JP2023061289A - Prediction and estimation system, learning device, and prediction and estimation device - Google Patents

Abstract

To improve accuracy of prediction and estimation utilizing more data.SOLUTION: A learning device according to one aspect includes: a latent variable calculation unit which, when the learning device inputs teacher data on an objective variable, teacher data on an explanatory variable which acts on the objective variable, and teacher data on an auxiliary variable which is a variable other than the objective variable and correlates with the explanatory variable and which is observed when the explanatory variable is given, calculates teacher data on a latent variable from at least the teacher data on the auxiliary variable and outputs latent variable calculation information; and a prediction and estimation model learning unit which, when the learning device inputs the teacher data on the objective variable, the teacher data on the explanatory variable, and the teacher data on the latent variable calculated by the latent variable calculation unit, creates a prediction and estimation model of the objective variable.SELECTED DRAWING: Figure 1

Description

The present invention relates to a prediction/estimation system, a learning device, and a prediction/estimation device.

In prediction/estimation of objective variables using supervised machine learning, objective variables and explanatory variables, which serve as teacher data, are generally used in learning prediction/estimation models. In addition, auxiliary variables (auxiliary variables) that are available at the time of prediction/estimation and are other than the objective variable may also be used as variables that have a causal relationship that results when the explanatory variable is the cause. be.

For example, the following method is known as a method of learning such a prediction/estimation model. First, the values of chemical substances in multiple experimental samples are used as teacher data for objective variables, and the near-infrared spectral data converted from the collected spectra of multiple experimental samples, which are teacher data for explanatory variables, are used as part of the data. A random selection is made into the calibration set and partly into the validation set.

Then, the calibration set and the validation set are subjected to principal component analysis to obtain a spectral feature space, and in this spectral feature space, the samples in which the calibration set and the validation set are most similar are selected as the calibration subset by the Mahalanobis distance method. Collect as a Syndrome Set.

After that, extract the number of principal components from the syndrome set as training data for auxiliary variables, build a regression model as an input layer of a backpropagation (BP) neural network, and learn a prediction/estimation model (Patent Document 1 reference).

Chinese Patent Application Publication No. 107655850

However, in the system disclosed in Patent Document 1, instead of near-infrared spectrum data, which is teacher data for explanatory variables, the number of principal components, which is teacher data for auxiliary variables, is used to construct a regression model and make predictions.・The estimation model is learned. In order to improve the accuracy of prediction/estimation, it is desirable to learn the prediction/estimation model with more teacher data. For this reason, it is conceivable to learn a prediction/estimation model using both explanatory variables and auxiliary variables as teacher data.

However, in the system as described above, since the auxiliary variables are correlated with the explanatory variables, the accuracy of prediction/estimation tends to deteriorate. Such a tendency is widely known as multicollinearity, especially in multiple regression analysis.

One aspect of the present invention is a prediction/estimation system, a learning device, and a prediction/estimation device capable of improving accuracy of prediction/estimation.

A prediction/estimation system according to an aspect of the present invention includes teacher data of an objective variable, teacher data of an explanatory variable acting on the objective variable, and variables other than the objective variable that are correlated with the explanatory variable, a learning device for inputting teacher data of an auxiliary variable observed when the explanatory variable is given, and creating a prediction/estimation model of the objective variable; and the explanation for predicting/estimating the objective variable. Input data for prediction/estimation of variables and data for prediction/estimation of the auxiliary variables for predicting/estimating the objective variable, and based on the prediction/estimation model created by the learning device, a prediction/estimation device that calculates the prediction/estimation value of the objective variable, wherein the learning device calculates teacher data of the latent variable from at least the teacher data of the auxiliary variable, and outputs latent variable calculation information. A first latent variable calculation unit, training data of the objective variable, training data of the explanatory variable, and training data of the latent variable calculated by the first latent variable calculation unit are input, and the objective variable and a prediction/estimation model learning unit that creates a prediction/estimation model of at least the auxiliary variable, based on the latent variable calculation information, from the prediction/estimation data of the auxiliary variable. a second latent variable calculation unit that calculates prediction/estimation data; and from the prediction/estimation data of the latent variables calculated by the second latent variable calculation unit and the prediction/estimation data of the explanatory variables, the a prediction/estimation unit that calculates a prediction/estimation value of the objective variable based on the prediction/estimation model.

According to the prediction/estimation system according to one aspect of the present invention, the teacher data of the objective variable, the teacher data of the explanatory variables, and the teacher data of the auxiliary variables are input, and at least the teacher data of the auxiliary variables are used to predict the potential variables. Teacher data is calculated, and latent variable calculation information and a prediction/estimation model of objective variables are created. For this reason, by using teacher data of latent variables in which the correlation between explanatory variables and teacher data or between each other is reduced more than that of auxiliary variables, more teacher data can be used, and between teacher data By eliminating correlation, it becomes possible to create a more accurate prediction/estimation model. Further, from at least the data for prediction/estimation of the auxiliary variable, the data for prediction/estimation of the latent variable is calculated based on the latent variable calculation information, and the calculated data for prediction/estimation of the latent variable and the explanatory variable are combined. Based on the prediction/estimation data, the prediction/estimation value of the objective variable is calculated based on the prediction/estimation model. For this reason, more data for prediction/estimation can be used by using prediction/estimation data for latent variables whose correlation between explanatory variables and data for prediction/estimation is reduced more than for auxiliary variables. and the elimination of the correlation between the prediction/estimation data, it becomes possible to obtain a more highly accurate prediction/estimation value of the objective variable.

According to one aspect of the present invention, it is possible to improve the accuracy of prediction/estimation using more data.

FIG. 1 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the first embodiment of the present invention. FIG. 2 is an explanatory diagram showing a basic hardware configuration of a learning device and/or a prediction/estimation device of the prediction/estimation system. FIG. 3 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the second embodiment of the present invention. FIG. 4 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the third embodiment of the present invention. FIG. 5 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the fourth embodiment of the present invention. FIG. 6 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the fifth embodiment of the present invention. FIG. 7 is an explanatory diagram showing the basic configuration of a welding strength estimation system for laser welding to which a prediction/estimation system according to a sixth embodiment of the present invention is applied. FIG. 8 is an explanatory diagram showing the basic configuration of a learning device of a nondestructive measurement system to which the prediction/estimation system according to the seventh embodiment of the present invention is applied. FIG. 9 is an explanatory diagram showing the basic configuration of a nondestructive testing device of the nondestructive measurement system. FIG. 10 is an explanatory diagram showing the basic configuration of a quality judgment/prediction system for cutting quality of a laser cutting apparatus to which the prediction/estimation system according to the eighth embodiment of the present invention is applied. FIG. 11 is a graph showing the relationship between welding temperature and elapsed time for explaining auxiliary variables in an example in which the prediction/estimation system according to the fourth embodiment shown in FIG. 5 is applied to actual laser welding. FIG. 12 is a graph showing the relationship between estimated values and measured values and the frequency distribution of errors of estimated values with respect to estimated values in Comparative Example and Example.

Hereinafter, a prediction/estimation system, a learning device, and a prediction/estimation device according to embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the following embodiments do not limit the invention according to each claim, and all combinations of features described in the embodiments are essential to the solution of the invention. Not exclusively. Further, in the following embodiments, the scale and dimensions of each component may be exaggerated, and some components may be omitted.

[First embodiment]
FIG. 1 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the first embodiment of the present invention. FIG. 2 is an explanatory diagram showing a basic hardware configuration of a learning device and/or a prediction/estimation device of the prediction/estimation system.

As shown in FIG. 1, a prediction/estimation system 100 according to the first embodiment is a system for obtaining prediction/estimation values 8 of various types of objective variables. The prediction/estimation system 100 includes teacher data 1 of the objective variable, teacher data 2 of the explanatory variable acting on the objective variable, and variables other than the objective variable that are correlated with the explanatory variable, and when the explanatory variable is given A learning device 10 that inputs the teacher data 3 of the auxiliary variables observed in , and creates a target variable prediction/estimation model 7 by, for example, machine learning, and an explanatory variable prediction/estimation model 7 for predicting/estimating the target variable. Estimation data 4 and auxiliary variable prediction/estimation data 5 for predicting/estimating the objective variable are input, and the objective variable is predicted based on the prediction/estimation model 7 created by the learning device 10. a prediction/estimation device 20 for calculating the estimated value 8;

The learning device 10 generates, from at least the auxiliary variable teacher data 3, latent variable teacher data (reduced correlation) that has a smaller correlation (reduced correlation) with the explanatory variable teacher data 2 than the auxiliary variable teacher data 3 ( (not shown), and outputs latent variable calculation information 6, the teacher data 1 of the objective variable, the teacher data 2 of the explanatory variable, and the first latent variable calculator 11 and a prediction/estimation model learning unit 12 for inputting latent variable teacher data and, for example, performing machine learning to create a target variable prediction/estimation model 7 .

The prediction/estimation device 20 extracts at least the auxiliary variable prediction/estimation data 5 from the explanatory variable prediction/estimation data 4 rather than the auxiliary variable prediction/estimation data 5 based on the latent variable calculation information 6. A second latent variable calculation unit 21 that calculates prediction/estimation data (not shown) of latent variables with a small correlation (reduced correlation) between or between a prediction/estimation unit 22 that calculates a prediction/estimation value 8 of the objective variable based on the prediction/estimation model 7 from the data for prediction/estimation of the latent variable and the data 4 for prediction/estimation of the explanatory variable; include.

The prediction/estimation system 100 is a system capable of obtaining prediction/estimation values of various types of objective variables as described above. In this case, for example, the target variable is the welding strength F (N/mm ² ). Further, for example, explanatory variables include welding conditions (laser output value, laser irradiation time, etc.). Further, the auxiliary variables correlated with the explanatory variables include, for example, physical quantities observed during or after welding (welding temperature, sound, light, color, etc.). Note that the objective variable, explanatory variable, and auxiliary variable used in the prediction/estimation system 100 are not limited to these examples.

Further, the latent variable teacher data and the latent variable prediction/estimation data calculated by the first latent variable calculation unit 11 of the learning device 10 and the second latent variable calculation unit 21 of the prediction/estimation device 20 are respectively used as auxiliary variables An auxiliary variable, or an auxiliary variable and an explanatory variable, are, for example, dimensionality-reduced or dimensionality-compressed so as to reduce the correlation between or between the explanatory variables than the explanatory variables.

That is, in the present embodiment, a latent variable is a variable that characterizes an objective variable like an explanatory variable, and is a variable that is estimated from an observed variable (auxiliary variable) but not directly observed. In other words, a latent variable means a variable that characterizes the observed variable and is dimensionality-reduced or dimensionality-compressed from at least the auxiliary variables.

The explanatory variables, objective variables, latent variables, and latent variable calculation information 6 used in the prediction/estimation system 100 can each consist of one or more variables or parameters. In this specification, "prediction/estimation of the objective variable" is used to mean "prediction of the objective variable" or "estimation of the objective variable". "Objective variable prediction" means to assume an objective variable that is expected to be realized in the future (future), and "objective variable estimation" is an objective variable that is currently realized but cannot be directly observed. means to estimate "Correlation" is not limited to a linear relationship between multiple variables, but regardless of whether it is linear or non-linear, a relationship between multiple variables in which a change in one variable causes a change in the other variables. means.

The first latent variable calculation unit 11 of the learning device 10 calculates at least the auxiliary variable out of the data including the objective variable teacher data 1, the explanatory variable teacher data 2, and the auxiliary variable teacher data 3 input to the learning device 10. The teacher data of the latent variables are calculated from the data including the teacher data 3 . The first latent variable calculator 11 calculates latent variable training data using, for example, a predetermined algorithm. Algorithms include, for example, regression analysis (RA), principal component analysis (PCA), singular value decomposition (SVD), linear discriminant analysis (LDA), independent component analysis (Independent Component Analysis: ICA), and at least one of Gaussian Process Latent Variable Model (GPLVM). Note that the algorithms are not limited to these. The latent variable calculation information 6 means various information that is determined in the first latent variable calculation unit 11 when calculating teacher data for latent variables and used when calculating additional latent variables. or algorithms are included.

The prediction/estimation model learning unit 12 of the learning device 10 receives the objective variable teacher data 1 and the explanatory variable teacher data 2 input to the learning device 10, and the latent variable teacher data calculated by the first latent variable calculator 11. By inputting data, the target variable prediction/estimation model 7 is created by, for example, machine learning. Learning methods include, but are not limited to, algorithms such as regression, classification, clustering, discrimination, interpolation, feature quantity extraction, and time series modeling. The prediction/estimation model 7 learned by the prediction/estimation model learning unit 12 includes various parameters, various formulas and/or algorithms for prediction/estimation that characterize the learning model. Then, the learning device 10 outputs at least the latent variable calculation information 6 and the prediction/estimation model 7 to the prediction/estimation device 20 .

On the other hand, the second latent variable calculator 21 of the prediction/estimation device 20 calculates prediction/estimation data for latent variables based on the latent variable calculation information 6 from at least the prediction/estimation data 5 for the auxiliary variables. Therefore, when the latent variable calculation information 6 includes an algorithm, the second latent variable calculation unit 21 uses the same algorithm as the one used to calculate latent variable teacher data in the first latent variable calculation unit 11 of the learning device 10. Calculate data for predicting and estimating latent variables using algorithms. In this way, prediction/estimation using the same prediction/estimation model 7 becomes possible by calculating latent variables using the same algorithm as that used in learning during prediction/estimation.

The prediction/estimation unit 22 of the prediction/estimation device 20 uses the prediction/estimation data 4 of the explanatory variables input to the prediction/estimation device 20, the prediction/estimation model 7, and the second latent variable calculation unit 21. Based on the latent variable prediction/estimation data, the target variable prediction/estimation value 8 is calculated and output.

The latent variable calculation information 6 passed from the learning device 10 side to the prediction/estimation device 20 side can be selected by the learning device 10 or the prediction/estimation device 20 so as to be suitable for calculating latent variables. For example, when the learning device 10 calculates the teacher data of the latent variables by principal component analysis, the latent variable parameters may be principal component vectors and eigenvalues. It is also possible to select latent variable parameters excluding principal component vectors whose rates are higher than a predetermined value. As a result, it is possible to use the latent variables that have a smaller correlation with or between the explanatory variables than the auxiliary variables, together with the explanatory variables, for the teacher data and prediction/estimation data. aggravation can be avoided.

Objective variable teacher data 1, explanatory variable teacher data 2, auxiliary variable teacher data 3, explanatory variable prediction/estimation data 4, and auxiliary variable prediction/estimation data 5 in the prediction/estimation system 100. can be stored, for example, in a stationary or portable storage device or storage medium (not shown). Further, each data can be transmitted and received via an information communication medium such as the Internet, and may be raw data acquired by a measuring device such as a sensor (not shown). Further, the latent variable calculation information 6 and the prediction/estimation model 7 may be input/output between the learning device 10 and the prediction/estimation device 20 via the above storage medium or information communication medium.

As shown in FIG. 2, the learning device 10 and/or the prediction/estimation device 20 of the prediction/estimation system 100 has a basic hardware configuration such as a CPU 201, a RAM 202, a ROM 203, and a HDD (hard disk drive) 204. , an SSD (Solid State Drive) 205 and a memory card 206 . The learning device 10 and/or the prediction/estimation device 20 also includes an input I/F (interface) 207, an output I/F (interface) 208, and a communication I/F (interface) 209, for example. Each component 201 to 209 is interconnected by a bus 200, respectively.

The CPU 201 controls the learning device 10 and/or the prediction/estimation device 20 by executing various programs stored in the RAM 202, ROM 203, HDD 204, SSD 205, and the like. The CPU 201 implements the functions of the first latent variable calculation unit 11 and the prediction/estimation model learning unit 12 by executing a learning program in the learning device 10 . Further, the CPU 201 implements the functions of the second latent variable calculation unit 21 and the prediction/estimation unit 22 by executing the prediction/determination program in the prediction/estimation device 20 . Note that the CPUs 201 of the learning device 10 and the prediction/estimation device 20 can be configured to control the entire prediction/estimation system 100 by cooperating.

The RAM 202 can be used as a work area for arithmetic processing of the CPU 201 . The ROM 203 stores at least the various programs described above in a readable manner. The HDD 204 and SSD 205 store the various data described above in a readable and writable manner. The memory card 206 stores these various data in a readable and writable manner, and constitutes a removable storage medium for each device 10 , 20 . The HDD 204, SSD 205, and memory card 206 implement the functions of the storage device or storage medium described above.

For example, a sensor 212 is connected to the input I/F 207 to acquire detection information. The sensor 212 includes various sensors such as a temperature sensor, an optical sensor, an acoustic sensor, and an image sensor. The input I/F 207 is connected to the touch panel 211 functioning as an operation unit or an input unit of the learning device 10 and/or the prediction/estimation device 20, and receives information accompanying an operation input from the user of the prediction/estimation system 100. . Various input devices such as a keyboard and a mouse (including a trackball mouse) (not shown) can also be connected to the input I/F 207 .

For example, a display 210 as a display device is connected to the output I/F 208, and various kinds of information displayed on the monitor by the learning device 10 and/or the prediction/estimation device 20 are output. The touch panel 211 may be provided on the display 210 . Also, the learning device 10 and/or the prediction/estimation device 20 can be indirectly or directly connected to a server device and external devices connected to a network such as the Internet (not shown) via the communication I/F 209 .

In the prediction/estimation system 100 of the first embodiment, the first latent variable calculation unit 11 of the learning device 10 calculates latent variable teacher data from data including at least the auxiliary variable teacher data 3, and calculates the latent variable teacher data. The prediction/estimation model learning unit 12 learns and creates the prediction/estimation model 7 using the data including Then, the prediction/estimation device 20 is provided with the latent variable calculation information 6 determined by the learning device 10 by calculating the teacher data of the latent variables. The second latent variable calculation unit 21 of the prediction/estimation device 20 to which the latent variable calculation information 6 is provided uses the same algorithm as that used by the first latent variable calculation unit 11 to calculate at least auxiliary variables for prediction/estimation. Calculate latent variable prediction/estimation data from data including data 5, and calculate target variable prediction/estimation value 8 in a prediction/estimation unit 22 using the data including the latent variable prediction/estimation data. . Thus, according to the prediction/estimation system 100 of the first embodiment, the accuracy of prediction/estimation can be improved by converting an auxiliary variable into a latent variable.

[Second embodiment]
FIG. 3 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the second embodiment of the present invention. In the following description including FIG. 3, the same reference numerals are given to the same constituent elements as those of the first embodiment and its modification, so redundant description will be omitted below.

As shown in FIG. 3, in the prediction/estimation system 100A according to the second embodiment, the first latent variable calculation unit 11 of the learning device 10 receives teacher data 2 of explanatory variables and teacher data 3 of auxiliary variables. , latent variable teacher data (not shown) is calculated, and latent variable calculation information 6 is output. Further, the second latent variable calculation unit 21 of the prediction/estimation device 20 inputs the prediction/estimation data 4 of the explanatory variables and the prediction/estimation data 5 of the auxiliary variables, and calculates the latent variables based on the latent variable calculation information 6 . Data for predicting/estimating variables (not shown) is calculated.

That is, in the prediction/estimation system 100A of the second embodiment, the first latent variable calculation unit 11 of the learning device 10 uses the teaching data 3 of the auxiliary variables as input data to calculate the teaching data of the latent variables. Furthermore, it is different from the learning device 10 of the prediction/estimation system 100 of the first embodiment in that it also includes teacher data 2 of explanatory variables. In addition to the auxiliary variable prediction/estimation data 5, the second latent variable calculation unit 21 of the prediction/estimation device 20 uses explanatory variables is different from the prediction/estimation device 20 of the prediction/estimation system 100 of the first embodiment in that it also includes the prediction/estimation data 4 of the first embodiment.

According to the second embodiment, the first latent variable calculation unit 11 of the learning device 10 and the second latent variable calculation unit 21 of the prediction/estimation device 20 also use the data related to the explanatory variables to obtain teacher data for each latent variable. and calculation of prediction/estimation data. Therefore, highly accurate prediction/estimation using latent variables with reduced correlation between explanatory variables and auxiliary variables is possible. Also, according to the second embodiment, more data is used to calculate latent variables, so latent variables can be calculated based on higher-dimensional and advanced algorithms. As a result, it is possible to further improve the accuracy of prediction/estimation using auxiliary variables while avoiding deterioration of prediction/estimation accuracy due to correlation.

[Third Embodiment]
FIG. 4 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the third embodiment of the present invention. As shown in FIG. 4, in the prediction/estimation system 100B according to the third embodiment, the second latent variable calculation unit 21 of the prediction/estimation device 20 includes predictive/estimation data 4 for explanatory variables and prediction/estimation data 4 for auxiliary variables. In addition to the estimation data 5, the explanatory variable teacher data 2, the auxiliary variable teacher data 3, and the latent variable teacher data 13 calculated by the first latent variable calculator 11 of the learning device 10, Based on this, latent variable prediction/estimation data 23 is calculated.

That is, in the prediction/estimation system 100B of the third embodiment, the second latent variable calculation unit 21 of the prediction/estimation device 20 uses auxiliary variable In addition to prediction/estimation data 5 and explanatory variable prediction/estimation data 4, explanatory variable teacher data 2, auxiliary variable teacher data 3, and latent variable teacher data 13 are also included. It is different from the estimation system 100A.

According to the third embodiment, in the second latent variable calculation unit 21 of the prediction/estimation device 20, the prediction/estimation data 23 of the latent variables are calculated using the teaching data 2 and 3 of the explanatory variables and the auxiliary variables. ing. Therefore, the latent variable prediction/estimation data 23 can be calculated using more data than in the second embodiment. As a result, it is possible to apply a more advanced algorithm and improve the accuracy of prediction/estimation using auxiliary variables while avoiding deterioration in prediction/estimation accuracy due to correlation.

[Fourth Embodiment]
FIG. 5 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the fourth embodiment of the present invention. In the prediction/estimation system 100C according to the fourth embodiment, the first latent variable calculation unit 11 of the learning device 10 calculates the latent variables using regression analysis and principal component analysis as a predetermined algorithm (latent variable calculation information 6). Calculate teacher data (not shown). The second latent variable calculation unit 21 of the prediction/estimation device 20 uses the algorithm (latent variable calculation information 6) used in the first latent variable calculation unit 11 to generate latent variable prediction/estimation data (not shown). ) are calculated.

More specifically, as shown in FIG. 5, in the prediction/estimation system 100C according to the fourth embodiment, the first latent variable calculation unit 11 of the learning device 10 converts the teacher data 3 of the auxiliary variables into the teacher data 3 of the explanatory variables. A first regression analysis unit 14 performs regression analysis on data 2 to create a regression model, and a regression error 15 of the regression analysis by the first regression analysis unit 14 is subjected to principal component analysis to calculate latent variable teacher data (not shown). and a first principal component analysis unit 16 . The latent variable calculation information 6 includes regression model parameters and regression equations, and principal component vectors and eigenvalues of principal component analysis. Further, the second latent variable calculation unit 21 of the prediction/estimation device 20 performs regression analysis on the auxiliary variable prediction/estimation data 5 with the explanatory variable prediction/estimation data 4 based on the latent variable calculation information 6 to create a regression model. A second regression analysis unit 24 to create, and a second regression analysis unit 24 for calculating latent variable prediction/estimation data (not shown) by principal component analysis of regression errors 25 by the second regression analysis unit 24 based on the latent variable calculation information 6 and a two-principal component analysis unit 26 .

That is, in the prediction/estimation system 100C of the fourth embodiment, the first latent variable calculation unit 11 of the learning device 10 and the second latent variable calculation unit 21 of the prediction/estimation device 20 are the first regression analysis unit 14 and the first principal It includes a component analysis unit 16 and performs regression analysis and principal component analysis to calculate teacher data for latent variables, and the second latent variable calculation unit 21 performs a second regression analysis unit 24 and a second principal component analysis unit 26. It is different from the prediction/estimation system 100A according to the second embodiment in that the latent variable prediction/estimation data is calculated by performing regression analysis and principal component analysis.

In the prediction/estimation system 100C of the fourth embodiment, the first regression analysis unit 14 of the first latent variable calculation unit 11 of the learning device 10 performs regression analysis (RA ) and a regression model is created. A regression error 15 of the regression model created by the first regression analysis unit 14 is subjected to principal component analysis (PCA) by the first principal component analysis unit 16, thereby calculating teacher data of latent variables. At this time, for example, a principal component vector having a small correlation between the explanatory variable and the auxiliary variable is selected as the principal component parameter. As a result, the first latent variable calculation unit 11 of the learning device 10 can calculate latent variable teacher data that has a small correlation with the explanatory variable teacher data 2 .

In the prediction/estimation model learning unit 12, the prediction/estimation model 7 is created by, for example, machine learning using the data including the teacher data of the latent variables with small correlation calculated in this way. This makes it possible to avoid deterioration in prediction/estimation accuracy due to the correlation between the teacher data 3 of the auxiliary variable and the teacher data 2 of the explanatory variable. In this case, the regression model included in the latent variable calculation information 6 may include data such as regression coefficients. Also, the principal component parameters included in the latent variable calculation information 6 may include data such as the eigenvalues of the principal component vector and the sample variance matrix.

Then, in the second latent variable calculation unit 21 of the prediction/estimation device 20, the latent variable calculation information 6 including the parameters and regression formula of the regression model in the first latent variable calculation unit 11 and the principal component parameters of the principal component analysis is Used. Using this latent variable calculation information 6, the second regression analysis unit 24 performs regression analysis (RA) on the auxiliary variable prediction/estimation data 5 with the explanatory variable prediction/estimation data 4 to create a regression model. be. The regression error 25 of the regression model generated by the second regression analysis unit 24 is subjected to principal component analysis (PCA) by the second principal component analysis unit 26, thereby calculating latent variable prediction/estimation data. Therefore, in the prediction/estimation device 20, the latent variable prediction/estimation data having a small correlation with the explanatory variable prediction/estimation data 4 is calculated using the latent variable calculation information 6 similar to that of the learning device 10. becomes possible.

In the prediction/estimation unit 22, the prediction/estimation value 8 of the objective variable is calculated using the data including the prediction/estimation data of the latent variable with small correlation calculated in this way. As a result, the deterioration of the prediction/estimation accuracy due to the correlation between the auxiliary variable prediction/estimation data 5 and the explanatory variable prediction/estimation data 4 is reduced, and the target variable prediction/estimation value 8 is increased. can be obtained with precision.

[Fifth embodiment]
FIG. 6 is an explanatory diagram showing the basic configuration of the prediction/estimation system according to the fifth embodiment of the present invention. As shown in FIG. 6, in the prediction/estimation system 100D according to the fifth embodiment, the first latent variable calculation unit 11 of the learning device 10 calculates latent variable teacher data 13 using a Gaussian process latent variable model (GPLVM). and the latent variable calculation information 6 includes hyperparameters of a Gaussian process latent variable model (GPLVM). Also, the prediction/estimation model learning unit 12 of the learning device 10 performs, for example, machine learning using Gaussian process regression (GPR) to create the prediction/estimation model 7 of the objective variable. Further, the second latent variable calculation unit 21 of the prediction/estimation device 20 calculates latent variable prediction/estimation data 23 using a Gaussian process latent variable model (GPLVM), and the prediction/estimation unit 22 of the prediction/estimation device 20 calculates the predicted/estimated value 8 of the objective variable by Gaussian process regression (GPR).

That is, in the prediction/estimation system 100D of the fifth embodiment, the first latent variable calculation unit 11 of the learning device 10 and the second latent variable calculation unit 21 of the prediction/estimation device 20 use the latent variable teacher data 13 and the latent variable The Gaussian process latent variable model (GPLVM) algorithm is used to calculate the prediction/estimation data 23, and the prediction/estimation model learning unit 12 of the learning device 10 and the prediction/estimation unit 22 of the prediction/estimation device 20 However, the prediction/estimation system 100B of the third embodiment differs from the prediction/estimation system 100B of the third embodiment in that a Gaussian process regression (GPR) algorithm is used for learning the target variable prediction/estimation model 7 and calculating the target variable prediction/estimation value 8. are different.

According to the fifth embodiment, in the first latent variable calculation unit 11 and the second latent variable calculation unit 21, the latent variable teacher data 13 and the latent variable prediction/estimation data 23 are converted into a Gaussian process latent variable model (GPLVM). The prediction/estimation model learning unit 12 and the prediction/estimation unit 22 learn and calculate the prediction/estimation model 7 and the prediction/estimation value 8 using Gaussian process regression (GPR). It is possible to achieve the same effects as the prediction/estimation system 100B of the embodiment, and prediction/estimation using a nonlinear model instead of a linear model such as principal component analysis (PCA) or independent component analysis (ICA) A specialized system can be realized.

[Sixth Embodiment]
FIG. 7 is an explanatory diagram showing the basic configuration of a welding strength estimation system for laser welding to which a prediction/estimation system according to a sixth embodiment of the present invention is applied. As shown in FIG. 7, the welding strength estimation system 100E of the sixth embodiment can be configured as an estimation system to which the configurations of the prediction/estimation systems 100 to 100D according to the first to fifth embodiments can be applied.

In the estimation system 100E of the sixth embodiment, the objective variable is the welding strength (welding strength data 1A) or welding quality (not shown) of laser welding by the laser welding machine 9, and the explanatory variable is the laser welding Welding conditions (welding condition data 2A) and auxiliary variables are physical quantities observable during or after laser welding (welding temperature data 3A). Further, for example, data for prediction/estimation of explanatory variables includes welding condition data 4A set in the laser welder 9, and data for prediction/estimation of auxiliary variables is data during laser welding in the laser welder 9 or Welding temperature data 5A, which is one of the physical quantities observed after welding, can be mentioned.

The welding strength data 1A is determined by the allowable stress and material strength of the welded portion of the workpiece according to the type of the workpiece to be laser-welded based on the following welding condition data 2A by the laser welder 9, for example. It represents the strength F (N/mm ² ). Welding quality is a predetermined quality of welding (excellent, good, acceptable, unsatisfactory, etc.) based on, for example, the welding strength F obtained according to the type of workpiece to be laser-welded as described above. represents

The welding condition data 2A includes, for example, the specifications of the workpiece (material, thickness, etc.), the laser output value of the laser welder 9, the laser irradiation time, the focal position of the laser light, and pulse conditions such as the frequency and duty ratio of the laser light. , welding speed, time from start of irradiation to peak power, elapsed time from peak power to end of irradiation, lens focal length, lens focal position, laser diameter at irradiation point, fiber diameter, and spot diameter (lens, fiber, focus position combination) and is an arbitrary parameter set for laser welding. Welding temperature data 3A represents, for example, the welding temperature detected during laser welding based on welding condition data 2A.

The welding condition data 4A is data relating to welding conditions actually set in the laser welder 9, and is configured in the same manner as the welding condition data 2A. The welding temperature data 5A represents the welding temperature detected and acquired during actual laser welding detected by the welding temperature sensor 17 provided in the laser welder 9 . Welding temperature sensor 17 is included in sensor 212 described above. The welding strength data 1A, the welding condition data 2A, and the welding temperature data 3A are stored in a readable state in an external storage device or storage medium (not shown) connected to the network 90, for example. Note that the network 90 may be a medium for exchanging data, and is not limited to networks such as the Internet and LAN, but may be a bus, USB, or a portable storage medium such as an HDD that constitutes a device. There are no particular restrictions on its form.

The estimation system 100E includes a learning server 30 functioning as the learning device 10 described above and an estimation edge server 40 functioning as the prediction/estimation device 20 described above, which are communicably connected to each other via a network 90 . The estimation system 100E also includes a storage device 50 that stores (at least one of) the latent variable calculation information 6 and the welding strength estimation model 7A. In this embodiment, the learning server 30 is also provided with a storage unit 50A having the same function as the storage device 50. FIG. Although not shown, a storage unit equivalent to the storage unit 50A may be provided in the estimated edge server 40. FIG.

The learning server 30 has the teacher data of the objective variable, the teacher data of the explanatory variables acting on the objective variable, and the variables other than the objective variable, which are correlated with the explanatory variables and observed when the explanatory variables are given. It is configured to be able to input teacher data of auxiliary variables. In addition, the learning server 30 calculates latent variable teacher data (not shown) having a smaller correlation between or with the explanatory variable teacher data than the auxiliary variable teacher data from at least the auxiliary variable teacher data. , a latent variable calculation unit (first latent variable calculation unit 31) that outputs latent variable calculation information 6, teacher data of objective variables, teacher data of explanatory variables, and a latent variable calculation unit (first latent variable calculation unit 31 ) and the teacher data of the latent variables calculated by ), and perform, for example, machine learning to create a prediction/estimation model of the objective variable (welding strength estimation model 7A). and a model learning unit 32).

On the other hand, the estimated edge server 40 has the teacher data of the objective variable, the teacher data of the explanatory variables acting on the objective variable, and the variables other than the objective variable, which have correlation with the explanatory variables, and when the explanatory variables are given, From the observed auxiliary variable teacher data, the latent variable teacher data calculated based on the latent variable calculation information 6, and the target variable prediction/estimation model (welding strength estimation model 7A) learned in advance using Based on this, the prediction/estimation device 20 calculates the prediction/estimation value of the objective variable (estimated value 8A of welding strength). In addition, the estimation edge server 40 is based on the latent variable calculation information 6 from at least the data for prediction/estimation of the auxiliary variables, and the data for prediction/estimation of the explanatory variables rather than the data for prediction/estimation of the auxiliary variables. A latent variable calculation unit (second latent variable calculation unit 41) that calculates prediction/estimation data (not shown) of latent variables with low mutual correlation, and a latent variable calculation unit (second latent variable calculation unit 41) From the prediction/estimation data of the latent variables calculated by and the prediction/estimation data of the explanatory variables, the prediction/estimation value of the objective variable (welding strength and a prediction/estimation unit (welding strength estimation unit 42) that calculates the estimated value 8A) of.

In addition, since the learning server 30 and the estimation edge server 40 can function as the learning device 10 and the prediction/estimation device 20 described above, respectively, all aspects already described in the first to fifth embodiments can be realized. is possible. Therefore, hereinafter, explanations that overlap with the components of the learning device 10 and the prediction/estimation device 20 will be omitted.

[Operation of estimation system]
The welding strength data 1A, the welding condition data 2A, and the welding temperature data 3A stored in an external storage device or storage medium are input to the learning server 30 as teacher data for objective variables, explanatory variables, and auxiliary variables, respectively, and stored in the storage unit 50A. remembered. The learning server 30 calculates latent variable teacher data in the first latent variable calculator 31 based on at least the welding temperature data 3A among the input data. In addition, the learning server 30 inputs the welding strength data 1A, the welding condition data 2A, and the teacher data of the latent variables calculated by the first latent variable calculation unit 31 in the welding strength estimation model learning unit 32. For example, Machine learning is performed to create the welding strength estimation model 7A. The learning server 30 stores the latent variable calculation information 6 and the welding strength estimation model 7A determined when calculating latent variable teacher data in the storage unit 50A, and outputs them to the storage device 50 via the network 90. The storage device 50 stores the input latent variable calculation information 6 and welding strength estimation model 7A. The welding strength data 1A, the welding condition data 2A, and the welding temperature data 3A may be stored, for example, in the storage device 50 instead of the external storage device or storage medium.

On the other hand, in the laser welder 9, laser welding is performed based on the set welding condition data 4A. Welding temperature data 5A during laser welding is acquired. The welding condition data 4A of the laser welder 9 and the welding temperature data 5A from the welding temperature sensor 17 are input to the estimation edge server 40 as explanatory variable prediction/estimation data and auxiliary variable prediction/estimation data, respectively. The latent variable calculation information 6 and the welding strength estimation model 7A stored in the storage device 50 are also input to the estimation edge server 40 .

The estimated edge server 40 stores these input data in the storage unit, and from the data including at least the welding temperature data 5A, the second latent variable calculation unit 41 calculates the latent variable calculation information 6. Calculate data for predicting and estimating latent variables. In addition, the estimation edge server 40 uses the welding condition data 4A and the latent variable prediction/estimation data calculated by the second latent variable calculation unit 41 in the welding strength estimation unit 42, based on the welding strength estimation model 7A. , to calculate the estimated weld strength 8A. The calculated prediction/estimation data of the latent variables and the estimated value 8A of the weld strength are stored in the storage unit, and the estimated value 8A of the weld strength is output from the estimated edge server 40 in a form that can be appropriately used such as display and printing. can be output.

In the estimation system 100E of the sixth embodiment, the latent variable teacher data is calculated based on the welding temperature data 3A, which is the auxiliary variable teacher data of the welding strength data 1A, and then the welding strength is estimated using the latent variable teacher data. A model 7A is created. Further, based on the latent variable calculation information 6 determined when calculating the teacher data of the latent variables, the data for prediction/estimation of the latent variables is calculated from the welding temperature data 5A, which is the data for prediction/estimation of the auxiliary variables. Then, an estimated welding strength value 8A is calculated based on the welding strength estimation model 7A from the welding condition data 4A, which is explanatory variable prediction/estimation data, and the latent variable prediction/estimation data.

Therefore, according to the estimation system 100E of the sixth embodiment, the welding strength of the estimation target (welded portion of the workpiece) can be non-destructively estimated from the data regarding the welding conditions of laser welding and the data regarding the temperature during welding. becomes possible. Generally, in laser welding, even if laser welding is performed under the same welding conditions, the welding strength varies depending on factors such as the material of the workpiece and variations in the welding environment, so the product must be destroyed. However, there is a problem that the weld strength cannot be measured. However, the estimation system 100E of the sixth embodiment adopts a configuration that can improve prediction/estimation accuracy using more data and avoid deterioration of prediction/estimation accuracy due to correlation, so this problem is solved. It is possible to solve the problem and achieve the same effects as those of the first to fifth embodiments.

The estimation system 100E of the sixth embodiment aims at estimating the welding strength of laser welding, but is not limited to this, and predicts and estimates various physical phenomena related to welding quality such as spatter and porosity. is also possible. In this case, the objective variable should be the welding quality of laser welding. Also, although the welding temperature data 3A and 5A are used as examples of auxiliary variables, the auxiliary variables are not limited to these. It may be a physical quantity observed after welding, such as the weld width, color, and number of spatters of the welded portion after welding. The estimation system 100E can install the learning server 30, the estimation edge server 40, and the storage device 50, for example, centrally at the same site as the installation location of the laser welding machine 9, or separately at a remote location. A wide variety of operations are possible, assuming global deployment, regarding the estimation of welding strength.

[Seventh Embodiment]
FIG. 8 is an explanatory diagram showing the basic configuration of a learning device of a nondestructive measurement system to which the prediction/estimation system according to the seventh embodiment of the present invention is applied. FIG. 9 is an explanatory diagram showing the basic configuration of a nondestructive testing device of the nondestructive measurement system. As shown in FIGS. 8 and 9, the nondestructive measurement system 100F of the seventh embodiment is configured as a nondestructive measurement system to which the configurations of the prediction/estimation systems 100 to 100D according to the first to fifth embodiments can be applied. can be

The nondestructive measurement system 100F includes a learning device 33 as shown in FIG. 8 and a nondestructive testing device 49 including a prediction/estimation device 43 as shown in FIG. The learning device 33 includes a first latent variable calculation unit 34 and a prediction/estimation model learning unit 35 . The prediction/estimation device 43 of the nondestructive testing device 49 includes a second latent variable calculator 44 and a prediction/estimation unit 45 .

In the nondestructive measurement system 100F, the objective variable is the concrete strength (strength property data 1B), and the explanatory variable is the measurable concrete strength influence factor/condition (influence factor/condition data 2B). , the auxiliary variable is the elastic wave propagation characteristic (elastic wave propagation characteristic data 3B) at the time of intensity measurement. Further, for example, explanatory variable prediction/estimation data includes impact factor/condition data 4B in concrete for strength prediction/estimation, and auxiliary variable prediction/estimation data includes a measuring device (acoustic sensor 18 ), the elastic wave propagation characteristic data 5B based on the data acquired in ). The strength property data 1B, the strength influence factor/condition data 2B, and the elastic wave propagation characteristic data 3B are input to the learning device 33, and the strength influence factor/condition data 4B and the elastic wave propagation characteristic data 5B are input to the nondestructive testing apparatus. 49 is input to the prediction/estimation device 43 .

The strength property data 1B can include, for example, compressive strength, tensile strength, bending strength, shear strength, bearing strength, adhesion strength, fatigue strength, etc. as data representing the strength of concrete. represents F (N/mm ² ). The strength influence factor/condition data 2B includes, for example, various types of information such as cement type, cement water ratio, cement additive, material age, temperature and humidity from concrete placement to strength measurement. However, it is not limited to these.

The elastic wave propagation characteristic data 3B represents the elastic wave propagation velocity, the elastic wave reflection characteristic, etc., but is not limited to these. are applicable. The strength influence factor/condition data 4B is data relating to the strength influence factor/condition of the concrete whose strength is predicted/estimated, and is configured in the same manner as the strength influence factor/condition data 2B input to the learning device 33. . The elastic wave propagation characteristic data 5B represents the elastic wave propagation characteristic derived from the acoustic data detected by the acoustic sensor 18, and is configured similarly to the elastic wave propagation characteristic data 3B input to the learning device 33. FIG. Note that the acoustic sensor 18 is included in the sensor 212 described above.

[Operation of non-destructive measurement system]
The strength property data 1B, the strength influence factor/condition data 2B, and the elastic wave propagation characteristic data 3B are input to the learning device 33 as teacher data for objective variables, explanatory variables, and auxiliary variables, respectively. The learning device 33 calculates teacher data of latent variables in the first latent variable calculator 34 based on data including at least intensity influence factor/condition data 2B and elastic wave propagation characteristic data 3B among these input data. . In addition, the learning device 33 receives the strength property data 1B, the strength influence factor/condition data 2B, and the teacher data of the latent variables calculated by the first latent variable calculation unit 34 in the prediction/estimation model learning unit 35. Then, for example, machine learning is performed to create a concrete strength prediction/estimation model 7 . The learning device 33 outputs to the outside the latent variable calculation information 6 and the concrete strength prediction/estimation model 7 that are determined when calculating latent variable teacher data.

On the other hand, the non-destructive testing device 49 is provided with the latent variable calculation information 6 and the prediction/estimation model 7 output from the learning device 33 . The latent variable calculation information 6 and the prediction/estimation model 7 are stored in a storage medium possessed by the nondestructive testing device 49 or in a portable storage medium detachable from the nondestructive testing device 49, or can be transmitted via the Internet or the like. It is obtained from a medium and provided in the non-destructive testing device 49 .

The non-destructive testing device 49 is provided with elastic wave propagation characteristic data 5B derived based on the acoustic data detected by the acoustic sensor 18 when concrete is struck. Then, the prediction/estimation device 43 of the nondestructive testing device 49 receives the latent variable calculation information 6, the prediction/estimation model 7, the elastic wave propagation characteristic data 5B, and the strength influence factor/condition data 4B.

The prediction/estimation device 43 converts the data including at least the strength influence factor/condition data 4B and the elastic wave propagation characteristic data 5B out of the input data into the latent variable calculation information 6 in the second latent variable calculation unit 44. Calculate data for prediction/estimation of latent variables based on The prediction/estimation device 43 also includes data for prediction/estimation of the latent variables calculated by the second latent variable calculation unit 44 in the prediction/estimation unit 45, strength influence factor/condition data 4B, elastic wave propagation characteristics Based on the data 5B and the concrete strength prediction/estimation model 7, a concrete strength prediction/estimation value 8 is calculated. The calculated predicted/estimated value 8 can be output from the non-destructive testing device 49 in a form that can be used as appropriate, such as display or printing.

In the non-destructive measurement system 100F of the seventh embodiment, a latent variable After calculating the teaching data of the latent variables, a concrete strength prediction/estimating model 7 is created using the teaching data of the latent variables. Also, based on the latent variable calculation information 6 determined when calculating the latent variable teacher data, elastic wave propagation characteristic data 5B that is data for prediction/estimation of auxiliary variables and data for prediction/estimation of explanatory variables. After calculating latent variable prediction/estimation data from strength influence factor/condition data 4B, concrete Based on the strength prediction/estimation model 7, a concrete strength prediction/estimation value 8 is calculated.

Therefore, according to the non-destructive measurement system 100F of the seventh embodiment, data related to factors and conditions affecting the strength of concrete representing factors at the time of placing and the like and elastic wave propagation characteristics in a non-destructive test using the elastic wave method From the data, it becomes possible to predict and estimate the strength of the estimation target (concrete) non-destructively. Therefore, it is possible to improve the prediction/estimation accuracy using more data and avoid deterioration of the prediction/estimation accuracy due to correlation, and it is possible to achieve the same effects as the first to fifth embodiments. becomes. When data other than elastic wave propagation characteristic data 3B and 5B are used as auxiliary variables, various sensors or the like corresponding to the auxiliary variables are used instead of the acoustic sensor 18 to obtain the auxiliary variables. Also good.

[Eighth Embodiment]
FIG. 10 is an explanatory diagram showing the basic configuration of a quality judgment/prediction system for cutting quality of a laser cutting apparatus to which the prediction/estimation system according to the eighth embodiment of the present invention is applied. As shown in FIG. 10, the quality judgment prediction system 100G of the eighth embodiment can be configured as a quality judgment prediction system to which the configurations of the prediction/estimation systems 100 to 100D according to the first to fifth embodiments can be applied.

The quality judgment prediction system 100G includes a learning device 36 and a laser cutting device 60. The learning device 36 includes a first latent variable calculation unit 37 and a quality judgment prediction model learning unit 38 . The laser cutting device 60 includes a quality judgment prediction device 46 and a data storage/transfer device 61 . The quality judgment prediction device 46 includes a second latent variable calculator 47 and a quality judgment prediction unit 48 .

In the quality judgment prediction system 100G, the objective variable is the laser cutting quality judgment result (quality judgment result data 1C) by the laser cutting device 60, and the explanatory variable is the laser cutting cutting condition (cutting condition data 2C). , and the auxiliary variable is a measurable physical quantity related to laser cutting (emission intensity data 3C). Further, for example, explanatory variable prediction/estimation data includes cutting condition data 4C set in the laser cutting device 60, and auxiliary variable prediction/estimation data is the physical quantity observed by the laser cutting device 60. (luminescence intensity data 5C).

Here, the quality judgment 59 of the cutting quality of the laser cutting device 60 means, for example, the physical state of the cut surface of the workpiece cut by the laser cutting device 60, that is, for example, surface roughness, dross adhesion state, gouging ( It is to automatically or manually determine whether or not the state of occurrence of defective cutting) can withstand practical use in using the workpiece. Dross means a deposit of metals, oxides, etc. melted and adhered to the lower surface of the cut material, and is synonymous with slag. Further, gouging in laser cutting means, for example, a state in which the laser beam does not penetrate during laser cutting and melted metal is ejected onto the surface of the material, resulting in a dirty appearance.

The quality judgment result data 1C is provided as discrete category information such as "optimum", "excellent", "good", "acceptable", and "impossible" as a result of the quality judgment 59 as described above. , or may be given as numerical information from 0 to 100, representing the determined laser cutting quality. The cutting condition data 2C can include various information such as the material and/or product name of the workpiece to be laser-cut, the laser output value, the laser irradiation time, and the cutting speed. Cutting conditions to be set in the device 60 are shown. Also, the cutting condition data 2C may include various kinds of information on environmental conditions such as temperature and humidity during laser cutting, but the various kinds of information in the cutting condition data 2C are not limited to these.

The emission intensity data 3C represents, for example, the emission intensity during piercing performed prior to laser cutting. However, the teaching data of the auxiliary variable is not limited to the light emission intensity, and may be physical quantity data that can be measured in association with laser cutting, such as the sound during piercing or the light emission spectrum. The cutting condition data 4C is, for example, data relating to the cutting condition set in accordance with the user's operation input of the cutting condition via the touch panel 211, and is configured in the same manner as the cutting condition data 2C. The emission intensity data 5C represents the emission intensity detected by the optical sensor 19 provided in the laser cutting device 60 during actual piercing. Note that the optical sensor 19 is included in the sensor 212 described above.

[Operation of quality judgment prediction system]
Of the quality determination result data 1C, the cutting condition data 2C, and the luminous intensity data 3C, at least the cutting condition data 2C and the luminous intensity data 3C are input to the learning device 36 as teacher data for explanatory variables and auxiliary variables, respectively. The quality judgment result data 1C can be input to the learning device 36 as teacher data for objective variables. The learning device 36 calculates teacher data of latent variables in the first latent variable calculator 37 based on the data including at least the cutting condition data 2C and the light emission intensity data 3C among the input data. In addition, the learning device 36 inputs the quality determination result data 1C, the cutting condition data 2C, and the teacher data of the latent variables calculated by the first latent variable calculation unit 37 in the quality determination prediction model learning unit 38, For example, machine learning is performed to create a quality determination predictive model 7B for cutting quality of laser cutting. The learning device 36 outputs to the laser cutting device 60 the latent variable calculation information 6 and the quality judgment prediction model 7B determined when calculating the teacher data of the latent variables.

The laser cutting device 60 is provided with the latent variable calculation information 6 and the quality judgment prediction model 7B output from the learning device 36 . The latent variable calculation information 6 and the quality judgment prediction model 7B are stored in a storage medium of the laser cutting device 60 or a portable storage medium detachably attached to the laser cutting device 60, or can be downloaded from an information transmission medium such as the Internet. It is acquired and provided in the laser cutting device 60 .

Further, the data storage/transfer device 61 of the laser cutting device 60 is supplied with the emission intensity data 5C during the piercing process detected by the optical sensor 19 and the cutting condition data 4C operated by the touch panel 211 and temporarily stored. ing. The latent variable calculation information 6 and the quality judgment prediction model 7B are input to the quality judgment prediction device 46 of the laser cutting device 60 together with the cutting condition data 4C and the emission intensity data 5C stored in the data storage/transfer device 61. be.

The quality judgment predicting device 46 uses at least the cutting condition data 4C and the luminescence intensity data 5C among the input data to generate latent variable prediction/estimation data based on the latent variable calculation information 6 in the second latent variable calculator 47. Calculate In addition, the quality judgment prediction device 46 predicts the quality from the latent variable prediction/estimation data calculated by the second latent variable calculation unit 47 in the quality judgment prediction unit 48, the cutting condition data 4C, and the light emission intensity data 5C. Based on the judgment prediction model 7B, quality judgment prediction of cutting quality is performed to calculate a quality judgment prediction value 8B. The calculated quality determination predicted value 8B can be output from the laser cutting device 60 in a form that can be used as appropriate, such as display or printing. Since the laser cutting device 60 can output various information about the actual cutting result 58, the quality judgment 59 described above is performed based on the various information about this cutting result 58, and the quality judgment result data 1C is created. .

Further, the learning device 36 converts the cutting condition data 4C and the light emission intensity data 5C stored in the data storage/transfer device 61 into explanatory variables and auxiliary variables when calculating teacher data of latent variables and creating the quality judgment prediction model 7B. may be used in addition to the cutting condition data 2C and the emission intensity data 3C as teaching data. In this way, more data can be used to contribute to improving the accuracy of prediction/estimation. Furthermore, since the learning device 36 can additionally use the quality determination result data 1C as teacher data for the objective variable, in this case, it is possible to further improve the accuracy of prediction/estimation.

In the quality judgment prediction system 100G of the eighth embodiment, latent light emission intensity data 3C (5C), which is teacher data for auxiliary variables of quality judgment result data 1C, and cutting condition data 2C (4C), which is teacher data for explanatory variables. After calculating the teacher data of the variables, the quality judgment prediction model 7B of the cutting quality using the teacher data of the latent variables is created. Further, based on the latent variable calculation information 6 determined when calculating the supervised data of the latent variables, the luminous intensity data 5C, which is the data for prediction/estimation of the auxiliary variables, and the cutting condition, which is the data for prediction/estimation of the explanatory variables After calculating latent variable prediction/estimation data from data 4C, cutting condition data 4C, luminous intensity data 5C, and latent variable prediction/estimation data are used to determine the quality of cutting quality based on quality judgment prediction model 7B. A judgment predicted value 8B is calculated.

Therefore, according to the quality judgment prediction system 100G of the eighth embodiment, the quality prediction target (the cut portion of the workpiece) is determined before laser cutting from the data regarding the cutting conditions for laser cutting and the data regarding the emission intensity during cutting. Cut quality can be predicted. Generally, in laser cutting, even when laser cutting is performed under the same cutting conditions, cutting quality varies depending on factors such as the material of the workpiece and variations in the cutting environment. There is a problem that the cutting quality cannot be determined unless the cutting is performed. However, the quality judgment prediction system 100G of the eighth embodiment adopts a configuration that can improve the quality judgment prediction accuracy using more data and avoid deterioration of the quality judgment prediction accuracy due to correlation. This problem can be solved by predicting the cutting quality before cutting, and the same effects as those of the first to fifth embodiments can be obtained.

In the quality judgment prediction system 100G of the eighth embodiment, the learning device 36 may be incorporated in the laser cutting device 60, and the quality judgment prediction device 46 may be a separate device from the laser cutting device 60. (For example, an edge server or the like) may be independent. For machine learning and quality judgment prediction in the learning device 36 and the quality judgment prediction device 46, the algorithms include regression analysis (RA), principal component analysis (PCA), singular value decomposition (SVD), and linear discriminant analysis as described above. (LDA), independent component analysis (ICA), Gaussian process latent variable model (GPLVM), Logistic Regression (LR), Support Vector Machine (SVM), Discriminant Analysis (DA) ), Random Forest (RF), Ranking Support Vector Machine (RSVM), Gradient Boosting (GB), Naive Bayes (NB), K-Nearest Neighbor Algorithm : K-NN) can be used.

[Other embodiments]
In the above-described first embodiment, the learning method of the prediction/estimation model 7 and the prediction/estimation method of the prediction/estimation value 8 in the learning device 10 and the prediction/estimation device 20 may be a program for executing these methods and/or this It can be implemented as a computer-readable storage medium storing a program. In addition, the learning method and the prediction/estimation method may be realized using various hardware resources such as an electronic circuit that executes these methods and a logic circuit that uses other physical media. and/or may be implemented as a device having hardware resources such as logic circuits. In addition, the learning method and the prediction/estimation method are realized using various hardware resources such as a computer and/or electronic circuit equipped with a storage medium storing a program etc. to be executed as one system, and other logic circuits. Alternatively, they may be realized using various hardware resources such as one or more computers and/or electronic circuits each having a different storage medium, and other logic circuits. In the case of execution as one system, for example, the learning method may be implemented on the cloud side and the prediction/estimation method may be implemented on the edge side, or vice versa.

[Example]
FIG. 11 is a graph showing the relationship between welding temperature and elapsed time for explaining auxiliary variables in an example in which the prediction/estimation system according to the fourth embodiment shown in FIG. 5 is applied to actual laser welding. FIG. 12 is a graph showing the relationship between estimated values and measured values and the frequency distribution of errors of estimated values with respect to estimated values in Comparative Example and Example.

In the laser welding of the example, welding conditions (output intensity and irradiation time of laser used for laser welding) were used as explanatory variables, and welding strength was used as an objective variable. In the learning device 10, the prediction/estimation model 7 is created using multiple regression analysis. In the comparative example, no latent variable as described above was used.

As shown in FIG. 11, according to a graph in which the vertical axis represents the welding temperature at the welding point of the laser welding and the horizontal axis represents the elapsed time of the laser welding, the temperature at the welding point varies from the start of the laser irradiation at the elapsed time T1. At the same time, it begins to rise sharply. Then, at a certain point after the elapsed time T1 (elapsed time T2), the temperature rise slows down, and at the same time the laser irradiation ends after the elapsed time T2 (elapsed time T3), the temperature drops sharply, and after the elapsed time T3 At a certain point (elapsed time T4), the temperature returns to the temperature before the start of laser irradiation. The welding temperature of laser welding changes with such characteristics. Therefore, as auxiliary variables for estimating the welding strength, the temperature TP1 at the time when the temperature rise slows down (elapsed time T2) and the temperature TP2 at the end of laser irradiation (elapsed time T3) are used.

The estimated value of the welding strength and the measured value of the actual welding strength according to the comparative example are shown in FIG. 12(a). On the other hand, the estimated value of the welding strength according to the example and the measured value of the actual welding strength are shown in FIG. 12(b). In FIGS. 12(a) and (b), the horizontal axis represents the measured values of the welding strength when the values of the welding conditions are changed to various conditions, and the vertical axis represents the estimated values of the welding strength corresponding to the welding conditions. represents. In FIGS. 12A and 12B, the plots are normalized so that the maximum welding strength is 1 for both the measured values and the estimated values.

Also, the frequency distribution of the errors of the estimated values with respect to the measured values according to the comparative example is shown in FIG. 12(c). On the other hand, the error frequency distribution of the estimated values with respect to the measured values according to the example is shown in FIG. 12(d). In FIGS. 12C and 12D, the horizontal axis represents the error range, and the vertical axis represents the frequency corresponding to the error.

Then, the error in the comparative example and the error in the example were evaluated by mean absolute error (MAE) and root mean squared error (RMSE), and the results are shown in Table 1 below. became.

The errors according to the example took smaller values in both MAE and RMSE than the errors according to the comparative example. Also, as is clear from FIGS. 12(c) and (d), the estimated values are closer to the measured values in the example than in the comparative example, and the tail (range) of the error distribution is narrower. I understand. Therefore, in the prediction/estimation system of the example, it was proved that the accuracy of prediction/estimation was improved even when more data including auxiliary variables were used.

Although several embodiments of the present invention have been described above, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, replacements, and modifications can be made without departing from the scope of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are included in the scope of the invention described in the claims and equivalents thereof.

1 Objective variable training data 2 Explanatory variable training data 3 Auxiliary variable training data 4 Explanatory variable prediction/estimation data 5 Auxiliary variable prediction/estimation data 6 Latent variable calculation information 7 Prediction/estimation model 8 Prediction/estimation Value 10 learning device 11 first latent variable calculation unit 12 prediction/estimation model learning unit 13 latent variable teacher data 20 prediction/estimation device 21 second latent variable calculation unit 22 prediction/estimation unit 23 latent variable prediction/estimation data 100 Prediction/estimation system

Claims (12)

teacher data of an objective variable, teacher data of an explanatory variable acting on the objective variable, and auxiliary variables other than the objective variable that are correlated with the explanatory variable and observed when the explanatory variable is given. a learning device that inputs teacher data of variables and creates a prediction/estimation model of the objective variable;
input data for prediction/estimation of the explanatory variable for predicting/estimating the objective variable and prediction/estimation data for the auxiliary variable for predicting/estimating the objective variable; a prediction/estimation device that calculates a prediction/estimation value of the objective variable based on the created prediction/estimation model;
with
The learning device
a first latent variable calculation unit that calculates latent variable teacher data from at least the auxiliary variable teacher data and outputs latent variable calculation information;
The teacher data of the objective variable, the teacher data of the explanatory variable, and the teacher data of the latent variable calculated by the first latent variable calculation unit are input to create a prediction/estimation model of the objective variable. a prediction/estimation model learning unit;
including
The prediction/estimation device is
a second latent variable calculation unit that calculates prediction/estimation data of a latent variable based on the latent variable calculation information from at least the prediction/estimation data of the auxiliary variable;
Prediction/estimation of the objective variable based on the prediction/estimation model from the prediction/estimation data of the latent variables calculated by the second latent variable calculation unit and the prediction/estimation data of the explanatory variables a prediction/estimation unit that calculates a value;
Forecasting and estimating systems, including The first latent variable calculation unit of the learning device calculates teacher data of the latent variables based on teacher data of the explanatory variables and teacher data of the auxiliary variables,
The second latent variable calculation unit of the prediction/estimation device calculates prediction/estimation data of the latent variables based on the prediction/estimation data of the explanatory variables and the prediction/estimation data of the auxiliary variables. The prediction/estimation system according to claim 1. The second latent variable calculation unit of the prediction/estimation device further predicts/estimates the latent variables based on the teacher data of the explanatory variables, the teacher data of the auxiliary variables, and the teacher data of the latent variables. The prediction/estimation system according to claim 2, wherein the data for the calculation is calculated. The latent variable calculation information includes algorithms, formulas and/or parameters,
The prediction/estimation system according to any one of claims 1 to 3, wherein the algorithm includes dimensionality reduction or dimensionality compression processing. The algorithm comprises at least one of regression analysis (RA), principal component analysis (PCA), singular value decomposition (SVD), linear discriminant analysis (LDA), independent component analysis (ICA), and Gaussian process latent variable model (GPLVM). The prediction/estimation system of claim 4, comprising: The first latent variable calculation unit of the learning device includes a first regression analysis unit that performs regression analysis on the teacher data of the auxiliary variables with the teacher data of the explanatory variables to create a regression model, and regression by the first regression analysis unit. a first principal component analysis unit that performs principal component analysis on the regression error of the analysis to calculate teacher data for the latent variable;
The latent variable calculation information includes parameters and regression equations of the regression model and principal component vectors and eigenvalues of the principal component analysis,
The second latent variable calculation unit of the prediction/estimation device performs regression analysis on the prediction/estimation data of the auxiliary variables with the prediction/estimation data of the explanatory variables to create a regression model based on the latent variable calculation information. a second regression analysis unit, and a second principal component analysis unit for calculating prediction/estimation data of the latent variables by performing principal component analysis on the regression error by the second regression analysis unit based on the latent variable calculation information. The prediction/estimation system according to claim 2. The first latent variable calculation unit of the learning device calculates teacher data of the latent variables by a Gaussian process latent variable model (GPLVM),
The latent variable calculation information includes hyperparameters of the Gaussian process latent variable model (GPLVM),
The prediction/estimation model learning unit of the learning device creates a prediction/estimation model of the objective variable by Gaussian process regression (GPR),
The second latent variable calculation unit of the prediction/estimation device calculates data for prediction/estimation of the latent variable by the Gaussian process latent variable model (GPLVM),
The prediction/estimation system according to claim 2, wherein the prediction/estimation unit of the prediction/estimation device calculates the prediction/estimation value of the objective variable by the Gaussian process regression (GPR). The objective variable is the welding strength or welding quality of laser welding by a laser welder,
The explanatory variable is a welding condition of the laser welding,
The prediction/estimation system according to any one of claims 1 to 7, wherein the auxiliary variable is a physical quantity observable during or after laser welding with respect to the laser welding. The objective variable is the strength of concrete,
The explanatory variable is a measurable factor/condition affecting the strength of the concrete,
The prediction/estimation system according to any one of claims 1 to 7, wherein the auxiliary variable is elastic wave propagation characteristics when measuring the intensity. The objective variable is a quality judgment result of laser cutting by a laser cutting device,
The explanatory variable is a cutting condition for the laser cutting,
the auxiliary variable is a measurable physical quantity related to the laser cutting;
The prediction/estimation system according to any one of claims 1 to 7. teacher data of an objective variable, teacher data of an explanatory variable acting on the objective variable, and auxiliary variables other than the objective variable that are correlated with the explanatory variable and observed when the explanatory variable is given. Enter the variable teacher data and
a latent variable calculation unit that calculates latent variable teacher data from at least the auxiliary variable teacher data and outputs latent variable calculation information;
A prediction/estimation model for creating a prediction/estimation model for the objective variable by inputting the teacher data for the objective variable, the teacher data for the explanatory variable, and the teacher data for the latent variable calculated by the latent variable calculator. an estimation model learning unit;
A learning device with teacher data of an objective variable, teacher data of an explanatory variable acting on the objective variable, and auxiliary variables other than the objective variable that are correlated with the explanatory variable and observed when the explanatory variable is given. Predicted/estimated value of the objective variable based on the predictive/estimated model of the objective variable learned in advance using the latent variable tutored data calculated based on the latent variable calculation information from the variable tutor data A prediction/estimation device that calculates
a latent variable calculation unit that calculates prediction/estimation data of the latent variable based on the latent variable calculation information from at least the prediction/estimation data of the auxiliary variable;
Based on the prediction/estimation model, the prediction/estimation value of the objective variable is calculated from the prediction/estimation data of the latent variable calculated by the latent variable calculation unit and the prediction/estimation data of the explanatory variable. a prediction/estimation unit that calculates;
Prediction/estimation device with

Images