control and training.
Smart Images

Figure JP2024000245_05122024_PF_FP_ABST
Abstract
Description
Skill analysis system and skill analysis method
[0001] The present disclosure relates to a skill analysis system and the like.
[0002] As a technique for extracting feature quantities that affect the quality of a predetermined work object, for example, the technique described in Patent Document 1 is known. That is, Patent Document 1 describes that "from among the representative feature quantities, a representative feature quantity that contributes to the variation of the product quality data is identified as a candidate quality influence factor."
[0003] Japanese Patent Application Laid-Open No. 2021-192137
[0004] The technology described in Patent Document 1 extracts feature quantities that are likely to affect the quality of a product (work object), but there is room for improvement in terms of analytical accuracy.
[0005] Therefore, an object of the present disclosure is to provide a skill analysis system or the like that aims to improve analysis accuracy.
[0006] In order to solve the above-mentioned problems, the skill analysis system according to the present disclosure includes a processing unit that performs a regression analysis using feature quantities of measurement data relating to a worker's work on a work object as explanatory variables and quality data of the work object as a target variable, and displays the results of extraction of important feature quantities based on the regression analysis on a display device, and the processing unit excludes outliers on the low quality side of the quality data and the measurement data corresponding to the quality data of the outliers from the regression analysis.
[0007] According to the present disclosure, a skill analysis system and the like that aims to improve analysis accuracy can be provided.
[0008] 1 is an explanatory diagram showing an example of an analysis target of the skill analysis system according to the first embodiment. FIG. 2 is a functional block diagram of the skill analysis system according to the first embodiment. FIG. 3 is a flowchart of processing executed by a processing unit of the skill analysis system according to the first embodiment. FIG. 4 is a box plot related to the exclusion of quality data of lower outliers in the skill analysis system according to the first embodiment. FIG. 5 is a display example related to threshold setting for lower outliers in the skill analysis system according to the first embodiment. FIG. 6 is an explanatory diagram related to correlation data of the skill analysis system according to the first embodiment. FIG. 7 is an explanatory diagram related to multiple regression analysis, which is an example of regression analysis in the skill analysis system according to the first embodiment. FIG. 8 is a display example of analysis results of the skill analysis system according to the first embodiment. FIG. 9 is a flowchart of processing executed by a processing unit of the skill analysis system according to the second embodiment. FIG. 10 is an explanatory diagram related to selection of feature quantities to be subjected to regression analysis in the skill analysis system according to the second embodiment. FIG. 11 is a functional block diagram of the skill analysis system according to the third embodiment. FIG. 12 is a flowchart of processing executed by a processing unit of the skill analysis system according to the third embodiment. FIG. 13 is a display example of a selection screen for a model used in regression analysis in the skill analysis system according to the third embodiment. FIG. 14 is an explanatory diagram related to extraction of important feature quantities when multiple types of models are used in the skill analysis system according to the third embodiment.
[0009] First Embodiment In the following, as an example, a case where an analysis is performed based on quality data and measurement data related to welding work performed by a worker M1 (see FIG. 1 ) will be described, but the type of work to be analyzed is not limited to welding. For example, various work such as painting, casting, forging, cutting, grinding, precision machining, plating, sheet metal processing, mold finishing, and lens polishing can be analyzed.
[0010] FIG. 1 is an explanatory diagram showing an example of an analysis target of the skill analysis system according to the first embodiment. FIG. 1 illustrates a state in which a worker M1 is welding a workpiece W1. The worker M1 uses a welding torch T1 connected to a welding power source P1 to perform arc welding with the tip of a welding wire (not shown) close to the workpiece W1. The welding wire is fed from a wire feeder (not shown). Images of the worker M1 welding are captured by cameras C1, C2, and C3 used for motion measurement. The voltage and current values of the welding power source P1 are also detected moment by moment. The captured images and the detected voltage and current values of the cameras C1, C2, and C3 are used as measurement data related to the welding work.
[0011] After worker M1 performs a series of tasks, another worker (not shown) performs the same tasks, and predetermined measurement data related to these tasks is acquired. The number of workers performing the predetermined tasks to be analyzed may be, for example, tens, hundreds, or even thousands. It is also preferable that the workers include skilled workers, workers with average skills, and workers with low skills. This makes it easier to identify the advantages of skilled workers compared to other workers. Furthermore, since data on various workers with different skill levels is obtained, the accuracy of the regression analysis described below is improved.
[0012] 2 is a functional block diagram of a skill analysis system 10. The skill analysis system 10 analyzes work performed by a worker on a predetermined work object, extracts important feature quantities, and presents the extracted feature quantities to a user. Such a skill analysis system 10 may be configured as a single computer (e.g., a server), or may be configured as a plurality of computers (not shown) connected in a predetermined manner via a signal line or a network.
[0013] 2, the skill analysis system 10 includes a storage unit 11, an input unit 12, a processing unit 13, and a display unit 14 (display device). Although not shown, the storage unit 11 includes a non-volatile memory such as a read-only memory (ROM) and a volatile memory such as a random access memory (RAM) or a register. The storage unit 11 stores predetermined programs in advance, as well as correlation data 11a (described later). In addition, the storage unit 11 also stores measurement data and quality data input via the input unit 12 and calculation results from the processing unit 13.
[0014] The input unit 12 is an interface used to input predetermined measurement data and quality data. For example, the measurement data and quality data are input via the input unit 12 from a predetermined database (not shown).
[0015] The term "measurement data" refers to predetermined data obtained by measuring the worker's movements from moment to moment. For example, when analyzing welding work, the measurement data used may include the tilt angle and welding speed of the welding torch T1 (see FIG. 1), the welding target position, the amount of welding wire (not shown) supplied per unit time, and the voltage and current values of the welding power source P1 (see FIG. 1). Other measurement data may include the worker's movement data acquired by motion capture (not shown) and the temperature and pressure values of the workpiece. Furthermore, the measurement data may include the worker's gaze and posture, as well as biological data such as the worker's breathing, heartbeat, and brain waves.
[0016] Furthermore, "quality data" refers to data that indicates the quality of the workpiece. Examples of such quality data include the number of defects in the workpiece, the amount of deviation from a predetermined target value, the degree of irregularity in the weld bead, and predetermined scoring results. The quality data is associated with the above-mentioned multiple types of measurement data. A set of measurement data and quality data is input via the input unit 12. For example, measurement data such as operation data and current values obtained in one welding operation and the number of defects (quality data) in the workpiece after the welding operation may be input as a set. Note that the greater the amount of measurement data and quality data, the higher the analytical accuracy tends to be.
[0017] 2 is, for example, a CPU (Central Processing Unit), which reads out a program stored in the storage unit 11 and executes predetermined processing. The display unit 14 displays the processing results of the processing unit 13 in a predetermined manner. For example, a display is used as the display unit 14. Note that the display unit 14 may also have a user interface function, such as a touch display.
[0018] The processing results of the skill analysis system 10 may also be transmitted to a user's information terminal (not shown) via a network (not shown). Examples of such information terminals include smartphones, mobile phones, tablets, personal computers, and wearable devices. In this case, the user's information terminal functions as the display unit 14.
[0019] As shown in FIG. 2 , the processing unit 13 includes a lower outlier removal unit 131, a feature extraction unit 132, a preprocessing unit 133, a correlation investigation unit 134, a feature selection unit 135, an analysis unit 136, and an important feature search unit 137.
[0020] The lower outlier exclusion unit 131 excludes quality data that corresponds to lower outliers, which are of extremely low quality, and measurement data associated with this quality data from the regression analysis by the analysis unit 136. This allows the regression analysis to be performed after excluding quality data and measurement data (data that may act as noise) related to, for example, a beginner worker with extremely low skills, thereby improving the accuracy of the analysis. Details of lower outliers will be described later.
[0021] The feature extraction unit 132 extracts feature quantities from each of the multiple types of measurement data. Examples of such feature quantities include average values and standard deviations. For example, the average welding speed of a single worker performing welding work may be used as a feature quantity (similarly for other workers). Furthermore, new feature quantities may be generated based on the calculation results of multiple types of measurement data. For example, in arc welding, current [A] × voltage [V] × 60 / welding speed [cm / min] equals heat input [J / cm]. This heat input [J / cm] often has a significant impact on the quality of the welding work (i.e., the quality of the workpiece). Therefore, heat input [J / cm] may be used as one of the feature quantities.
[0022] The preprocessing unit 133 performs predetermined preprocessing on the features extracted by the feature extraction unit 132. For example, the preprocessing unit 133 performs standardization (i.e., preprocessing) by dividing the deviation of the feature from its average value by the standard deviation. This standardization makes it possible to correctly evaluate the impact on quality even when features of measurement data have different units. Note that there is no particular need to perform preprocessing such as standardization on quality data. This is because the regression analysis described below uses one target variable (quality data).
[0023] The correlation investigation unit 134 investigates the correlation between feature quantities of the measurement data. A method such as calculating a correlation coefficient is used to investigate such correlation. The threshold value used as a criterion for determining whether or not there is a strong correlation between predetermined feature quantities is not particularly limited, but for example, it may be determined that there is a strong correlation when the absolute value of the correlation coefficient is 0.7 or more. Note that clustering such as spectral clustering can also be used as a method for investigating the correlation. The investigation results by the correlation investigation unit 134 are stored in the storage unit 11 as correlation data 11a. The correlation data is data in which multiple types of feature quantities that are highly correlated are associated and grouped.
[0024] The feature selection unit 135 refers to the correlation data 11a and selects one type of feature from a group of feature values having a strong correlation to be used in the regression analysis of the analysis unit 136. The regression analysis described below analyzes the influence of explanatory variables (multiple types of measurement data) on a target variable (quality data), but if feature values that are highly correlated with each other are mixed, a problem of so-called multicollinearity occurs, and the results of the regression analysis often become unstable.
[0025] The method for selecting the feature is not particularly limited, but for example, the feature having the largest correlation coefficient with the quality data (objective variable) may be selected from among multiple types of feature belonging to one group.
[0026] The analysis unit 136 performs a predetermined regression analysis using the quality data as a response variable and multiple types of feature quantities corresponding to the measurement data as explanatory variables, and calculates the variable importance for each type of measurement data. For example, multiple regression analysis, Lasso regression, Ridge regression, and partial least squares method are used as methods for such regression analysis. Furthermore, when the number of data is large, methods such as gradient boosting decision trees and neural networks may also be used.
[0027] The "variable importance" calculated by the analysis unit 136 is a numerical value indicating the degree to which a predetermined explanatory variable (feature corresponding to measurement data) contributes to the objective variable (i.e., quality data). For example, when multiple regression analysis is performed, the t value is used as the variable importance. Furthermore, when partial least squares is performed, VIP (Variable Importance for Prediction) is used as the variable importance.
[0028] The important feature search unit 137 ranks multiple types of feature amounts in order of the degree of variable importance calculated by the analysis unit 136. This makes it possible to identify what types of feature amounts contribute to the quality of the work object. The search results by the important feature search unit 137 are displayed as important feature amounts on the display unit 14 in a predetermined manner.
[0029] Furthermore, among the important feature quantities identified based on the results of the regression analysis, those with which there are other feature quantities that are strongly correlated are displayed in association with the other feature quantities on the display unit 14. As described above, multiple types of feature quantities that are strongly correlated are grouped and stored as correlation data 11a in the storage unit 11. In this way, by also displaying other feature quantities that are strongly correlated with the important feature quantities, the user can fully grasp which feature quantities affect the quality of the workpiece when performing work such as welding.
[0030] 3 is a flowchart of the processing executed by the processing unit of the skill analysis system (see also FIG. 2 as appropriate). In step S101, the processing unit 13 accepts input of measurement data and quality data. As described above, the measurement data and quality data are input from a predetermined database (not shown) via the input unit 12. The measurement data includes multiple types of data, such as the angle of the welding torch T1 (see FIG. 1) and the welding speed. The quality data includes one predetermined type of data, such as the number of defects in the workpiece.
[0031] In step S102, the processing unit 13 uses the lower outlier exclusion unit 131 to exclude from the regression analysis any quality data that falls under the lower outlier category. Note that the "lower" in the phrase "lower outlier" means low quality. In other words, the processing unit 13 excludes from the regression analysis any quality data of the work object that falls under the lower quality outlier category, as well as the measurement data corresponding to the quality data of the outliers (outlier exclusion step).
[0032] Fig. 4 is a box plot showing the elimination of quality data of lower outliers. In the example of Fig. 4, the higher the quality of the work object after a predetermined operation, the larger the value of the quality data. The box plot shown in Fig. 4 is a statistical diagram showing the degree of dispersion of the data, and the minimum value Q min and the first quartile Q 1 and the median Q 2 and the third quartile Q 3 and the maximum value Q Max Here, the first quartile Q 1 The third quartile Q is the value of the data (i.e., quality data) when the number of data from the smallest value reaches 25% when a predetermined data group is sorted in ascending order of value. 3 The third quartile Q is the value of the data (i.e., quality data) when the number of data from the smallest value reaches 75% when a given data group is sorted in ascending order of value. 3 From the first quartile Q 1 The value obtained by subtracting (Q 3 -Q 1 ) is called the interquartile range.
[0033] For example, when a beginner performs welding work, the quality of the weld may be extremely low. In such a case, the measurement data may act as noise during regression analysis, which may lead to a decrease in the accuracy of the regression analysis. Therefore, in the example of Figure 4, the first quartile Q 1 to the interquartile range (Q 3 -Q 1The quality data values smaller than the value obtained by subtracting 1.5 times the average quality data (three quality data surrounded by dashed line D1) are excluded from the regression analysis.
[0034] Furthermore, measurement data corresponding to the excluded quality data are also excluded from the regression analysis. As a result, the regression analysis is performed after excluding quality data with extremely low quality, and the accuracy of the regression analysis is improved. In this way, when the quality of the work object is higher, the value of the quality data is larger, the processing unit 13 calculates the first quartile Q of the quality data. 1 From the above, the interquartile range of the quality data (Q 3 -Q 1 The value obtained by subtracting the product of the above and a predetermined value (for example, 1.5) is set as a threshold, and quality data whose value is smaller than the threshold and the measurement data corresponding to the quality data are excluded from the regression analysis.
[0035] Furthermore, for example, if the quality data is a defect rate in a work object, the lower the defect rate, the higher the quality. In this case, for example, the interquartile range (Q 3 -Q 1 ) is 1.5 times the third quartile Q 3 In other words, when the quality data value is smaller as the quality of the work object is higher, the processing unit 13 excludes the quality data value that is larger than the value obtained by adding 3 , the interquartile range of the quality data (Q 3 -Q 1 The value obtained by multiplying the value of the quality data by a predetermined value (for example, 1.5) is set as a threshold value, and quality data whose value is greater than the threshold value and the measurement data corresponding to the quality data are excluded from the regression analysis.
[0036] On the other hand, when a skilled worker performs welding, the quality of the weld may be extremely high compared to the others. In this case, there is almost no possibility that the measurement data contains noise, and there is a high possibility that the measurement data contains factors that have improved the quality of the weld. Therefore, in the example of Figure 4, the interquartile range (Q 3 -Q 1) is 1.5 times the third quartile Q 3 The quality data whose value is greater than the value obtained by adding (the two quality data surrounded by dashed line D2) is included in the regression analysis. That is, the processing unit 13 includes quality data that falls under outliers on the high quality side in the regression analysis, and the measurement data corresponding to that quality data. This makes it possible to include the measurement data and quality data of people with particularly high skills in the regression analysis.
[0037] In addition, for quality data, the threshold value that is the criterion for determining whether or not to exclude data from regression analysis is 1.5 × (Q 3 -Q 1 ) or, as will be explained next, the threshold value may be set by the user.
[0038] 5 is a display example for setting a threshold for the lower outlier. In the example of FIG. 5, a box-and-whisker plot of predetermined quality data is displayed on the display unit 14 (see FIG. 2), and a pull-down list K1 for setting a threshold for the lower outlier is also displayed. Specifically, the first quartile Q 1 to the interquartile range (Q 3 -Q 1 The user can set the "a times" part when the value obtained by subtracting a times the value of the lower outlier is used as the threshold value for the lower outlier. For example, as shown in Figure 5, when "1.7" is selected as the value of a and the enter button B1 is pressed, the value of a is set to "1.7".
[0039] In this way, the processing unit 13 displays the box-and-whisker plot of the quality data on the display unit 14 (display device), and further allows the user to adjust the magnitude of the predetermined value (the value of a) through input operations. This increases the degree of freedom for the user in setting the threshold value for the lower outlier. Note that the value of a set by the user may be greater than 1 or may be less than or equal to 1.
[0040] Returning to FIG. 3 , the description will be continued. After excluding the lower outliers of the quality data from the regression analysis in step S102, the processing unit 13 proceeds to step S103. In step S103, the processing unit 13 extracts feature quantities of the measurement data using the feature quantity extraction unit 132. That is, the processing unit 13 extracts feature quantities such as the average value and standard deviation for each of the multiple types of measurement data. Note that, like the heat input [J / cm] described above, new feature quantities may be generated by performing a predetermined calculation on multiple types of measurement data.
[0041] In step S104, the processing unit 13 performs predetermined preprocessing using the preprocessing unit 133. For example, the processing unit 13 performs standardization as preprocessing, such as dividing the deviation of the measurement data from its average value by the standard deviation. After standardization, the average value of the data becomes 0 and the standard deviation becomes 1, so that the impact on quality, etc., can be correctly evaluated even between data in different units.
[0042] In step S105, the processing unit 13 generates correlation data 11a by performing a correlation investigation between feature quantities using the correlation investigation unit 134, and stores this correlation data 11a in the storage unit 11. That is, the processing unit 13 generates correlation data 11a by associating and grouping multiple types of feature quantities that are highly correlated, and stores this correlation data 11a in the storage unit 11.
[0043] FIG. 6 is an explanatory diagram of the correlation data 11a. In the example of the correlation data 11a shown in FIG. 6, a predetermined group ID is associated with a combination of highly correlated features (i.e., a group). For example, in a group with a group ID of G0001, feature α and feature β are associated as a combination of highly correlated features. In addition, in a group with a group ID of G0002, feature γ, feature δ, and feature ε are associated as a combination of highly correlated features. In this way, three or more types of features may be included in one group.
[0044] In step S106 of FIG. 3 , the processing unit 13 uses the feature selection unit 135 to select one feature from each group of highly correlated features. This prevents regression analysis from being performed in a state where highly correlated features are mixed. As a result, the so-called multicollinearity problem can be suppressed, and regression analysis can be performed appropriately. For example, one feature is selected from each group, such as the feature α selected from group G0001, the feature γ selected from group G0002, and the feature ζ selected from group G0003 shown in FIG. 6 .
[0045] In this way, the processing unit 13 performs a predetermined correlation investigation on multiple types of feature quantities (S105 in FIG. 3), selects one type of feature quantity from each group representing combinations of feature quantities with relatively strong correlations, and includes the selected feature quantities in the regression analysis (S106).
[0046] If there are no other types of features that are strongly correlated with a given feature, the feature will exist alone without forming any particular group. Features selected one for each strongly correlated group and features that do not form any particular group are weakly correlated data (i.e., independent variables), and are therefore suitable for regression analysis.
[0047] Next, in step S107 of FIG. 3 , the processing unit 13 performs a regression analysis using the quality data as the objective variable using the analysis unit 136. The explanatory variables used in the regression analysis are the feature quantities selected in step S106 and feature quantities that do not form any particular group. By performing the regression analysis in this manner, it is possible to identify which of the multiple types of feature quantities has a particular impact on the quality of the work object. Furthermore, the regression analysis calculates the variable importance, which indicates the degree to which each explanatory variable (i.e., feature quantity based on the measurement data) contributes to the objective variable (i.e., the quality data).
[0048] FIG. 7 is an explanatory diagram of multiple regression analysis, an example of regression analysis. The vertical axis of FIG. 7 represents feature values based on predetermined measurement data, and the horizontal axis represents quality data. For simplicity, FIG. 7 shows a single feature value; however, in practice, multiple regression analysis is performed based on multiple types of feature values. The multiple points plotted in FIG. 7 represent data corresponding to multiple workers. Line L1 is a line expressed by a predetermined regression formula based on the least squares method or the like. As a result of the multiple regression analysis, a t value corresponding to variable importance is calculated. The t value is a numerical value indicating the degree of influence that the feature value has on the quality data, and is calculated for each of multiple types of feature values. The regression analysis method is not limited to multiple regression analysis; as described above, other methods such as Lasso regression, Ridge regression, and partial least squares may also be used.
[0049] After performing the regression analysis in this manner, in step S108 of Fig. 3, the processing unit 13 searches for important features using the important feature search unit 137. That is, the processing unit 13 ranks the features in descending order of the variable importance for each feature calculated as a result of the regression analysis. For example, the important feature ranked first has the highest variable importance and can therefore be said to contribute most to the quality of the work object.
[0050] 3, the processing unit 13 displays the results of the analysis on the display unit 14. In this way, the processing unit 13 performs a regression analysis using the feature quantities of the measurement data related to the work of the worker on the work object as explanatory variables and the quality data of the work object as a response variable (S108), and displays the extraction results of the important feature quantities based on the regression analysis on the display unit 14 (display device) (display step).
[0051] FIG. 8 is a display example of the analysis results. In the example of FIG. 8, the analysis results are displayed when the height of the weld bead is used as quality data. Specifically, the feature amounts ranked from 1st to 5th are displayed vertically as important feature amounts on the display unit 14 (see FIG. 2). In the example of FIG. 8, the feature amount α and the feature amount β are listed as important feature amounts that have the greatest influence on the height of the weld bead (quality data). Here, the feature amounts α and β belong to one group with a strong correlation, and are read from the correlation data 11a (see FIG. 2).
[0052] As described above, one feature is selected from each group of strongly correlated features before the regression analysis is performed (S106 and S107 in FIG. 3 ), and therefore the important feature identified based on the variable importance is one of the feature α and the feature β (for example, the feature α). The processing unit 13 refers to the correlation data 11 a stored in the storage unit 11, and if there is another feature (for example, the feature β) associated with the important feature, displays this feature as well.
[0053] That is, when processing unit 13 displays an important feature extracted by regression analysis on display unit 14 (display device), if the important feature belongs to a predetermined group, processing unit 13 displays other feature values belonging to the group in association with the important feature, thereby enabling the user to easily understand that both feature value α and feature value β, which are highly correlated, affect the height of the weld bead.
[0054] 8, the upward arrow next to the feature quantity α indicates that the larger the value of the feature quantity α, the higher the quality of the work object. Also, the downward arrow next to the feature quantity β indicates that the smaller the value of the feature quantity β, the higher the quality of the work object. By showing these arrows, the user can understand at a glance whether the larger the value of the feature quantity based on the measurement data, the higher the quality, or whether the smaller the value of the feature quantity, the higher the quality.
[0055] In the example of Figure 8, the feature quantity γ, the feature quantity δ, and the feature quantity ε are shown as the second most important feature quantities. One of these feature quantities γ, δ, and ε is identified by regression analysis, and the rest are read from the correlation data 11a (see Figure 2). The feature quantity ρ is shown as the third most important feature quantity. This feature quantity ρ is not included in the correlation data 11a (see Figure 2) because there are no other features with which it has a particularly strong correlation.
[0056] <Effects> According to the first embodiment, quality data of lower outliers, which indicate extremely low quality of the workpiece, and the measurement data corresponding to this quality data are excluded from the regression analysis (S102 in FIG. 3). This eliminates data that acts as noise, thereby improving the analytical accuracy when performing regression analysis. Furthermore, when the processing unit 13 displays the important feature values identified by the regression analysis, it also displays other feature values that have a strong correlation with these important feature values. This allows the user to fully understand which feature values (measurement data) have a strong impact on the quality of the workpiece.
[0057] Furthermore, according to the first embodiment, by presenting to the user important features that are likely to affect the quality of the work object, it is possible to support the task of expressing the tacit knowledge (unverbalized knowledge) of skilled workers as explicit knowledge (verbalized knowledge). In other words, by automating the extraction of important features, at least a part of the task of converting the tacit knowledge of skilled workers into explicit knowledge can be made non-personal, and this task can be completed in a short period of time. Furthermore, the results of the regression analysis can be used for predetermined skill training, quality control, and assistance with work operations.
[0058] Second Embodiment The second embodiment differs from the first embodiment in that the processing unit 13 (see FIG. 2) performs regression analysis for all combinations when selecting one feature from a group of highly correlated features. Other aspects (such as the configuration of the skill analysis system 10: see FIG. 2) are the same as those of the first embodiment. Therefore, only the differences from the first embodiment will be described, and overlapping aspects will not be described.
[0059] 9 is a flowchart of the processing executed by the processing unit of the skill analysis system according to the second embodiment (see also FIG. 2 as needed). Note that steps S101 to S105 in FIG. 9 are the same as those in the first embodiment (see FIG. 3), and therefore a description thereof will be omitted. After investigating the correlation between feature quantities and storing the correlation data 11a in step S105 in FIG. 9, the processing unit 13 proceeds to step S201.
[0060] In step S201, the processing unit 13 determines whether all combinations have been examined. That is, the processing unit 13 determines whether all combinations have been examined when selecting one feature from a group of highly correlated features. In step S201, if there is a feature combination that has not yet been examined (i.e., a feature combination for which regression analysis has not been performed) (S201: No), the processing unit 13 proceeds to step S106. Here, how to select a feature will be described with reference to FIG. 10.
[0061] 10 is an explanatory diagram for selecting feature quantities to be used in regression analysis. In FIG. 10, a total of six groups with group IDs "G0001" to "G0006" correspond to the correlation data 11a shown in FIG. 6. For example, when selecting feature quantities one by one from the six groups shown in FIG. 10, the combinations are as follows: 4 ×3 2 = 144 combinations. The processing unit 13 refers to the correlation data 11a (see FIG. 2) in the storage unit 11, selects one feature from each of the six groups, and performs regression analysis after also including the feature values ρ, τ, σ, ω, ... that do not form a group as explanatory variables (S107: see FIG. 9). The processing unit 13 then performs regression analysis on all combinations of feature values selected one by one from the six groups.
[0062] Returning to FIG. 9 , the description will be continued. In step S106, the processing unit 13 selects one feature from each group of highly correlated features. The combination of features selected in step S106 is a combination for which regression analysis has not yet been performed. In step S107, the processing unit 13 performs regression analysis using the quality data as the response variable, using the analysis unit 136. That is, the processing unit 13 performs regression analysis using the feature of the combination selected in step S106 as well as a feature not included in the correlation data 11a (a feature with which there is no other feature with a strong correlation) as the explanatory variable, and the quality data as the response variable. After performing the processing of step S107, the processing unit 13 returns to step S201.
[0063] Furthermore, if all combinations of feature quantities have been examined in step S201 (S201: Yes), the processing of the processing unit 13 proceeds to step S202. In step S202, the processing unit 13 searches for important feature quantities based on a predetermined combination of feature quantities. That is, the processing unit 13 identifies, from among the combinations when selecting one type of feature quantity each from a plurality of groups, a combination for which the correlation coefficient between the estimated value of the objective variable (quality data) based on the combination and the actual value of the objective variable (quality data) is closest to 1. Then, based on the result of regression analysis for the combination for which the correlation coefficient is closest to 1, the processing unit 13 ranks the feature quantities in order of variable importance, and identifies, for example, the feature quantities ranked first to fifth as important feature quantities.
[0064] In step S109, the processing unit 13 displays the analysis results on the display unit 14. Note that an example of displaying the analysis results is the same as in the first embodiment (see FIG. 8), and therefore a description thereof will be omitted.
[0065] According to the second embodiment, the processing unit 13 performs a comprehensive regression analysis on all combinations of feature quantities selected from each group, thereby enabling highly accurate identification of feature quantities that affect the quality of the workpiece from among multiple types of feature quantities based on measurement data.
[0066] Third Embodiment The third embodiment differs from the first embodiment in that the processing unit 13A (see FIG. 11 ) performs regression analysis using each of a plurality of types of models and searches for important features based on the magnitude of the total value of scores based on variable importance. The rest of the third embodiment is the same as the first embodiment. Therefore, only the differences from the first embodiment will be described, and a description of the overlapping parts will be omitted.
[0067] FIG. 11 is a functional block diagram of a skill analysis system 10A according to the third embodiment. The processing unit 13A of the skill analysis system 10A shown in FIG. 11 includes a multiple-model analysis unit 136A instead of the analysis unit 136 (see FIG. 2) included in the configuration described in the first embodiment. The multiple-model analysis unit 136A performs regression analysis based on multiple models employing different analytical methods. Examples of such multiple models include multiple regression analysis, Lasso regression, partial least squares, and Ridge regression. When a large amount of data is involved, a gradient boosting decision tree or a neural network may also be used.
[0068] 12 is a flowchart of the processing executed by the processing unit of the skill analysis system (see also FIG. 11 as appropriate). Note that steps S101 to S106 in FIG. 12 are the same as those in the first embodiment (see FIG. 3), and therefore their explanation will be omitted. After selecting one feature from each group of highly correlated features in step S106, the processing unit 13A proceeds to step S307.
[0069] In step S307, the processing unit 13A performs regression analysis using each of the multiple models using the multiple model analysis unit 136A. For example, the processing unit 13A performs regression analysis using each of three methods: gradient boosting decision tree, partial least squares, and Lasso regression. Note that the multiple models used in step S307 may be set in advance, or may be selectable by the user, as described below.
[0070] FIG. 13 shows an example of a display screen for selecting a model to be used in regression analysis. As shown in FIG. 13, a predetermined model for regression analysis may be selected from a pull-down list K2 based on a user's input operation. In the example of FIG. 13, the partial least squares method is selected from among the candidate models to be used in regression analysis. The name of the model selected by the user is displayed in a presentation field K3 below the pull-down list K2. That is, the processing unit 13A (see FIG. 11) displays on the display unit 14 (display device) a selection screen for selecting, based on a user's input operation, from among the candidate models for regression analysis, multiple types of models to be used for extracting important features.
[0071] After selecting multiple models, the user presses the confirm button B2. In this way, by allowing the user to select multiple models to actually use from among the candidate regression analysis models, the degree of freedom in user settings is increased.
[0072] Next, in step S308 of FIG. 12 , the processing unit 13A searches for important features based on the total score values using the important feature search unit 137. For example, the processing unit 13A ranks the features in order of variable importance based on the results when a predetermined regression analysis model is used, and then assigns scores corresponding to the ranking. The processing unit 13A also performs similar processing when regression analysis is performed using other models. Then, the processing unit 13A calculates the total score value corresponding to multiple types of models for each feature, and extracts a predetermined number of feature values in order of the magnitude of the total value as important features.
[0073] 14 is an explanatory diagram regarding extraction of important features when multiple types of models are used. In the example of FIG. 14, for example, when a gradient boosting decision tree is used as the regression analysis method, the features α, γ, ζ, κ, and μ are ranked from 1st to 5th in order of decreasing variable importance. The variable with the highest variable importance is assigned 5 points, 2nd place is assigned 4 points, 3rd place is assigned 3 points, 4th place is assigned 2 points, and 5th place is assigned 1 point. The same applies when partial least squares or Lasso regression is performed.
[0074] The processing unit 13A then calculates the total score for each regression analysis method for each of the feature quantities α, γ, ζ, κ, μ, τ, and ω. In the example of FIG. 14 , the feature quantity ζ has the highest total score, followed by the feature quantities γ, α, τ, κ, ω, and μ (however, the total scores for the feature quantities κ and ω are equal). In this way, the processing unit 13 performs regression analysis for each of multiple types of models and scores the feature quantities for each model based on the variable importance, which indicates the degree to which the explanatory variable (measurement data) contributes to the target variable (quality data). The processing unit 13 then extracts important feature quantities based on the magnitude of the total score. Note that the scoring method and feature ranking method shown in FIG. 14 are merely examples and are not limited thereto.
[0075] After searching for important features in this way, the processing unit 13A displays the analysis results in step S109 of Fig. 12. Note that the display example of the analysis results is the same as that of the first embodiment (see Fig. 8), and therefore description thereof will be omitted.
[0076] According to the third embodiment, the processing unit 13A uses a plurality of models as a regression analysis method. Each regression analysis method (i.e., a plurality of models) has advantages and disadvantages, such as being superior to other methods in some respects and inferior in others, but by comparing the total scores based on the plurality of models, it is possible to obtain highly versatile (i.e., robust) and highly accurate analysis results.
[0077] Although the skill analysis system 10, 10A and the skill analysis method according to the present disclosure have been described in the above embodiments, the present disclosure is not limited to these descriptions and various modifications can be made. For example, although the embodiments have been described with respect to a case where one type of quality data is used in the regression analysis, multiple types may be used. In this case, the series of processes shown in FIG. 3 and the like are performed for each type of quality data.
[0078] 3, the order of steps S102 and S103 may be reversed. That is, the processing unit 13 may extract feature quantities based on the measurement data and then exclude lower outliers from the quality data. The same applies to the second and third embodiments.
[0079] Furthermore, the processing (skill analysis method) executed by the skill analysis system 10, 10A may be executed as a predetermined computer program. The program may be provided via a communication line or may be written to a recording medium such as a CD-ROM and distributed.
[0080] Furthermore, the present disclosure is not limited to each embodiment and includes various modifications. For example, each embodiment has been described in detail to clearly explain the present disclosure, and is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another example to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.
[0081] Furthermore, the above-mentioned components, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The above-mentioned components, functions, etc. may also be implemented in software, with a processor interpreting and executing a program that implements each function. Information such as programs, tables, and files that implement each function can be stored in memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD. Furthermore, the control lines and information lines shown are those considered necessary for explanation, and do not necessarily represent all control lines and information lines in the product. In reality, it can be assumed that almost all components are interconnected.
[0082] 10, 10A Skill analysis system 11 Memory unit 11a Correlation data 12 Input unit 13, 13A Processing unit 14 Display unit (display device) 131 Lower outlier removal unit 132 Feature extraction unit 133 Preprocessing unit 134 Correlation investigation unit 135 Feature selection unit 136 Analysis unit 136A Multiple model analysis unit 137 Important feature search unit M1 Worker W1 Work object S102 Step (outlier removal step) S109 Step (display step)
Claims
1. A skill analysis system comprising a processing unit which performs a regression analysis using feature quantities of measurement data relating to a worker's work on a work object as explanatory variables and quality data of the work object as a response variable, and displays the extraction results of important feature quantities based on the regression analysis on a display device, wherein the processing unit excludes outliers on the low quality side of the quality data and the measurement data corresponding to the quality data of the outliers from the regression analysis.
2. The skill analysis system according to claim 1, wherein the processing unit includes in the regression analysis outliers on the high quality side of the quality data and the measurement data corresponding to the quality data.
3. The skill analysis system according to claim 1, wherein the processing unit performs a predetermined correlation investigation on a plurality of types of the feature quantities, selects one type of feature quantity from each group showing combinations of types of feature quantities with relatively strong correlations, and includes the selected feature quantities in the regression analysis.
4. The skill analysis system of claim 3, wherein when the processing unit causes the display device to display the important feature extracted by the regression analysis, if the important feature belongs to a specific group, the processing unit causes other features belonging to the group to be displayed in association with the important feature.
5. The skill analysis system of claim 1, wherein the processing unit, when the higher the quality of the work object, the larger the value of the quality data, sets a threshold value obtained by subtracting the multiplication of the interquartile range of the quality data and a predetermined value from the first quartile of the quality data, and excludes the quality data whose value is smaller than the threshold value and the measurement data corresponding to the quality data from the subject of the regression analysis.
6. The skill analysis system of claim 1, wherein the processing unit, when the higher the quality of the work object, the smaller the value of the quality data, sets a threshold value obtained by adding the third quartile of the quality data to a multiplied value of the interquartile range of the quality data and a predetermined value, and excludes the quality data whose value is greater than the threshold value and the measurement data corresponding to the quality data from the regression analysis.
7. The skill analysis system according to claim 5 or 6, wherein the processing unit causes a box-and-whisker plot of the quality data to be displayed on the display device, and further causes the magnitude of the predetermined value to be adjusted by a user's input operation.
8. The skill analysis system according to claim 3, wherein the processing unit identifies, from among combinations when selecting one type of feature from each of the multiple groups, a combination in which the correlation coefficient between the estimated value of the objective variable based on the combination and the actual value of the objective variable is closest to 1, and performs the regression analysis based on the feature including the combination.
9. The skill analysis system of claim 1, wherein the processing unit performs the regression analysis for each of a plurality of types of models, scores the features for each model based on variable importance indicating the degree to which the explanatory variable contributes to the target variable, and extracts the important features based on the magnitude of the total score.
10. The skill analysis system of claim 9, wherein the processing unit causes the display device to display a selection screen for selecting, by a user's input operation, from among candidate models for the regression analysis, multiple types of models to be used for extracting the important features.
11. A skill analysis method comprising: an outlier exclusion step of excluding from regression analysis outliers on the low quality side in quality data of a work object and measurement data related to work by a worker on the work object that corresponds to the quality data of the outliers; and a display step of performing the regression analysis using features of the measurement data as explanatory variables and the quality data as a response variable, and displaying the extraction results of important features based on the regression analysis on a display device.
Citation Information
Patent Citations
Data processing method, data processing program, recording medium recording same program, and data processor
JP2007329415A
Analysis device, analysis method, and program
JP2018151883A
Information processing apparatus, quality related formula generation method and quality related formula generation program
JP2019197245A
Analysis system
JP2020126454A
Identification device, identification method, and identification program
JP2022152605A
