JP2020184111A - Learning program, learning method and learning device - Google Patents

Abstract

To reduce deterioration of prediction accuracy.SOLUTION: A quality prediction device calculates a first rate of abnormal values which shows a rate at which abnormal products are manufactured in a testing process testing production of a manufacture. The quality prediction device calculates a second rate of abnormal values which shows a rate at which abnormal products are manufactured on the basis of the manufacture manufactured during prescribed time after starting the manufacture in a mass production process manufacturing the manufacture. The quality prediction device calculates similarities of the testing process and the mass production process on the basis of the first rate of abnormal values and the second rate of abnormal values. The quality prediction device alters the rate which is utilized for learning data causing a learning model to learn among testing data acquired at the testing process based upon the similarity.SELECTED DRAWING: Figure 1

Description

The present invention relates to learning programs, learning methods and learning devices.

In the manufacturing industry, there are situations where judgment based on empirical rules affects quality. For this reason, as a visualization of judgments based on empirical rules, efforts are being made to acquire IoT (Internet of Things) data of manufacturing equipment, machine-learn quality trends, and generate quality models for judging quality. ..

For example, a quality prediction technique for predicting quality from machining data by calculating the correlation between machining data obtained from a machining device and quality data obtained from an inspection device is known. In this quality prediction technique, a learning model such as a linear model that predicts quality from machining data and represents the correlation between machining data and quality data is generated in a preliminary test of a mass production process. Then, the inspection process is reduced by inputting the processing data into the learning model generated in advance and determining whether the product is non-defective or defective.

Japanese Unexamined Patent Publication No. 2017-215832 Japanese Unexamined Patent Publication No. 2018-156415 JP-A-2017-130100

However, in the above technique, when the learning model learned from the data of the test process in advance is applied to the mass production process, the quality may not be accurately predicted, and the prediction accuracy is not good.

Specifically, since the processing environment such as warm-up operation time and air temperature differs between the test process and the mass production process, the learning model generated in the test process may not be suitable in the mass production process. In addition, since the data acquired in an unstable operating environment such as the existence of manufacturing variations and human judgments and operations is used as learning data, high quality judgment accuracy can be achieved by simply learning the data by machine learning. I can't expect it.

For example, in order to obtain IoT data for learning that reflects only specific conditions, a learning model is generated using the data acquired with the production line in the test process stopped. In this case, the data acquired in the mass production process is the data acquired in the continuous operation state, and since the acquisition environment is different between the training data and the prediction target data (mass production data), it is generated from the training data having different properties. High accuracy cannot be obtained even if the learned model is applied to mass production data.

The difference between the test process and the mass production process is that the influence of warm-up is a candidate, but it is not realistic because it takes too much man-hours to investigate the cause.

In one aspect, it is an object of the present invention to provide a learning program, a learning method and a learning device capable of reducing deterioration of prediction accuracy.

In the first proposal, the learning program causes the computer to perform a process of calculating a first outlier ratio, which indicates the percentage of anomalous products manufactured in the test process of testing the manufacture of the product. The learning program tells the computer a second outlier that indicates the percentage of anomalous products produced based on the product manufactured within a predetermined time from the start of production in the mass production process of manufacturing the product. The process of calculating the ratio is executed. The learning program causes a computer to execute a process of calculating the similarity between the test process and the mass production process based on the first abnormal value ratio and the second abnormal value ratio. The learning program causes the computer to execute a process of changing the ratio of the test data obtained in the test step to be used for the learning data to be trained by the learning model based on the similarity.

On one side, the deterioration of prediction accuracy can be reduced.

FIG. 1 is a diagram illustrating a quality prediction device according to the first embodiment. FIG. 2 is a functional block diagram showing a functional configuration of the quality prediction device according to the first embodiment. FIG. 3 is a diagram showing an example of information stored in the test data DB. FIG. 4 is a diagram showing an example of information stored in the model list DB. FIG. 5 is a diagram illustrating a method for detecting an abnormal value ratio. FIG. 6 is a diagram for explaining the pattern A of the similarity determination. FIG. 7 is a diagram for explaining the pattern B of the similarity determination. FIG. 8 is a diagram for explaining the pattern C of the similarity determination. FIG. 9 is a diagram illustrating the determination of the application model. FIG. 10 is a diagram illustrating the extraction of learning data. FIG. 11 is a flowchart showing the overall processing flow. FIG. 12 is a flowchart showing the flow of the abnormality ratio calculation process. FIG. 13 is a flowchart showing the flow of the index calculation process. FIG. 14 is a diagram illustrating an example when machine learning with randomness is adopted. FIG. 15 is a diagram illustrating a hardware configuration example.

Hereinafter, examples of the learning program, learning method, and learning device disclosed in the present application will be described in detail with reference to the drawings. The present invention is not limited to this embodiment. In addition, each embodiment can be appropriately combined within a consistent range.

[Explanation of quality prediction device]
FIG. 1 is a diagram for explaining the quality prediction device 10 according to the first embodiment. As shown in FIG. 1, the quality prediction device 10 generates a learning model (prediction model) for predicting the quality of a product generated in a factory such as a manufacturing industry, and uses the trained learning model to measure the quality. It is a computer device that predicts.

Specifically, the quality prediction device 10 learns a learning model by using test data obtained in a test process of manufacturing a product, quality inspection, or the like in advance as learning data. Then, the quality prediction device 10 inputs the mass production data about the product manufactured in the mass production process in which the product is actually manufactured after the test is completed into the learned learning model, and determines the quality of the product manufactured in the mass production process. Predict.

The reasons for carrying out the test process are as follows. For example, to confirm the behavior of the grinding machine. Even with the same type of processing machine, there are machine differences due to variations when each processing machine is manufactured, and a test process is carried out in order to know the characteristics of each. In addition, this is to deliberately create a rare case (corner case) in the mass production process and confirm the behavior at that time. In addition, since there are cases that cannot be fully considered only by the simulation on the desk, the behavior is confirmed when the grinding conditions are changed to improve the productivity and reduce the cost.

In general, the environment of the test process and the environment of the mass production process are different, such as whether the manufacturing equipment is warmed up, the room temperature, the climate, and the operating time of the manufacturing equipment. Therefore, the learning model learned from the test process data is , Overfit to the test process. Therefore, when the learning model learned from the data of the test process is applied to the test process, the quality prediction accuracy may deteriorate.

Therefore, the quality prediction device 10 according to the first embodiment calculates the similarity between the test data of the test process and the mass production data of the mass production process, and learns the learning model by using the amount of test data according to the similarity. By doing so, overfitting to the test process is suppressed.

For example, in the test process, the processing device generates processing data that is data related to the test environment, and the inspection device generates quality data such as the ratio of abnormal products to the products generated in the test process. Similarly, in the mass production process, the products are mass-produced for a certain period of time, processing data related to the mass production environment is generated, and quality data such as the ratio of abnormal products in the products generated in the mass production process is generated.

Then, the quality prediction device 10 calculates the similarity between the quality data of the test process and the quality data of the mass production process. Subsequently, the quality prediction device 10 extracts learning data of a ratio based on the similarity from the learning data composed of the processing data and the quality data of the test process, and learns the learning model using the extracted learning data. To do. Here, the quality prediction device 10 executes learning with the processing data as an explanatory variable and the quality data as an objective variable.

After that, the quality prediction device 10 applies the learning model learned after the ratio control of the learning data is executed to the mass production process. That is, the quality prediction device 10 inputs the data of the product manufactured by the processing device in the mass production process into the trained learning model, and predicts the quality of the mass production process based on the output result. As a result, overfitting to the test process can be suppressed, and deterioration of prediction accuracy can be reduced.

[Functional configuration]
FIG. 2 is a functional block diagram showing a functional configuration of the quality prediction device 10 according to the first embodiment. Since each commonly used device can be used as the processing device and the inspection device in each process, detailed description thereof will be omitted. As shown in FIG. 2, the quality prediction device 10 includes a communication unit 11, a storage unit 12, and a control unit 20.

The communication unit 11 is a processing unit that controls communication between other devices, such as a communication interface. For example, the communication unit 11 receives machining data from a machining device in a test process or a mass production process, receives inspection data from an inspection device in each process, and transmits a learned learning model, a prediction result, or the like to an administrator terminal. ..

The storage unit 12 is a storage device that stores various data, various programs executed by the control unit 20, and the like, such as a memory and a hard disk. The storage unit 12 stores the test data DB 13, the mass production data DB 14, the model list DB 15, and the learning result DB 16.

The test data DB 13 is a database that stores processing data and quality data generated in the test process. FIG. 3 is a diagram showing an example of information stored in the test data DB 13. As shown in FIG. 3, the test data DB 13 stores "product, processing data, quality data" in association with each other.

The "product" stored here is information that identifies the product manufactured in the test process, such as a product name and a product ID. The "machining data" is data indicating the environment of the test process, such as cutting speed and temperature. The "quality data" is data indicating the quality of the product generated in the test process, such as the flatness and the squareness of the product. The data stored here is data measured by a processing device or an inspection device in the test process, and can be generated by each device or can be generated by an administrator or the like.

The mass production data DB 14 is a database that stores processing data and quality data generated in the mass production process. Since the stored information is the same as that in FIG. 3, detailed description thereof will be omitted. The data stored here is data measured by a processing device or an inspection device in a mass production process, and can be generated by each device or can be generated by an administrator or the like.

The model list DB 15 is a database that stores information used for determining the usage ratio of learning data. Specifically, the model list DB 15 stores a model list that can specify the ratio of the test data generated in the test process to be used as learning data, based on the similarity between the test process and the mass production process.

FIG. 4 is a diagram showing an example of information stored in the model list DB 15. As shown in FIG. 4, the model list DB 15 stores "Metalix, similarity rank, model (learning%, test%)" in association with each other. The “Metalix” stored here is information indicating an index of the degree of similarity between the test data and the mass production data. The "similarity rank" is information that identifies the degree of similarity between the test data and the mass production data. "Model (learning%)" is information indicating the ratio of all test data to be used for learning, and "model (test%)" is used to verify a learned learning model among all test data. It is information which shows the ratio to do.

In the example of FIG. 4, when "Metalix" is "0.4", the similarity rank is 4, and it is defined that 60% of the test data is used as training data and 40% is used as verification data. There is. The information defined here is defined in advance by the user or the like. Further, in the example of FIG. 4, an example of defining a similarity rank of 10 levels has been described, but this is just an example, and the setting can be arbitrarily changed. In addition, each ratio is just an example, and the setting can be changed arbitrarily.

The learning result DB 16 is a database that stores the learning results. For example, the learning result DB 16 stores various parameters of the learning model learned by the learning unit 24, which will be described later. By using various parameters stored here, a trained learning model can be constructed.

The control unit 20 is a processing unit that controls the entire quality prediction device 10, and is, for example, a processor. The control unit 20 includes an abnormality calculation unit 21, an index calculation unit 22, a rank determination unit 23, a learning unit 24, and a quality prediction unit 25. The abnormality calculation unit 21, the index calculation unit 22, the rank determination unit 23, the learning unit 24, and the quality prediction unit 25 are examples of electronic circuits included in the processor and examples of processes executed by the processor.

The abnormality calculation unit 21 is a processing unit that calculates the abnormality ratio of the manufactured product for each of the test process and the mass production process. Specifically, the abnormality calculation unit 21 calculates the ratio of abnormal values (abnormal products) in the quality data obtained in the test process for each parameter of the quality data. Similarly, the abnormality calculation unit 21 calculates the ratio of abnormal values (abnormal products) in the quality data obtained in the mass production process for each parameter of the quality data. Then, the abnormality calculation unit 21 stores the calculation result in the storage unit 12 or outputs it to the index calculation unit 22.

For example, the anomaly calculation unit 21 can use a commonly used method such as anomaly detection by gamma distribution or anomaly detection assuming a mixed normal distribution. Here, as an example, a parameter value is used as a threshold value. An example of performing anomaly detection will be described by comparing.

Anomaly detection in the test process will be described with reference to FIG. FIG. 5 is a diagram illustrating a method for detecting an abnormal value ratio. As shown in FIG. 5, the abnormality calculation unit 21 is based on a plurality of parameters such as flatness and squareness of the quality data corresponding to M products (M is a number of 1 or more) generated in the test process. Select one parameter (parameter A). Then, the abnormality calculation unit 21 determines whether or not the value of the parameter A corresponding to each product is equal to or greater than the threshold value, and counts the number of products having the parameter A which is equal to or greater than the threshold value and is determined to be abnormal.

Here, when the number of products that is equal to or greater than the threshold value is counted as Ma, the abnormality calculation unit 21 calculates the abnormal value ratio of the parameter A as "abnormal number / total number = Ma / M". In this way, the abnormality calculation unit 21 executes the above-described processing for each parameter included in the quality data to calculate the abnormality value ratio. That is, when the abnormality calculation unit 21 has P parameters (P is a number of 1 or more) in the quality data, the abnormality calculation unit 21 calculates the abnormal value ratio for each of the P parameters.

Further, the abnormality calculation unit 21 calculates the ratio of abnormal values of each parameter by the same processing for the quality data of the products manufactured in the mass production process. In the mass production process, in order to measure the similarity with the test process, the production line is operated for a predetermined time such as 10 minutes, and quality data is collected.

The index calculation unit 22 is a processing unit that indexes the degree of similarity between the test process and the mass production process. Specifically, the index calculation unit 22 compares the test process with the mass production process based on the abnormal value ratio of each parameter in the test process and the abnormal value ratio of each parameter in the mass production process calculated by the abnormality calculation unit 21. Calculate the result (diff).

For example, the index calculation unit 22 calculates "diff" for each parameter by clustering using the mixed normal distribution shown in the equation (1). After that, the index calculation unit 22 divides the total value obtained by adding the diffs of each parameter by the number of parameters (N: N is a number of 1 or more) using the formula (2), thereby performing the test process and mass production. An index "Metalix" indicating the degree of similarity with the process is calculated. Then, the index calculation unit 22 outputs the calculated index “Metalix” to the rank determination unit 23.

The rank determination unit 23 is a processing unit that determines the similarity rank based on the result calculated by the index calculation unit 22. Specifically, the rank determination unit 23 refers to the model list DB 15 and specifies the similarity rank corresponding to the “Metirx” input from the index calculation unit 22. Then, the rank determination unit 23 outputs the specified similarity rank as the determined rank to the learning unit 24.

The learning unit 24 is a processing unit that executes machine learning and generates a learning model based on the ratio corresponding to the similarity rank determined by the rank determination unit 23. For example, when the similarity rank “6” is accepted, the learning unit 24 refers to the model list DB 15 and specifies “learning% = 40” and “test% = 60%”.

Then, the learning unit 24 extracts "40%" of the data included in the test data and determines them as learning data. After that, the learning unit 24 learns the learning model by machine learning using the processed data of the read test data as the “explanatory variable” and the quality data as the “target variable (label)”. After that, the learning unit 24 stores the learning result in the learning result DB 16. The extraction method can be arbitrarily set, such as extracting the latest 40% of the data from all the test data or randomly extracting 40% of the data.

The quality prediction unit 25 is a processing unit that executes quality prediction using a learned learning model. For example, the quality prediction unit 25 reads various parameters stored in the learning result DB 16 and generates a learning model. Then, the quality prediction unit 25 inputs the processing data of the mass production process into the trained learning model, and acquires the output result as the quality prediction result. After that, the quality prediction unit 25 displays the quality prediction result on a display unit such as a display or transmits the quality prediction result to the administrator terminal. For example, the quality prediction unit 25 acquires, as a quality prediction result, a probability or a numerical value indicating whether each parameter for specifying quality is abnormal or normal from the learned learning model. Then, the quality prediction unit 25 predicts that the quality is abnormal when one or more or a predetermined number or more of the abnormal parameters are present.

[Concrete example]
Here, using FIGS. 6 to 10, calculation of the outlier ratio, calculation of the comparison result (diff) between the test process and the mass production process, and calculation of the index "Metalix" indicating the degree of similarity between the test process and the mass production process. , A specific example of the learning method will be described.

(Pattern A)
FIG. 6 is a diagram for explaining the pattern A of the similarity determination. In FIG. 6, for each of the quality data parameters X0, X1, X2, and X3, the abnormal value threshold value (2.3), the abnormal value threshold value (2.4), and the abnormal value threshold value (0.5), which are the respective threshold values, are shown. An example of calculating the outlier ratio using the outlier threshold (-0.8) is shown.

Specifically, in the test step, the abnormal value ratio of the parameter X0 is "10%", the abnormal value ratio of the parameter X1 is "5%", the abnormal value ratio of the parameter X2 is "20%", and the abnormal value of the parameter X3. The value ratio is "3%". Further, in the mass production process, the abnormal value ratio of the parameter X0 is "10%", the abnormal value ratio of the parameter X1 is "6%", the abnormal value ratio of the parameter X2 is "15%", and the abnormal value ratio of the parameter X3 is "15%". It is "2%".

In such a state, the index calculation unit 22 has the maximum value of "10" among the abnormal value ratios of the test process and the mass production process for the parameter X0, and the difference of the abnormal value ratios of each process is "10-10 =". Since it is "0", it is calculated as "diff = 0/10 = 0". Similarly, the index calculation unit 22 has a maximum value of "6" among the abnormal value ratios of the test process and the mass production process for the parameter X1, and the difference of the abnormal value ratios of each process is "6-5 = 1". Therefore, it is calculated as "diff = 1/6 ≒ 0.17".

Similarly, the index calculation unit 22 has a maximum value of "20" among the abnormal value ratios of the test process and the mass production process for the parameter X2, and the difference of the abnormal value ratios of each process is "20-15 = 5". Therefore, it is calculated as "diff = 5/20 = 0.25". Further, the index calculation unit 22 has a maximum value of "3" among the abnormal value ratios of the test process and the mass production process for the parameter X3, and the difference of the abnormal value ratios of each process is "3-2 = 1". Therefore, it is calculated as "diff = 1/3 ≒ 0.3".

Then, the index calculation unit 22 divides the total value of the diffs of the parameters X0, X1, X2, and X3 by the number of parameters "4" to obtain "Metalix = (0 + 0.17 + 0.25 + 0.3) /. 4 = 0.18 "is calculated. As a result, the rank determination unit 23 refers to the model list and identifies the similarity rank (2), which is the lowest rank among the Metrics “0.18” and above. That is, it is specified that the mass production process of the pattern A is an environment close to the test process, and in the pattern A, learning is performed with an increased fit rate to the test process.

(Pattern B)
FIG. 7 is a diagram for explaining the pattern B of the similarity determination. In FIG. 6, similarly to FIG. 6, for each of the quality data parameters X0, X1, X2, and X3, the abnormal value threshold value (2.3), the abnormal value threshold value (2.4), and the abnormal value threshold value, which are the respective threshold values, are shown. (0.5) shows an example in which the outlier ratio is calculated using the outlier threshold (-0.8).

Specifically, in the test step, the abnormal value ratio of the parameter X0 is "10%", the abnormal value ratio of the parameter X1 is "5%", the abnormal value ratio of the parameter X2 is "20%", and the abnormal value of the parameter X3. The value ratio is "3%". Further, in the mass production process, the abnormal value ratio of the parameter X0 is "11%", the abnormal value ratio of the parameter X1 is "30%", the abnormal value ratio of the parameter X2 is "10%", and the abnormal value ratio of the parameter X3 is "10%". It is "2%".

In such a state, the index calculation unit 22 has the maximum value of "11" among the abnormal value ratios of the test process and the mass production process for the parameter X0, and the difference of the abnormal value ratios of each process is "11-10 =". Since it is "1", it is calculated as "diff = 1/11 ≒ 0.1". Similarly, the index calculation unit 22 has a maximum value of "30" among the abnormal value ratios of the test process and the mass production process for the parameter X1, and the difference of the abnormal value ratios of each process is "30-5 = 25". Therefore, it is calculated as "diff = 10/30 ≒ 0.83".

Similarly, the index calculation unit 22 has a maximum value of "20" among the abnormal value ratios of the test process and the mass production process for the parameter X2, and the difference of the abnormal value ratios of each process is "20-10 = 10". Therefore, it is calculated as "diff = 10/20 = 0.5". Further, the index calculation unit 22 has a maximum value of "3" among the abnormal value ratios of the test process and the mass production process for the parameter X3, and the difference of the abnormal value ratios of each process is "3-2 = 1". Therefore, it is calculated as "diff = 1/3 ≒ 0.3".

Then, the index calculation unit 22 divides the total value of the diffs of the parameters X0, X1, X2, and X3 by the number of parameters "4" to obtain "Metalix =" (0.1 + 0.83 + 0.5 + 0.3). ) / 4 = 0.43 ". As a result, the rank determination unit 23 refers to the model list and identifies the similarity rank (5), which is the lowest rank among the Metrics “0.43” and above.

(Pattern C)
FIG. 8 is a diagram for explaining the pattern C of the similarity determination. In FIG. C, as in FIG. 6, for each of the quality data parameters X0, X1, X2, and X3, the abnormal value threshold value (2.3), the abnormal value threshold value (2.4), and the abnormal value threshold value, which are the respective threshold values, are shown. (0.5) shows an example in which the outlier ratio is calculated using the outlier threshold (-0.8).

Specifically, in the test step, the abnormal value ratio of the parameter X0 is "10%", the abnormal value ratio of the parameter X1 is "5%", the abnormal value ratio of the parameter X2 is "20%", and the abnormal value of the parameter X3. The value ratio is "3%". Further, in the mass production process, the abnormal value ratio of the parameter X0 is "40%", the abnormal value ratio of the parameter X1 is "10%", the abnormal value ratio of the parameter X2 is "80%", and the abnormal value ratio of the parameter X3 is "80%". It is "30%".

In such a state, the index calculation unit 22 has the maximum value of "40" among the abnormal value ratios of the test process and the mass production process for the parameter X0, and the difference of the abnormal value ratios of each process is "40-10 =". Since it is "30", it is calculated as "diff = 30/40 = 0.75". Similarly, the index calculation unit 22 has a maximum value of "10" among the abnormal value ratios of the test process and the mass production process for the parameter X1, and the difference of the abnormal value ratios of each process is "10-5 = 5". Therefore, it is calculated as "diff = 5/10 = 0.5".

Similarly, the index calculation unit 22 has a maximum value of "80" among the abnormal value ratios of the test process and the mass production process for the parameter X2, and the difference of the abnormal value ratios of each process is "80-20 = 60". Therefore, it is calculated as "diff = 60/80 = 0.75". Further, the index calculation unit 22 has a maximum value of "30" among the abnormal value ratios of the test process and the mass production process for the parameter X3, and the difference of the abnormal value ratios of each process is "30-3 = 27". Therefore, it is calculated as "diff = 27/30 = 0.9".

Then, the index calculation unit 22 divides the total value of the diffs of the parameters X0, X1, X2, and X3 by the number of parameters "4" to obtain "Metalix =" (0.75 + 0.5 + 0.75 + 0.9). ) / 4 = 0.73 ". As a result, the rank determination unit 23 refers to the model list and identifies the similarity rank (8), which is the lowest rank among the Metrics “0.73” and above. That is, the mass production process of the pattern C is an environment that is not similar to the test process, and in the pattern C, learning with a reduced fit rate to the test process is performed.

(Explanation of learning process)
FIG. 9 is a diagram illustrating the determination of the application model. Here, the determination of the application model for each of the above patterns A, B, and C will be described. As shown in FIG. 9, the learning unit 24 executes learning using 80% of the test data because “2” is determined as the similarity rank in the pattern A. Further, since the learning unit 24 has determined "5" as the similarity rank in the pattern B, the learning unit 24 executes learning using 50% of the test data. Similarly, the learning unit 24 executes learning using 20% of the test data because "8" is determined as the similarity rank in the pattern C. In this way, for each pattern, the generation of the quality determination model in which the fit rate to the test process is adjusted according to the similarity rank is executed.

Here, the extraction image of the learning data will be described. FIG. 10 is a diagram illustrating the extraction of learning data. As shown in FIG. 10, in Example 1, the similarity between the test data and the mass-produced data is estimated, and when the degree of agreement of X% is estimated, the number of X% of the training data is extracted and the extracted learning. Learning using the data is performed.

The general technique is to use all the test data to create a highly accurate quality prediction model. On the other hand, in Example 1, only the amount of test data according to the similarity is used. For example, if the similarity is 80, 80% of the test data is randomly extracted and used. As a result, the prediction accuracy for the test data is lowered, but a quality prediction model that is resistant to data variation can be created.

[Processing flow]
FIG. 11 is a flowchart showing the overall processing flow. As shown in FIG. 11, when the process start is instructed, the quality prediction device 10 calculates the abnormal value ratio in the test process (S101) and calculates the abnormal value ratio in the mass production process (S102).

Subsequently, the quality prediction device 10 calculates an index using the abnormal value ratio in the test process and the abnormal value ratio in the mass production process (S103), and determines the similarity rank based on the calculated index (S104). ).

Then, the quality prediction device 10 learns the learning model using the number of training data based on the determined similarity rank (S105), and when the learning is completed, uses the learning result and the processing data of the mass production process. Then, the quality prediction of the mass production process is executed (S106).

[Abnormal rate calculation process]
FIG. 12 is a flowchart showing the flow of the abnormality ratio calculation process. This process is a process executed in S101 and S102 of FIG.

As shown in FIG. 12, the abnormality calculation unit 21 acquires each parameter included in the quality data of each process (S201). Subsequently, the abnormality calculation unit 21 extracts one parameter from the acquired parameters (S202).

Then, the abnormality calculation unit 21 calculates the abnormal value ratio of the extracted parameter (S203), and stores the calculation result in the storage unit 12 or the like (S204). After that, when the remaining parameters that have not been processed exist (S205: Yes), the abnormality calculation unit 21 repeats S202 and subsequent steps for the unprocessed parameters. On the other hand, the abnormality calculation unit 21 ends the process when the remaining parameters do not exist (S205: No).

[Index calculation process]
FIG. 13 is a flowchart showing the flow of the index calculation process. This process is a process executed in S103 and S104 of FIG.

As shown in FIG. 13, the index calculation unit 22 selects one parameter from each parameter included in the quality data of each process (S301), and acquires the ratio of abnormal values in the test process and the mass production process for the selected parameter. (S302). Subsequently, the index calculation unit 22 calculates and records a comparison value (diff) of the selected parameters (S303).

Then, when the remaining parameters that have not been calculated exist (S304: Yes), the index calculation unit 22 repeats S301 and subsequent steps for the remaining parameters.

On the other hand, when there is no remaining uncalculated parameter (S304: No), the rank determination unit 23 calculates an index (Metrix) using the comparison value (diff) of each parameter and stores it in the storage unit 12 or the like. Record (S305).

Then, the rank determination unit 23 determines the similarity rank based on the calculated index (Metrix) (S306). After that, the learning unit 24 executes learning of the learning model based on the ratio of the learning data corresponding to the similarity rank.

[effect]
As described above, the quality prediction device 10 controls the ratio of the test data used for the training data based on the similarity of the quality data, and executes the learning of the learning model. Therefore, the quality prediction device 10 can prevent the learning model from overfitting to the test process for the mass production process in which the environment is different from that of the test process. Therefore, the quality prediction device 10 mass-produces the learning model learned from the data of the test process in advance. Even when applied to the process, deterioration of prediction accuracy can be suppressed. Further, the quality prediction device 10 can suppress deterioration of prediction accuracy by generating a learning model using data of many test processes for a mass production process having a similar environment to the test process.

Therefore, the quality prediction device 10 can dynamically control the ratio of the test data used for learning according to the environment of the mass production process, and is suitable for each of a plurality of mass production processes from the data of one test process. It is possible to generate a learning model.

Although the examples of the present invention have been described so far, the present invention may be implemented in various different forms other than the above-described examples.

[Selection of learning model]
In the above embodiment, an example of learning using the test data of the ratio corresponding to the similarity rank has been described, but the present invention is not limited to this. For example, the quality prediction device 10 learns and prepares in advance a plurality of learning models using the test data of the ratio corresponding to each similarity rank. Then, the quality prediction device 10 specifies the similarity rank based on the similarity between the target mass production process and the test process, and selects a learning model corresponding to the specified similarity rank from the learning models held in advance. You can also do it. By doing so, the quality prediction device 10 can shorten the learning time and can quickly execute the prediction using the learning model suitable for the mass production process.

[Learning with randomness]
In the above embodiment, a learning model using a neural network, a linear model, or the like has been illustrated, but the present invention is not limited to this, and machine learning with randomness such as a random forest can also be used. In machine learning with randomness, even if the same learning data is used, the test data is highly biased, and randomness occurs in the selection of test data for each model construction, and the result may change significantly. That is, even if the same ratio of test data is used for learning data, the accuracy is not uniform. Therefore, the quality prediction device 10 can select a suitable learning model from a plurality of accuracy learning models.

FIG. 14 is a diagram illustrating an example when machine learning with randomness is adopted. As shown in FIG. 14, the quality prediction device 10 calculated "0.6" as an index (Metrix) from the ratio of the abnormal value in the quality data of the test process and the ratio of the abnormal value in the quality data of the mass production process. And. In this case, the quality prediction device 10 determines the similarity rank as "6", the learning ratio as "40%", and the verification ratio as "60%".

Then, the quality prediction device 10 learns a learning model by a random forest using 40% of the test data (processing data and quality data) of the test process. At this time, it is assumed that models 1, model 2, and model 3 having different accuracy are generated. Then, the quality prediction device 10 uses the remaining test data (60% remaining processing data and quality data) of the test data of the test process that was not used for learning, and is used for each of the model 1, model 2, and model 3. The accuracy verification is executed, and the accuracy of the model 1 is calculated as 90%, the accuracy of the model 2 is calculated as 80%, and the accuracy of the model 3 is calculated as 85%. Here, the accuracy is a ratio that can be correctly predicted (identified).

Then, the quality prediction device 10 calculates the average value "85" of 90%, 80%, and 85% of each accuracy, and determines the model 3 corresponding to the average value as the learning model for the mass production process. By doing so, the quality prediction device 10 can execute model selection based on the similarity even in machine learning with randomness. Not only the average value but also the maximum value and the minimum value can be adopted.

[Data, numerical values, etc.]
The data example, numerical example, display example, etc. used in the above embodiment are merely examples and can be arbitrarily changed. In addition, the input data and the learning method are just examples, and can be changed arbitrarily. The abnormality calculation unit 21 is an example of the first calculation unit and the second calculation unit, the index calculation unit 22 is an example of the third calculation unit, and the rank determination unit 23 is an example of the change unit.

[Similarity calculation]
Various known methods can be adopted for calculating the similarity, and for example, similarity detection by principal component analysis (PCA) can be adopted. By performing PCA, the parameters can be converted to different axes, and the degree of coincidence of the types of parameters included in the axes of PC1, PC2, and PC3 at this time is averaged to obtain the degree of similarity. If the data generation properties are completely the same, the parameters included in the PC1 axis, the PC2 axis, and the PC3 axis are the same. Also, only the type of variable is targeted, and the value is not considered.

[system]
Information including processing procedures, control procedures, specific names, and various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified.

Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific forms of distribution and integration of each device are not limited to those shown in the figure. That is, all or a part thereof can be functionally or physically distributed / integrated in any unit according to various loads, usage conditions, and the like.

Further, each processing function performed by each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

[hardware]
FIG. 15 is a diagram illustrating a hardware configuration example. As shown in FIG. 15, the quality prediction device 10 includes a communication device 10a, an HDD (Hard Disk Drive) 10b, a memory 10c, and a processor 10d. Further, the parts shown in FIG. 15 are connected to each other by a bus or the like.

The communication device 10a is a network interface card or the like, and communicates with another server. The HDD 10b stores a program or DB that operates the function shown in FIG.

The processor 10d reads a program that executes the same processing as each processing unit shown in FIG. 2 from the HDD 10b or the like and expands the program into the memory 10c to operate a process that executes each function described in FIG. 2 or the like. For example, this process executes the same function as each processing unit of the quality prediction device 10. Specifically, the processor 10d reads a program having the same functions as the abnormality calculation unit 21, the index calculation unit 22, the rank determination unit 23, the learning unit 24, the quality prediction unit 25, etc. from the HDD 10b or the like. Then, the processor 10d executes a process of executing the same processing as the abnormality calculation unit 21, the index calculation unit 22, the rank determination unit 23, the learning unit 24, the quality prediction unit 25, and the like.

In this way, the quality prediction device 10 operates as an information processing device that executes the quality prediction method by reading and executing the program. Further, the quality prediction device 10 can realize the same function as that of the above-described embodiment by reading the program from the recording medium by the medium reading device and executing the read program. The program referred to in the other embodiment is not limited to being executed by the quality prediction device 10. For example, the present invention can be similarly applied when another computer or server executes a program, or when they execute a program in cooperation with each other.

10 Quality prediction device 11 Communication unit 12 Storage unit 13 Test data DB
14 Mass production data DB
15 Model list DB
16 Learning result DB
20 Control unit 21 Abnormality calculation unit 22 Index calculation unit 23 Rank determination unit 24 Learning unit 25 Quality prediction unit

Claims (8)

On the computer
Calculate the first outlier ratio, which indicates the percentage of anomalous products manufactured in the test process of testing the manufacture of the product.
In the mass production process of manufacturing the product, a second abnormal value ratio indicating the ratio of abnormal products manufactured is calculated based on the products manufactured within a predetermined time from the start of production.
Based on the first abnormal value ratio and the second abnormal value ratio, the similarity between the test process and the mass production process is calculated.
A learning program characterized in that a process of changing the ratio of the test data obtained in the test step to be used for the learning data to be trained by the learning model is executed based on the similarity. By referring to the list in which the ratio is associated with each similarity, the ratio corresponding to the calculated similarity is specified.
The computer is made to execute a process of learning the learning model for predicting the quality of a product in the mass production process by extracting the specified ratio data from the test data and using the extracted data as the training data. The learning program according to claim 1, wherein the learning program is characterized in that. In the process to be learned, the learning model is used with the processing data indicating the manufacturing environment in the test step as an explanatory variable and the quality data indicating the quality of the product manufactured in the test step as an objective variable among the test data. The learning program according to claim 2, wherein the learning program is characterized by learning. Processing data indicating the manufacturing environment in the mass production process is input to the trained learning model, and
The learning program according to claim 3, wherein the computer executes a process of predicting the quality of a product in the mass production process based on the output result of the learned learning model. Hold multiple learning models trained using test data corresponding to each of the multiple ratios associated with each similarity as training data.
With reference to the list in which the ratio is associated with each similarity, the ratio corresponding to the calculated similarity is specified.
A learning model associated with a specified ratio is selected from the plurality of learning models, and the learning model is selected.
The learning program according to claim 1, wherein the computer is used to perform a process of predicting the quality of a product in the mass production process using the selected learning model. When a plurality of learned learning models are learned by machine learning with randomness using the test data of the ratio based on the similarity between the test process and the mass production process, the learning of the test data Using the remaining test data excluding the test data used for the trained training model, the accuracy of the plurality of trained training models is verified.
The learning program according to any one of claims 1 to 5, wherein one learning model is selected from the plurality of learned learning models based on the verification result, and the processing is executed by the computer. .. The computer
Calculate the first outlier ratio, which indicates the percentage of anomalous products manufactured in the test process of testing the manufacture of the product.
In the mass production process of manufacturing the product, a second abnormal value ratio indicating the ratio of abnormal products manufactured is calculated based on the products manufactured within a predetermined time from the start of production.
Based on the first abnormal value ratio and the second abnormal value ratio, the similarity between the test process and the mass production process is calculated.
A learning method characterized by executing a process of changing the ratio of the test data obtained in the test step to be used for the learning data to be trained by the learning model based on the similarity. The first calculation unit that calculates the first outlier ratio, which indicates the percentage of abnormal products manufactured in the test process for testing the manufacture of products,
A second calculation for calculating a second abnormal value ratio indicating the ratio of abnormal products manufactured based on the products manufactured within a predetermined time from the start of production in the mass production process for manufacturing the product. Department and
A third calculation unit that calculates the similarity between the test process and the mass production process based on the first abnormal value ratio and the second abnormal value ratio.
A learning device characterized by having a changing unit that changes the ratio of the test data obtained in the test step to be used for the learning data to be trained by the learning model based on the similarity.

Images