Method for analyzing a manufacturing process, analysis apparatus, analysis program, and computer-readable storage medium storing the analysis program.

Abstract

For products manufactured through multiple process groups, the manufacturing process that is thought to have contributed to the quality of each process group can be identified with high accuracy and easily. [Solution] The analysis method includes a process of obtaining two or more pre-generated machine learning models 51 that associate each of a plurality of feature quantities P with one or more of a plurality of manufacturing processes Q across one or more of a plurality of process groups G, and a process of estimating, for each type of machine learning model 51, a sequence of one or more manufacturing processes Q that is estimated to contribute to quality determination, based on a specific feature quantity Ps and two or more machine learning models 51, wherein the machine learning model 51 includes a batch model 511 that generates all manufacturing equipment 121 at once, and one or more subdivided models 512 that associate selected manufacturing equipment 121 for each process group G across one or more process groups G.

Description

Technical Field

[0001] The present disclosure relates to a method for analyzing a manufacturing process, an analysis apparatus, an analysis program, and a computer-readable storage medium storing the analysis program.

Background Art

[0002] For example, Patent Document 1 discloses a manufacturing process of a product including a characteristic adjustment step for imparting characteristic control processing conditions and a characteristic inspection step reached through at least one other step after the specific adjustment step. Here, the characteristic adjustment step is provided in the middle of a number of steps.

[0003] According to the description of Patent Document 1, in the characteristic adjustment step, by using a pre-generated learning model, optimal characteristic control processing conditions are searched from the intermediate characteristics obtained in the steps before the characteristic adjustment step.

[0004] In addition, the learning model described in Patent Document 1 takes intermediate characteristics and characteristic control processing conditions as inputs and outputs product characteristics obtained in the characteristic inspection step. By using such a learning model, it is possible to automatically and appropriately search for characteristic control processing conditions for obtaining desired product characteristics from the intermediate characteristics obtained each time.

[0005] On the other hand, Patent Document 2 discloses performing a first discrimination step and a classification step on a plurality of product data. The first discrimination step is a process of performing discrimination between good products and defective products based on a first learning model with teacher data. The classification step is a process of grouping the types of defects for the product data determined to be defective based on a cluster analysis without teacher data.

[0006] Furthermore, Patent Document 2 discloses that a second discrimination step is performed after the classification step. The second discrimination step is a process that, based on a second learning model, distinguishes between good products and defective products, and distinguishes the type of defect in each defective product.

[0007] Here, the second learning model is generated by performing machine learning using the good product data identified in the first discrimination step and the defective product data grouped by type of defect in the classification step as training data.

[0008] According to the aforementioned Patent Document 2, by performing classification using a first learning model followed by cluster analysis, it becomes possible to distinguish between good and defective products, as well as to distinguish the type of defect in each defective product. Furthermore, by generating a second learning model using the classification results from the first learning model and cluster analysis as training data, it becomes possible to perform two types of discrimination using the second learning model. [Prior art documents] [Patent Documents]

[0009] [Patent Document 1] Japanese Patent Publication No. 2002-287803 [Patent Document 2] Japanese Patent Publication No. 2019-204232 [Overview of the Initiative] [Problems that the invention aims to solve]

[0010] However, the method described in Patent Document 1 can only be used in cases where a specific characteristic adjustment process is assumed. In more general cases, simply changing the characteristic control processing conditions in a specific characteristic adjustment process may not achieve the desired product characteristics.

[0011] When faced with such problems, product manufacturers need to consider which of the multiple processes should be adjusted. However, Patent Document 1 does not disclose or suggest any such considerations.

[0012] On the other hand, Patent Document 2 discloses a method for determining the type of defect in each defective product based on a machine learning model. However, even if the type of defect is determined, the process that caused that defect is not necessarily uniquely determined.

[0013] For example, in the case of a product manufactured through a pre-processing stage where raw materials are processed and a post-processing stage where the processed materials are assembled, even if a defect is found in the product, it is not easy to determine whether the defect occurred during the pre-processing stage or the post-processing stage. Even if it were possible to determine this by inspecting the length, depth, etc., of such defects, such determination would likely require expertise, which is still inconvenient.

[0014] These inconveniences become more pronounced, for example, in products manufactured through multiple factories. In such cases, since one or more manufacturing processes are carried out at each factory, it would be advantageous to be able to identify the processes at each factory. However, such identification becomes more difficult than when analyzing a single factory.

[0015] Furthermore, the aforementioned problem is not limited to products manufactured through multiple factories, but is considered a common issue for products manufactured through multiple process groups in general.

[0016] This disclosure has been made in view of the above, and its purpose is to enable high-precision and easy identification of the manufacturing process that is considered to have contributed to the quality of a product manufactured through multiple process groups, for each process group. [Means for solving the problem]

[0017] A first aspect of this disclosure relates to a method for analyzing a manufacturing process, which is performed using a computer having a memory unit and an arithmetic unit, and a result of determining the quality of a product manufactured through a group of processes, each consisting of one or more manufacturing processes.

[0018] Furthermore, according to the first embodiment, the plurality of process groups are arranged in the order of progression of the product manufacturing stages, and at least one of the plurality of process groups is performed individually by each of the plurality of manufacturing units.

[0019] Furthermore, if, among the multiple feature quantities used for determining the quality, one or more feature quantities estimated to contribute to determining the quality are designated as specific feature quantities, and a sequence of manufacturing processes selected from one or more manufacturing processes constituting each of the multiple process groups, and each estimated to contribute to determining the quality, in accordance with the order of progression, is designated as a specific process group, then the method for analyzing the manufacturing process involves the calculation unit obtaining a pre-generated machine learning model from the storage unit so associating each of the multiple feature quantities with one or more of the multiple manufacturing processes across one or more of the multiple process groups, the calculation unit estimating the specific process groups based on the specific feature quantities and the machine learning model, and the machine learning model includes a batch model that generates all units of the manufacturing unit at once, and one or more subdivided models that associate the manufacturing units selected for each process group across one or more of the process groups.

[0020] Here, the term “manufacturing unit” is used in a broad sense. In this disclosure, “manufacturing unit” includes manufacturing lines, manufacturing plants, and manufacturers and suppliers in general, which are units capable of performing a series of manufacturing processes.

[0021] According to the first embodiment described above, the analysis method uses two or more machine learning models that associate each of the multiple features with one or more of the multiple manufacturing processes across one or more groups of processes.

[0022] Here, let's consider the appearance of a product as an example of quality. In this example, if we can determine the features that are estimated to have contributed to the decrease in appearance, such as the presence and location of scratches, we can use a machine learning model to estimate the specific manufacturing process in one process group and the specific manufacturing process in another process group connected to that process group that may have caused the scratches. This estimation using a machine learning model can be performed with high accuracy and easily, even by someone without specialized skills.

[0023] Furthermore, in the case of products consisting of numerous parts, it is likely that there are countless features that could be used for analysis. In such cases, even a highly skilled worker would find it difficult to identify the process that caused the difference in quality. Machine learning models can perform estimation with high accuracy and ease, even when there are a large number of features.

[0024] Furthermore, instead of simply using machine learning models, we estimate specific process groups using both a batch model and a sub-model, each as a machine learning model. This allows for a more multifaceted analysis and enables more accurate analysis.

[0025] For example, by using both a batch model and a subdivision model in combination, a more multifaceted analysis can be achieved when estimating a specific group of processes, such as determining whether or not a common trend appears across all manufacturing equipment 121.

[0026] Furthermore, according to a second aspect of this disclosure, the process group may be assigned a sample dataset that constitutes the training data for the machine learning model, the sample dataset is quantified for each manufacturing process and is composed of a combination of multiple sub-datasets belonging to different manufacturing units, the calculation unit generates the overall model based on the sample dataset, and the calculation unit generates the sub-models based on the sub-datasets.

[0027] According to the second embodiment described above, a whole model and a subdivided model can be generated by switching between using the entire sample dataset or using a subdivided dataset corresponding to a part of the sample dataset. This makes it easy to perform analysis using the whole model and the subdivided model as machine learning models.

[0028] Furthermore, according to the second embodiment, the fractional dataset used to generate the fractional model has fewer data points than the sample dataset used to generate the whole model. The fractional model is generated with fewer training data points than the whole model.

[0029] Therefore, by using both a batch model and a segmented model, it is possible to achieve both high-precision estimation using a batch model with a large amount of training data and multifaceted estimation using a segmented model that places emphasis on differences between manufacturing units.

[0030] Furthermore, according to a third aspect of this disclosure, the sub-dataset may include data values ​​corresponding to a manufacturing process specific to the manufacturing unit to which the sub-dataset belongs, and the sample dataset may be composed of a combination of the sub-datasets with the data values ​​corresponding to the specific manufacturing process excluded.

[0031] According to the third embodiment described above, the sample dataset used to generate the batch model has fewer manufacturing processes than the sub-dataset used to generate the sub-model. The batch model is generated by fewer factors than the sub-model.

[0032] Therefore, by using both a batch model and a sub-model in combination, it is possible to achieve both multifaceted estimation using a sub-model generated by numerous factors and general estimation using a batch model generated by factors common to all manufacturing units.

[0033] Furthermore, according to a fourth aspect of this disclosure, if, among a plurality of process groups, one process group performed at the end of the sequence is designated as the first process group, and other process groups performed before the first process group are designated as the second process group, then the batch model and the sub-model are pre-generated machine learning models that include, in addition to the association between a plurality of features and one or more manufacturing processes constituting the first process group, the association between each manufacturing process constituting the first process group and one or more manufacturing processes constituting the second process group, and the calculation unit estimates the specific process group for the batch model and the sub-model, respectively, based on the specific features and the machine learning model.

[0034] By using a machine learning model as described in the fourth embodiment, even when multiple process groups are connected, it is possible to estimate a specific manufacturing process in one process group and a specific manufacturing process in another process group connected to that process group. This enables easy and highly accurate analysis.

[0035] Furthermore, according to a fifth aspect of this disclosure, the machine learning model includes a first machine learning model based on a decision tree algorithm, and the calculation unit calculates the importance of each manufacturing process included in the first process group that corresponds to each of the plurality of features in the first machine learning model, and the importance of each manufacturing process included in the second process group that is associated with each manufacturing process in the first process group, and estimates the specific process group by determining one or more manufacturing processes for each process group based on each importance.

[0036] According to the fifth embodiment described above, the importance of each manufacturing process can be mechanically calculated for each process group, depending on the selection of the machine learning model. For example, in the case of a random forest model, it is sufficient to calculate the importance specific to that model. In this way, the scope for user input in determining a particular process group can be reduced. This enables more accurate analysis.

[0037] Furthermore, according to a sixth aspect of this disclosure, the first machine learning model may be a model constructed by sequentially generating a decision tree model composed of multiple submodels, and another decision tree model composed of multiple submodels, by performing a random forest generation process for all of the multiple target variables, where each of the multiple target variables is used as a new explanatory variable and each of the multiple explanatory variables is used as a new target variable.

[0038] According to the sixth embodiment described above, a specific group of processes is determined based on a chain of interconnected decision tree models. Since the importance level can be defined for each model, the specific group of processes can be easily estimated.

[0039] Furthermore, according to a seventh aspect of this disclosure, the machine learning model may include the first machine learning model and a second machine learning model based on a non-decision tree algorithm, and the calculation unit may perform the generation of both the batch model and the subdivision model in at least one of the first machine learning model and the second machine learning model.

[0040] According to the seventh embodiment described above, by using models constructed from different perspectives, such as whether or not they are decision tree systems, a more multifaceted analysis can be performed. This makes it possible to achieve a more accurate analysis.

[0041] Furthermore, according to an eighth aspect of this disclosure, if the estimation result of the specific process group by the batch model differs from the estimation result of the specific process group by the subdivision model, the calculation unit may notify the estimation results of the batch model and the subdivision model, respectively, and the number of samples used in the generation of the batch model and the subdivision model, respectively.

[0042] According to the eighth embodiment described above, the estimation results for each model are notified to the user without overwriting or rejecting the estimation results of one model with those of another. By also notifying the number of samples, a more comprehensive analysis can be performed, such as an analysis of the learning status of each model.

[0043] Furthermore, according to a ninth aspect of this disclosure, the quality determination result is quantified as quality data that increases or decreases depending on the quality, the manufacturing processes constituting the first and second process groups are each quantified to characterize the content of each manufacturing process, the storage unit pre-stores a first correlation coefficient showing the correlation between each of the plurality of feature quantities and the quality data, a second correlation coefficient showing the correlation between each manufacturing process in the first process group and each of the plurality of feature quantities, and a third correlation coefficient showing the correlation between each manufacturing process in the second process group and each manufacturing process in the first process group, and after estimating the specific process group using the batch model and the subdivision model, the calculation unit may display the breakdown of the specific process group using the batch model and the subdivision model, and methods for improving the specific process group to improve the quality, on a display unit capable of displaying the calculation results of the calculation unit based on the first, second, and third correlation coefficients, for each of the batch model and the subdivision model.

[0044] According to the ninth embodiment described above, it becomes possible not only to determine a specific group of processes but also to propose a quality improvement plan. This makes it possible for even unskilled workers to easily improve product quality.

[0045] Furthermore, according to a tenth aspect of this disclosure, the calculation unit may read, as the quality determination result, the determination result by a worker, or the determination result by rule-based determination using a plurality of the features as input, or the output from a pre-generated machine learning model.

[0046] As shown in the tenth aspect described above, the quality determination results used in this disclosure are not limited to the output from a machine learning model. The quality determination can be performed by a worker, and the subsequent processes can be performed by a computer as described above, thereby enabling a division of labor in the manufacturing process. This improves the usability of the analysis method.

[0047] Furthermore, according to an eleventh aspect of this disclosure, the calculation unit may acquire at least one of the following as a plurality of feature quantities: image data characterizing the product, audio data, text data, and mechanical or electrical data.

[0048] As shown in the eleventh aspect described above, the features used in this disclosure are not limited to electrical data. For example, the analysis method according to this disclosure can be performed based on text data written by a worker (e.g., text describing the condition of scratches, etc.). This improves the usability of the analysis method.

[0049] Furthermore, according to a twelfth aspect of the present disclosure, the product is a metal product having an insulating layer on its surface, the quality of the product is the rust-preventive performance of the insulating layer, and the calculation unit applies a voltage while a corrosive factor is in contact with the surface of the metal product, thereby obtaining a change in current over time caused by the voltage, and obtaining a plurality of characteristic quantities that characterize the waveform of the change over time.

[0050] As described in the twelfth aspect above, this disclosure is particularly effective for analyzing the corrosion prevention performance of metal products.

[0051] Furthermore, according to a thirteenth aspect of this disclosure, the plurality of manufacturing units may each represent different manufacturing lines, manufacturing plants, or manufacturers or suppliers, each configured to perform the process group individually.

[0052] As shown in the 13th aspect above, the manufacturing units used in this disclosure may be different manufacturing lines, different manufacturing plants, or different manufacturers or suppliers. This disclosure is applicable to various manufacturing units capable of performing a series of manufacturing processes.

[0053] Furthermore, a fourteenth aspect of this disclosure relates to a manufacturing process analysis device that uses a computer having a storage unit and an arithmetic unit, and a result of determining the quality of a product manufactured through a group of processes, each consisting of one or more manufacturing processes.

[0054] According to the 14th embodiment, the plurality of process groups are arranged in order of progression of the manufacturing stages of the product, and at least one of the plurality of process groups is performed individually by each of the plurality of manufacturing units.

[0055] Furthermore, if, among the multiple feature quantities used for determining the quality, one or more feature quantities estimated to contribute to determining the quality are designated as specific feature quantities, and a sequence of manufacturing processes selected from one or more manufacturing processes constituting each of the multiple process groups, and each estimated to contribute to determining the quality, is designated as a specific process group, then the manufacturing process analysis device includes a model acquisition means for acquiring a pre-generated machine learning model from the storage unit so as to associate each of the multiple feature quantities with one or more of the multiple manufacturing processes across one or more of the multiple process groups, and an estimation means for estimating the specific process groups according to the type of machine learning model based on the specific feature quantities and the machine learning model, wherein the machine learning model includes a batch model generated for all units of the manufacturing unit at once, and one or more subdivided models that associate the manufacturing units selected for each process group across one or more of the process groups.

[0056] Furthermore, a 15th aspect of this disclosure relates to a computer having a storage unit and an arithmetic unit, and a manufacturing process analysis program that is executed using the quality determination results of a product manufactured through a group of processes, each consisting of one or more manufacturing processes.

[0057] According to the 15th embodiment, the plurality of process groups are arranged in order of progression of the manufacturing stages of the product, and at least one of the plurality of process groups is performed individually by each of the plurality of manufacturing units.

[0058] Furthermore, if, among the multiple feature quantities used for determining the quality, one or more feature quantities estimated to contribute to determining the quality are designated as specific feature quantities, and a sequence of manufacturing processes selected from one or more manufacturing processes constituting each of the multiple process groups, and each estimated to contribute to determining the quality, in accordance with the order of progress, is designated as a specific process group, then the analysis program causes the computer to execute a process in which the calculation unit obtains from the storage unit a machine learning model that has been generated in advance so that each of the multiple feature quantities and one or more of the multiple manufacturing processes are associated with one or more of the multiple process groups, and a process in which the calculation unit estimates the specific process groups according to the type of machine learning model based on the specific feature quantities and the machine learning model, the machine learning model includes a batch model that generates all units of the manufacturing unit at once, and one or more sub-models that associate the manufacturing units selected for each process group with one or more of the process groups.

[0059] Furthermore, a sixteenth aspect of this disclosure relates to a computer-readable storage medium that stores the analysis program. [Effects of the Invention]

[0060] As explained above, according to this disclosure, for products manufactured through multiple process groups, it is possible to easily and accurately identify, for each process group, the manufacturing process that is considered to have contributed to the quality of the product. [Brief explanation of the drawing]

[0061] [Figure 1] Figure 1 is a system diagram illustrating the configuration of a manufacturing management system. [Figure 2] Figure 2 illustrates the configuration of a measuring device in a manufacturing system. [Figure 3] Figure 3 is a diagram illustrating the hardware configuration of the analysis device. [Figure 4] Figure 4 is an example of the software configuration of the analysis device. [Figure 5] Figure 5 is a flowchart illustrating the steps of the analysis method. [Figure 6] Figure 6 illustrates the change in electric current over time and the characteristic quantities obtained from its waveform. [Figure 7] Figure 7 is a flowchart illustrating the measurement process. [Figure 8] Figure 8 is a flowchart detailing the inspection data analysis process. [Figure 9] Figure 9 is a flowchart showing the details of the quality assessment process. [Figure 10] Figure 10 is a diagram illustrating the quality judgment model. [Figure 11] Figure 11 is a flowchart illustrating the feature selection process. [Figure 12] Figure 12 is an example of process data in one process group. [Figure 13] Figure 13 is a diagram illustrating the first machine learning model. [Figure 14A] Figure 14A is a diagram illustrating the first decision tree model. [Figure 14B] Figure 14B is a diagram illustrating the second decision tree model. [Figure 15] Figure 15 illustrates the sub-dataset and sample dataset. [Figure 16A] Figure 16A is a flowchart illustrating the procedure for generating a batch model. [Figure 16B] Figure 16B is a flowchart illustrating the procedure for generating a subdivision model. [Figure 17] Figure 17 is a flowchart illustrating the process for estimating a specific group of processes. [Figure 18] Figure 18 is a diagram illustrating a specific group of processes. [Figure 19] Figure 19 is a flowchart illustrating the correlation analysis process. [Figure 20] Figure 20 is an example of the first list of correlation coefficients. [Figure 21] Figure 21 is an example of the second list of correlation coefficients. [Figure 22] Figure 22 is an example of a display screen on a display. [Figure 23] Figure 23 is an example of a display screen on a display. [Figure 24] Figure 24 is a diagram illustrating the second machine learning model. [Figure 25A] Figure 25A is a schematic diagram showing the overall model. [Figure 25B] Figure 25B is a schematic diagram showing the subdivision model. [Figure 25C] Figure 25C is a schematic diagram illustrating the subdivision model. [Modes for carrying out the invention]

[0062] The embodiments of this disclosure will be described below with reference to the drawings. Note that the following description is illustrative.

[0063] <1. System Configuration> Figure 1 is a system diagram illustrating the configuration of the manufacturing management system 100 according to this disclosure. This manufacturing management system 100 is composed of multiple systems. For example, in this embodiment, the manufacturing management system 100 includes an analysis system 101 and N manufacturing systems 102.

[0064] The analysis system 101 is configured to analyze the manufacturing process Q of product W, which is manufactured through a group of multiple processes G, based on the quality determination results of the product W. In this embodiment, the quality determination of product W is performed by the computer 1 of the analysis system 101 based on the feature quantities P acquired by the analysis system 101. The group of multiple processes G referred to here consists of one or more manufacturing processes Q.

[0065] Multiple process groups G are arranged in the order Oc of the manufacturing stages of product W (see the white arrows in Figure 1). Each process group G is linked to one of the manufacturing systems 102. Each process group G is executed by one of the manufacturing systems 102. Each manufacturing system 102 advances the manufacturing of product W toward the finished product by executing the process group G linked to that manufacturing system 102.

[0066] Hereafter, the mth (1 ≤ m ≤ N) manufacturing system 102 may be referred to as the mth system 102m. In relation to this, the nth (1 ≤ n ≤ N) process group G may be referred to as the nth process group G. n This can happen (see Figure 1).

[0067] For example, the first system 1021 executes the first process group G1. The nth system 102n executes the nth process group G n The Nth system 102N is the Nth process group G N Execute this.

[0068] As an example, if we consider the metal product 201 shown in Figure 2 as product W, then the first process group G1 is considered to be the process group carried out in the steel mill, the second process group G2 is considered to be the process group carried out in the metalworking plant, and the Nth process group G Ncan be regarded as a series of processes related to the painting of the product W, for example, performed just before the completion of the metal product 201. In the case of this example, the Nth process group G N can be regarded as a process group G performed by the manufacturer, which is so-called "Tier1".

[0069] Hereinafter, for the sake of comprehensive discussion, the mth process group G m is assumed to include m i manufacturing processes Q (i ≥ 2, and m i ≥ 1). Accordingly, for example, among the plurality of manufacturing processes constituting the mth process group G m the i-th manufacturing process Q may be referred to as the "m-i-th process Q mi " (see Fig. 1).

[0070] Furthermore, each manufacturing system 102 according to the present embodiment includes one or more manufacturing facilities 121. Each manufacturing facility 121 executes the process group G associated with the manufacturing system 102 that includes the manufacturing facility 121. Each manufacturing facility 121 executes the process group G by executing a series of manufacturing processes Q that constitute the process group G associated with the manufacturing system 102.

[0071] And, in the case of a manufacturing system 102 having a plurality of manufacturing facilities 121, such as the nth system 102n in Fig. 1, the process group G associated with the manufacturing system 102 is executed by each of the plurality of manufacturing facilities 121. As a result, the same number of products W as the number of manufacturing facilities 121 are manufactured or processed simultaneously.

[0072] For example, the nth system 102n in Fig. 1 includes L manufacturing facilities 121. The nth system 102n executes the nth process group G n by each of the L manufacturing facilities 121. As a result, L products W are manufactured simultaneously or substantially simultaneously.

[0073] Thus, at least one of the multiple process groups G is performed individually by each of the multiple manufacturing equipment 121. Each of the multiple manufacturing equipment 121 is an example of a "manufacturing unit" in this embodiment.

[0074] More generally, the multiple manufacturing facilities 121 can consist of different manufacturing lines, manufacturing plants, or manufacturers or suppliers, each configured to perform process group G individually.

[0075] For example, in the nth system 102 of Figure 1, the first manufacturing equipment 1211 can be considered as equipment installed in a first manufacturing plant, a first production line, or a first supplier. Similarly, the second manufacturing equipment 1212 can be considered as equipment installed in a second manufacturing plant, a second production line, or a second supplier.

[0076] The analysis system 101 comprises a computer 1 that functions as an analysis device and a measuring device 122. The analysis system 101 analyzes the manufacturing process Q in the manufacturing system 102 using the computer 1 and the quality determination result of product W.

[0077] Here, the quality of product W is obtained by the measuring device 122. Specifically, the measuring device 122 inspects product W, which has been manufactured through multiple manufacturing processes Q, and obtains one or more feature quantities P that characterize the quality of product W. Furthermore, the computer 1 of the analysis system 101 performs the quality determination of product W based on the feature quantities P obtained by the measuring device 122.

[0078] Furthermore, the quality inspection of product W (in other words, the process of acquiring one or more feature quantities P) may be performed by the manufacturing system 102 instead of the analysis system 101. Similarly, the quality determination of product W may be performed by the manufacturing system 102 instead of the analysis system 101.

[0079] More generally, the training data for various machine learning purposes, such as the first training data D1 described later, may be data acquired by the analysis system 101 or the manufacturing system 102, or it may be data acquired at facilities different from the analysis system 101 and the manufacturing system 102, such as the development facility, research facility, and experimental facility for product W.

[0080] Although the following explanation primarily pertains to the former type of data, this disclosure may also be applied to the latter type of data. When applied to the latter type of data, the term "manufacturing conditions" as described below may be replaced with terms such as "development conditions," "research conditions," or "experimental conditions."

[0081] Furthermore, even when experimental data of product W is used in the inspection data 31 described later, not limited to the manufacturing of product W, various pieces of information that characterize the manufacturing of product W (so to speak, "experimental conditions") can be considered as manufacturing conditions R.

[0082] In this embodiment, product W is a metal product. As a metal product, product W includes steel materials used in various parts that make up a vehicle. When product W is a metal product, the analysis system 101 may determine the rust prevention performance of the metal product as a quality. In that case, the metal product 201 may have an insulating layer 203, described later, on its surface. Below, the rust prevention performance of the metal product 201 will be described as an example of quality. In the following description, the word "quality" may be replaced with "performance," such as rust prevention performance, as appropriate.

[0083] In this case, the manufacturing management system 100 can be considered a rust prevention performance management system that manages the rust prevention performance of metal products manufactured through multiple process groups G. It should be noted that even when metal products are used for product W, the inclusion of an insulating layer 203 is not mandatory.

[0084] Furthermore, if product W is a metal product, the Nth process group G is performed by the manufacturing equipment 121 of the Nth system 102. N This may include one or more manufacturing processes Q related to the formation of the insulating layer 203, such as an electrodeposition coating process, a washing process, and a drying process.

[0085] First, we will describe an example of the configuration of the measuring device 122 when a metal product 201 having an insulating layer 203 on its surface is used as product W. Then, we will explain the method for analyzing the results obtained by the measuring device 122 through a description of the computer 1 in the analysis system 101.

[0086] <2. Example of measuring device configuration> Figure 2 shows an example of the configuration of the measuring device 122 according to this embodiment. As mentioned above, the product W to be inspected by this measuring device 122 is a metal product 201 in this embodiment. The metal product 201 has a base material 202 such as a steel plate and an insulating layer 203 located on the surface of the base material 202.

[0087] Although not shown in the diagram, a chemical conversion coating may be formed on the surface of the substrate 202, and an insulating coating may be provided on that surface. In that case, an insulating layer 203, as in this embodiment, will be formed by the insulating coating.

[0088] The measuring device 122 is configured to acquire the change in current over time caused by applying a voltage to the surface of the metal product 201, which is product W.

[0089] In detail, the measuring device 122 applies a voltage to the surface of the metal product 201 while the corrosion factor 205 is in contact with it. The corrosion factor 205 may be an electrolyte material containing a supporting electrolyte such as water or sodium chloride, and a clay mineral such as kaolinite.

[0090] Specifically, the measuring device 122 includes a container 220, an electrode 221, a power supply 222, a sealing material 223, and wiring 224.

[0091] Of these elements, the container 220 is placed on the surface of the metal product 201 (for example, the surface of the insulating layer 203) via a leak-proof sealing material 223. The corrosive factor 205 is contained within the container 220 and is in contact with the surface of the insulating layer 203. The shape and material of the container 220 are not particularly limited. The container 220 has a cylindrical shape, for example, made of a resin material such as acrylic resin or epoxy resin. The term "cylindrical shape" here includes cylindrical shapes, polygonal cylindrical shapes, and other cylindrical shapes with any cross-sectional shape.

[0092] The sealing material 223 is, for example, a sheet-like sealing material made of silicone resin. When the container 220 is placed on the metal product 201, the sealing material 223 improves the adhesion between the container 220 and the insulating layer 203 and fills the gap between them. In this way, leakage of corrosive factors 205 from between the container 220 and the insulating layer 203 can be effectively suppressed.

[0093] Electrode 221 is for applying a voltage between the substrate 202 and the insulating layer 203. Electrode 221 is configured such that at least its tip is embedded in and in contact with the corrosive factor 205 in the container 220. Electrode 221 can be an electrode suitable for electrochemical measurements, for example. Electrode 221 may be, for example, a carbon electrode or a platinum electrode.

[0094] The power supply 222 is connected to the electrode 221 and the substrate 202 via wiring 224, and applies a voltage between the electrode 221 and the substrate 202. At the same time, the power supply 222 measures the change in current flowing between the electrode 221 and the substrate 202 over time in response to the applied voltage. The application of voltage and measurement of current by the power supply 222 are controlled by the measuring device 122.

[0095] The measurement data (inspection data 31) from the power supply 222 is transmitted to the computer 1 of the analysis system 101 via wireless or wired communication.

[0096] The inspection data 31 may be data plotting the detected current value against time, or, if a gradually increasing voltage is applied, data plotting the detected current value against the applied voltage value. In addition to the detected current value, applied voltage value, and measurement time, the inspection data 31 may also include information identifying the manufacturing lot of the metal product 201 that was measured and the measurement location on the metal product 201.

[0097] Note that the configuration example shown in Figure 2 is merely an example of how to analyze the rust prevention performance of a metal product (product W) having an insulating layer 203. The configuration of the measuring device 122 may change depending on the type of product W and the type of quality being analyzed.

[0098] Furthermore, the measuring device 122 is not essential in this disclosure. For example, the system may be configured to input the quality judgment results of the craftsman, along with the inspection data 31 used by the craftsman during the inspection, into the analysis system 101.

[0099] The inspection data 31 includes not only measurement data obtained by the measuring device 122, but also data obtained through methods other than measurement, such as inspection by a craftsman. The inspection data 31 in this disclosure includes one or more combinations of data in general that are obtained by inspecting the product W.

[0100] <3. Analysis System> Figure 3 is a diagram illustrating the hardware configuration of the analysis device (computer 1) according to this disclosure, and Figure 4 is a diagram illustrating the software configuration of computer 1. This computer 1 is composed of a computer comprising a CPU 3 as an arithmetic unit and RAM 7 and SSD 9 as storage units.

[0101] As illustrated in Figure 3, computer 1 includes a CPU (Central Processing Unit) 3 that controls the entire computer 1, a ROM (Read Only Memory) 5 that stores boot programs and the like, a RAM (Random Access Memory) 7 that functions as main memory, and an SSD (Solid State Drive) 9 as secondary storage. Note that an HDD (Hard Disk Drive) or the like can be used instead of the SSD 9 as secondary storage.

[0102] Of these elements, the CPU3 executes various programs. The CPU3 functions as the arithmetic unit in this embodiment. The RAM7 and SSD9 temporarily or continuously store the programs executed by the CPU3. The RAM7 and SSD9 each function as the storage units in this embodiment.

[0103] Computer 1 also includes a display 11, a VRAM (Video RAM) 13 as graphics memory for storing image data displayed on the display 11, and a keyboard 15 and mouse 17 as a human-machine interface. The keyboard 15 and mouse 17 function as input receiving units for the operator. The display 11 functions as a display unit that displays a screen based on the calculation results of the CPU 3. Furthermore, computer 1 according to this embodiment can send and receive data with external devices via a communication interface 19.

[0104] As shown in Figure 4, the program memory of the SSD9 stores the analysis program 21 according to this embodiment, as well as an OS (Operating System) and application programs (not shown).

[0105] Herein, the analysis method according to this embodiment is performed using the computer (analysis device) 1 configured as described above, the quality determination result of product W, and performs an analysis of the manufacturing process Q.

[0106] The analysis program 21 is a program coded to realize such analysis, and is configured to cause the computer 1, which acts as an analysis device, to execute each process constituting the analysis method sequentially or simultaneously. The analysis program 21 is pre-stored in a computer-readable storage medium 18. This storage medium 18 is a tangible storage medium made up of a disk medium or the like.

[0107] Specifically, the analysis program 21 according to this embodiment consists of an inspection data analysis program 211, a quality judgment program 212, a feature selection program 213, a model acquisition program 214, a specific process group estimation program 215, a correlation analysis program 216, and a machine learning program 217.

[0108] These programs are merely convenient groupings that classify the analysis program 21 by function. These programs may be merged as needed.

[0109] Furthermore, the depiction of a single computer 1 in Figures 1 and 3 is merely illustrative. Each program constituting the analysis program 21 may be executed on two or more computers 1.

[0110] In the program memory of SSD9, each program constituting the analysis program 21 is launched in response to commands input from the keyboard 15, mouse 17, etc. At that time, each program is loaded from SSD9 into RAM7 and executed by CPU3.

[0111] On the other hand, as shown in Figure 4, the data memory of the SSD9 stores the inspection data 31 to be analyzed and the feature data 33. The inspection data 31 is the data measured by the measuring device 122, as described above. The feature data 33 is a dataset generated from the inspection data 31 and composed of multiple feature quantities P used for quality determination. One feature data 33 corresponds to one inspection data 31 obtained from one inspection.

[0112] The inspection data 31 and feature data 33 are stored in the data memory in CSV format. This improves the usability of the inspection data 31 and feature data 33 in various processes such as pre-training.

[0113] The term "test data 31" as used here includes both the test data 31 used for pre-training and the test data 31 used for analysis. The test data 31 used for pre-training constitutes the various training data described later.

[0114] Similarly, the term "feature data 33" as used here includes both the feature data 33 used for pre-training and the feature data 33 to be analyzed.

[0115] It is not essential to store the pre-training test data 31 and feature data 33 on computer 1. These datasets are introduced for the purpose of explaining the pre-training process described later, and are not necessarily used in the following processes.

[0116] Furthermore, the SSD9's data memory stores one or more quality judgment models 41 and multiple machine learning models 51. Both the one or more quality judgment models 41 and the one or more machine learning models 51 are machine learning models generated through prior machine learning.

[0117] In particular, in this embodiment, all of the multiple machine learning models 51 are configured as decision tree-based machine learning models. These machine learning models 51 all exemplify the "first machine learning model" in this embodiment. Hereinafter, this will also be simply referred to as "machine learning model 51".

[0118] The machine learning model 51 includes a whole model 511 representing one type of machine learning model and sub-models 512 representing other types of machine learning models. As shown in Figure 4, in this embodiment, one or more sub-models 512 are used. Details of these machine learning models will be described later.

[0119] In the case of machine learning models 51 such as random forest models or Bayesian networks, the SSD9's data memory will store numerical data that represents the machine learning model 51 itself, such as information indicating the decision boundary and information indicating the parent or child nodes connected to each node.

[0120] On the other hand, in the case of a machine learning model 51 such as a neural network, the data memory of SSD9 will store numerical data that characterizes the input and output of the machine learning model 51, such as the connection coefficients of each layer.

[0121] Furthermore, the data memory of the SSD9 stores the first, second, and third correlation coefficient lists 791, 792, and 793, which are used to provide specific suggestions for changes and adjustments to the manufacturing process Q. In addition, the data memory of the SSD9 stores quality data 35 generated by the quality judgment program 212. These details will be described later.

[0122] In addition, various data generated by each program and process that constitute the analysis program 21, as well as the execution results of application programs, etc., are stored in the data memory of the SSD9 or in the RAM7 as main memory, as needed.

[0123] <4. Outline of the Analysis Method> Figure 5 is a flowchart illustrating the steps of the analysis method. As shown in Figure 5, the analysis method is carried out by sequentially executing the inspection data analysis process (step S1), the quality judgment process (step S2), the feature selection process (step S3), the model acquisition process (step S4), the specific process group estimation process (step S5), and the correlation analysis process (step S6).

[0124] The analysis program 21 is configured to cause computer 1 to execute these processes. Specifically, among these processes, the inspection data analysis process is carried out by CPU 3 executing the aforementioned inspection data analysis program 211.

[0125] Similarly, the quality judgment process is performed by CPU3 executing the quality judgment program 212. The feature selection process is performed by CPU3 executing the feature selection program 213. The model acquisition process is performed by CPU3 executing the model acquisition program 214. The specific process group estimation process is performed by CPU3 executing the specific process group estimation program 215. The correlation analysis process is performed by CPU3 executing the correlation analysis program 216.

[0126] The CPU 3 executes the specific process group estimation program 215, etc., thereby configuring the analysis device by the computer 1. In other words, the computer 1 functions as an analysis device comprising: inspection data analysis means for executing an inspection data analysis process; quality determination means for executing a quality determination process; feature selection means for executing a feature selection process; model acquisition means for executing a model acquisition process; specific process estimation means for executing a specific process estimation process; and correlation analysis means for executing a correlation analysis process.

[0127] The following describes the processes (measurement processes) performed prior to the analysis of product W, referring to Figures 6 and 7. After that, we will return to Figure 5 and explain in detail the analysis method related to this disclosure.

[0128] The measurement and analysis methods described below are performed periodically, for example, for each manufacturing lot of product W. A manufacturing lot generally refers to a unit of production or order. In this embodiment, a manufacturing lot can be defined by, for example, a shipping unit, a production unit, a manufacturing month / day, a lot of raw materials (e.g., paint), a point of change in raw materials, or a point of change in the process (e.g., when the cutting blade of metal product 201 is replaced). It is not mandatory to perform the process for each manufacturing lot. The measurement process may be performed on each individual product.

[0129] <5. Measurement Process> Figure 6 illustrates the change in current over time and the characteristic quantity P obtained from its waveform. Figure 7 is a flowchart illustrating the measurement process. First, as shown in step S101 of Figure 7, the measuring device 122 extracts one or more products W from each manufacturing lot. The number of products extracted is adjusted as appropriate based on the size of the manufacturing lot, the expected probability of abnormalities occurring, etc.

[0130] In the subsequent step S102, the measuring device 122 performs an inspection on each product W extracted in step S101. By inspecting each product W, the measuring device 122 acquires inspection data 31 for each product W that is inspected.

[0131] The inspection data 31 obtained by the measurement process may be used for pre-training, or for product analysis corresponding to each of the processes in Step 1, Step S2, Step S3, Step S4, and Step S5. In other words, the measurement process exemplified in Figure 7 can be used for pre-training, product analysis, and retraining.

[0132] In detail, in step S102, the measuring device 122 applies a voltage between the insulating layer 203 and the substrate 202, for example, with the corrosive factor 205 in contact with the surface of the insulating layer 203. The manufacturing system 102 measures the change in current over time caused by the applied voltage. The measurement locations include the main surface, edges, and welds of the metal product 201.

[0133] In the subsequent step S103, the measuring device 122 converts the time-series changes measured in step S102 into inspection data 31 and transmits it to the computer 1 of the analysis system 101. The transmitted inspection data 31 is stored in the RAM 7 or SSD 9 of the computer 1.

[0134] In the case of a more general product W, the inspection data 31 is not limited to changes in current over time. The inspection data 31 may consist of at least one of the following: image data characterizing the product W, audio data, text data, and mechanical or electrical data.

[0135] Image data is data obtained, for example, by imaging product W. Audio data is data obtained, for example, by recording sounds produced when product W is in operation, or when product W is struck. Text data is data related to the quality of product W, such as inspection records by craftsmen. Mechanical or electrical data includes the weight and dimensions of product W, as well as voltage and current values ​​as in this embodiment.

[0136] The following explanation will primarily focus on the case of "Inspection Data 31 = Change in Current over Time," but it can be substituted with other data as appropriate, as mentioned above.

[0137] <6. Details of the analysis method> Next, we will explain the quality determination process, which is one of the processes that make up the analysis method. The following process is performed on each manufactured product W. It may also be performed on products W extracted from each manufacturing lot.

[0138] (6-1. Test Data Analysis Process) Figure 8 is a flowchart detailing the inspection data analysis process. When the control process proceeds to step S1 in Figure 5, the CPU 3 executes the flow shown in Figure 8 sequentially, starting from step S201.

[0139] First, in step S201 in Figure 8, the CPU 3 reads inspection data 31 from RAM 7 or SSD 9, etc. In the following step S202, the CPU 3 extracts multiple types of feature quantities P from the read inspection data 31.

[0140] More specifically, CPU3 acquires multiple types of parameters as multiple feature quantities P, which characterize the time-dependent change in the current (current generated due to the applied voltage) shown in the test data 31. Here, "time-dependent change in current" in this embodiment refers to "time-dependent change in current measured as a waveform." In other words, it refers to "current value corresponding to voltage value changed over time."

[0141] More specifically, CPU3 extracts three or more parameter sets as multiple feature quantities P, including the number of current peaks Np, the height of each peak Ip, and the slope of the current S (see Figure 6).

[0142] For example, if there is no abnormality or warning sign in the insulating layer 203 of the metal product 201, when a DC voltage that increases over time (gradually increases) is applied as described above, almost no current flows until the applied voltage reaches the dielectric breakdown voltage, and once the applied voltage reaches the dielectric breakdown voltage, the current increases rapidly.

[0143] Until the applied voltage reaches the dielectric breakdown voltage, the insulating layer 203 maintains its ability to block the corrosion factor 205. As a result, almost no current flows. On the other hand, once the applied voltage reaches the dielectric breakdown voltage, the penetration of the corrosion factor 205 into the insulating layer 203 is promoted, and the corrosion factor 205 can reach the surface of the substrate 202 at the most vulnerable points of the insulating layer 203, such as areas with relatively few cross-linked resin structures. As a result, the current can increase rapidly. In other words, a rapid increase in the detected current value indicates that the corrosion-preventive performance of the insulating layer 203 has been lost due to the corrosion factor 205 reaching the surface of the substrate 202.

[0144] On the other hand, consider a case where there are localized defects in the insulating layer 203 (for example, foreign matter such as gas pins, welding spatter and slag, burrs, iron powder, or surface irregularities of the base material 202), and there are localized areas in the insulating layer 203 where the effective film thickness is small. In this case, when a gradually increasing DC voltage is applied between the electrode 204 and the base material 202, the current generated between them is expected to show a change over time, for example, as shown in Figure 6.

[0145] In other words, if a localized defect exists in the insulating layer 203, localized corrosion factor 205 infiltration occurs at the location of the defect. When the corrosion factor 205 penetrates the insulating layer 203 at a certain defect location and reaches the substrate 202, conduction occurs, and the current value increases instantaneously. At this point, if a voltage higher than the voltage at which water electrolysis occurs is applied between the electrode and the substrate 202, electrochemical reactions such as water electrolysis proceed on the surface of the substrate 202 due to the conduction. As a result, the generated gas and electrolytic products accumulate in the defect, interrupting conduction and causing the current value to decrease. In other words, if a localized defect exists, the conduction at the defect location and the subsequent interruption cause peaks in the waveform of the time-dependent change data of the detected current value to appear, indicating instantaneous increases and decreases in the current value.

[0146] When multiple defects are present, a peak of several Np corresponding to the number of defects occurs, as shown in Figure 6. Furthermore, the current value at the peak, i.e., the peak height Ip, is related to the size of the conduction path through which the current flows, i.e., the size, type, and conductivity of the defect. In addition, since conduction occurs at lower applied voltage values ​​where the thickness of the insulating layer 203 is smaller due to the defect, the applied voltage value that gives the peak is correlated with the thickness of the film at the location of the defect.

[0147] Furthermore, if there are overall defects in the insulating layer 203 (for example, low crosslinking density due to insufficient catalyst in the paint, oxide film on the surface of the substrate 202, etc.), corrosive factors 205 penetrate the entire insulating layer 203 and reach the substrate 202. As a result, a gradual increase in the current value occurs in the waveform of the time-dependent change data of the detected current value. In other words, the slope S of the waveform of the time-dependent change data when the detected current value increases is related to the insulating quality (film quality) of the insulating layer 203.

[0148] The number, size, type, conductivity, film thickness, and film quality of the above-mentioned defects are all factors that affect the corrosion prevention performance of the insulating layer 203. Therefore, in this embodiment, the number of current value peaks Np, the peak height Ip, and the slope S of the waveform of the current change over time, which are related to these factors, are obtained as characteristic quantities P that characterize the corrosion prevention performance of the insulating layer 203. Furthermore, since the likelihood of peaks occurring in the current differs depending on the location where the current change over time is measured (including the main surface, edge, and welded part of the metal product 201), positional information indicating the location where the current change over time was measured may also be obtained from the measurement data as a parameter for evaluating the corrosion prevention performance of the insulating layer 203.

[0149] Furthermore, the parameters available as feature quantities P are not limited to the number of current peaks Np, the height of each peak Ip, and the slope of the current S. Any parameter can be used as long as it is related to the rust prevention performance of the metal product 201. When dealing with other products W or other qualities (performance), more general parameters can be used.

[0150] Furthermore, as mentioned above, if image data, audio data, text data, and mechanical or electrical data are used as the inspection data 31, it is sufficient to extract features corresponding to the type of data.

[0151] To conduct a comprehensive discussion, we will assume that there are a total of N features P (where N is a natural number greater than or equal to 2). Accordingly, the Nth feature P will be referred to as "the Nth feature P". N It is sometimes referred to as "[this term]". In the case of metal product 201, the first feature P1 corresponds to the number of peaks Np, the second feature P2 corresponds to the peak height Ip, and the third feature P3 corresponds to the slope S (see Figure 6).

[0152] Subsequently, in step S203, CPU3 converts multiple feature quantities P into a dimensionless risk score that is between 0 and 1, and combines each of the converted feature quantities P into feature data 33. CPU3 stores the feature data 33 in the RAM 7 or SSD 9 of computer 1. Then, returning from the flow in Figure 8 to the flow in Figure 5, CPU3 proceeds the control process to step S2.

[0153] The feature data 33 generated in this way may be used for pre-training or for product analysis as exemplified in steps 2 to S5. In other words, the inspection data analysis process exemplified in Figure 8 can be used for both pre-training and product analysis.

[0154] In this embodiment, the risk score indicates the level of risk of abnormalities occurring in the quality (corrosion prevention performance) of the metal product 201. In other words, the higher the risk score, the more likely it is that abnormalities will occur in the metal product 201.

[0155] The inventors of this application attempted to determine the quality of a metal product 201 based on the level of risk scores corresponding to each of several feature quantities P, and the level of the average value of the multiple feature quantities P. Furthermore, instead of simply dividing the quality of the metal product 201 into two categories, "normal" or "abnormal," they aimed to achieve three or more classifications, including a gray area (intermediate state) between white (normal) and black (abnormal), where there are "signs of an impending abnormality."

[0156] However, such judgments are not easy. Therefore, we decided to use a pre-generated machine learning model in the quality judgment process (step S2) that follows the inspection data analysis process (step S1). By using a machine learning model for judgment, even those who are not highly skilled craftsmen can make judgments with a certain degree of accuracy.

[0157] In order to keep the explanation concise, we will proceed by treating each risk score and its corresponding feature P as identical.

[0158] (6-2. Quality Assessment Process) Figure 9 is a flowchart detailing the quality determination process. Figure 10 is a diagram illustrating the quality determination model 41. When the control process proceeds to step S2 in Figure 5, the CPU 3 executes the flow shown in Figure 9 sequentially, starting from step S301.

[0159] In the performance evaluation process, CPU3 determines and outputs the degree of quality of product W based on multiple features (more precisely, risk scores corresponding to each feature P) P obtained in the inspection data analysis process. The term "degree of quality" here includes not only two states, "normal" and "abnormal," but also an intermediate state, "showing signs of abnormality," as mentioned above.

[0160] In the case of metal product 201, the quality referred to here is "corrosion prevention performance." The degree of quality (hereinafter also simply referred to as "quality") is quantified so as to increase or decrease according to the degree of quality. For example, in this embodiment, the quality exemplified as corrosion prevention performance is quantified as, for example, "normal = 2," "signs of abnormality = 1," and "abnormal = 0." Hereinafter, the quantified quality will also be referred to as "quality data 35."

[0161] Specifically, in step S301 of Figure 9, the CPU3 first reads multiple features P, i.e., feature data 33, which have been scored as risk scores. In the example described above, the features P read here will be the number of peaks Np as the first feature P1, the peak height Ip as the second feature P2, and the slope S as the third feature P3.

[0162] In the following step S302, the CPU3 reads the quality judgment model 41 from the SSD9, which serves as the memory unit. As shown in Figure 20, the quality judgment model 41 is a machine learning model that associates multiple feature quantities P with the quality judgment result (quality data 35). The quality judgment model 41 outputs the "quality grade" as defined above as the "quality judgment result".

[0163] In other words, the quality judgment model 41 is a machine learning model that takes multiple feature quantities P as input and outputs quality judgment results (quality data 35), such as an estimated value of the degree of rust prevention performance. The quality judgment model 41 has been pre-trained using supervised learning.

[0164] For pre-training, the quality judgment model 41 takes feature data 33 based on inspection data 31 acquired at development facilities, experimental facilities, etc. as input. For quality judgment for process estimation processing, the quality judgment model 41 takes feature data 33 based on inspection data 31 acquired by the measuring device 122 as input.

[0165] For example, in this embodiment, the quality judgment model 41 is generated by machine learning using the first training data D1. As shown in Figure 10, the first training data D1 is composed of multiple feature data 33 and quality data 35 that indicates the quality judgment result, which has been acquired in advance for each feature data 33. A dataset composed of multiple feature data 33 and the quality data 35 associated with them is created in advance, in quantities equal to the number of samples (training data) of the first training data D1.

[0166] The quality judgment model 41 may be a nonlinear regression model or a linear regression model. The quality judgment model 41 as a nonlinear regression model may be a machine learning model based on a decision tree algorithm. The quality judgment model 41 related to a decision tree may be an RF (Random Forest) model.

[0167] The quality determination model 41, like the RF model described above, determines quality based on multiple features P, and in addition, it only needs to be able to determine the importance of each feature P in that determination (see the first importance I1 in Figure 10). This importance is a parameter given for each feature P, and it indicates the contribution of each feature P to the quality determination.

[0168] As illustrated in Figure 10, in this embodiment, the RF model M1 is used for the quality judgment model 41. When using the RF model M1, the contribution of each feature P (so-called RF importance) can be used as the importance. The importance may be determined, for example, based on the Gini coefficient of the decision trees that constitute the random forest model.

[0169] Hereinafter, among the multiple feature quantities P used to determine quality, one or more feature quantities P that are estimated to contribute to the determination of quality will be referred to as specific feature quantities Ps (see Figure 10, etc.).

[0170] The specific feature Ps is a parameter used in the specific process group estimation process (step S5). In this embodiment, the specific feature Ps is determined in the quality determination process (step S2), but this configuration is merely illustrative. The specific feature Ps can be determined in a feature selection process (step S3) or other process before the execution of the specific process group estimation process (step S5).

[0171] In the subsequent step S303, the CPU 3 determines the degree of quality (e.g., rust prevention performance) based on multiple feature quantities P and the quality judgment model 41. Specifically, the CPU 3 inputs the multiple feature quantities P read in step S301, that is, the feature quantities P obtained from the product W to be analyzed, into the trained quality judgment model 41, and the model 41 outputs an estimated result of the degree of quality. When quantified as described above, this estimated result is output as quality data 35.

[0172] In the subsequent step S304, the CPU3 stores the output result of step S303 in RAM7 or SSD9, etc. After that, the control process returns from the flow in Figure 9 to the flow in Figure 5 and proceeds to step S3. At this time, the CPU3 stores the importance of each of the multiple feature quantities P as the first importance I1 exemplified in Figure 10, together with the quality data 35 that indicates the quality judgment result.

[0173] (6-3. Feature Selection Process) Figure 11 is a flowchart detailing the feature selection process. When the control process proceeds to step S3 in Figure 5, the CPU 3 executes the flow shown in Figure 11 sequentially, starting from step S401.

[0174] Hereinafter, among the multiple feature quantities P obtained during quality determination, one or more feature quantities P that are estimated to contribute to the quality determination will be referred to as specific feature quantities Ps. In the specific examples shown in Figures 2 and 6, the specific feature quantities Ps will be one or more feature quantities P that are estimated to have contributed to the determination of rust prevention performance.

[0175] Specifically, in the feature selection process, CPU3 selects specific features Ps from among multiple features P obtained from the inspection data 31. This selection can be made based on the importance of each of the multiple features P corresponding to the quality judgment result, i.e., the degree of quality.

[0176] In more detail, CPU3 determines the importance of each of the multiple features P, and selects one or more features Ps as specific features Ps in order of their importance. When using the RF model M1, the importance can be determined based on the Gini coefficient, as mentioned above.

[0177] Alternatively, the so-called SHAP (SHapley Additive exPlanations) process may be used. In this case, the quality judgment model 41 is locally approximated around the features P based on the multiple features P obtained from the inspection data 31 and the quality judgment model 41. The model generated by approximating the quality judgment model 41 (approximation model) may be, for example, a simple model that easily explains the contribution of each feature P. Subsequently, based on the approximation model, the contribution of each feature P to the predicted value output from the approximation model is expressed using the so-called Shapley value used in cooperative game theory, etc. This contribution may be used as the importance for selecting specific features Ps. In addition, various image processing techniques for 2D images, such as gradient processing, can be applied.

[0178] In specific examples such as Figure 2, the specific feature Ps is one or more feature Ps that are estimated to have contributed to the determination of rust prevention performance. In this case, from among the three parameters—the number of peaks Np as the first feature P1, the peak height Ip as the second feature P2, and the slope S as the third feature P3—which parameter is selected to have contributed to the determination of rust prevention performance, including abnormalities and their precursors.

[0179] In the specific example shown in Figure 2, the RF model M1 is used, as illustrated in Figure 10. In this example, the specific feature Ps is composed of three elements: the number of peaks Np as the first feature P1, and the peak height Ip as the second feature P2.

[0180] Specifically, in step S401, the CPU3 reads the importance of each of the multiple features P. In the following step S402, the CPU3 selects one or more features P as specific features Ps, in order of importance.

[0181] Here, the importance of each feature P may be scored as a dimensionless number between 0 and 1, and if that score is equal to or greater than a predetermined value (for example, 0.5), it may be considered to have "high importance." Furthermore, if there are multiple features P whose scores are equal to or greater than the predetermined value, only the top three features (for example, three) with the highest scores may be selected.

[0182] Furthermore, as mentioned above, when the quality of product W is quantified as a dimensionless number between 0 and 2, a correlation coefficient can be given that shows the correlation between each feature P and the quality data 35, separate from importance (see Figure 20). In this embodiment, such correlation coefficients (hereinafter also referred to as "first correlation coefficients") are calculated in advance and stored in the data memory of the SSD9 as the first correlation coefficient list 791 shown in Figures 4 and 20.

[0183] In the subsequent step S403, the CPU3 stores the specific feature Ps selected in step S402 in RAM7 or SSD9, etc. After that, the control process returns from the flow in Figure 11 to the flow in Figure 5 and proceeds to step S4.

[0184] (6-4. Model Acquisition Process) Figure 12 is a diagram illustrating process data 39 in one process group G. Figure 13 is a diagram explaining the machine learning model 51. Figure 14A is a diagram explaining the first decision tree model 511. Figure 14B is a diagram explaining the second decision tree model 512.

[0185] In addition, Figure 15 illustrates the sub-dataset 802 and the sample dataset 801. Figure 16A is a flowchart illustrating the procedure for generating the batch model 511. Figure 16B is a flowchart illustrating the procedure for generating the sub-model 512. Furthermore, Figure 25A is a schematic diagram of the batch model 511, and Figures 25B and 25C are schematic diagrams of the sub-model 512.

[0186] During the model acquisition process, CPU3 acquires multiple types of machine learning models 51. The processing performed during the model acquisition process is, for example, carried out by CPU3 reading electronic data that characterizes each model from SSD9.

[0187] More specifically, CPU3 acquires two types of machine learning models 51: a batch model 511 and a subdivided model 512. Both the batch model 511 and the subdivided model 512 associate each of the multiple features P with one or more of the aforementioned manufacturing processes Q across one or more process groups G. Both machine learning models 51 are pre-trained.

[0188] Furthermore, each manufacturing process Q constituting each process group G is quantified in a way that characterizes the content of each manufacturing process Q. Each manufacturing process Q also contains one or more control factors. Each control factor is a factor that characterizes the content of each manufacturing process Q. These control factors are used as "manufacturing conditions R" as exemplified in Figure 12, and are also quantified as "process data 39".

[0189] As an example, a predetermined process group G includes an electrodeposition coating process as the first manufacturing process Q and a drying process using a drying oven as the second manufacturing process Q.

[0190] In this case, the former manufacturing process (electrodeposition coating process) Q includes control factors such as conductivity during electrodeposition coating (μS / cm), paint temperature, voltage value during electrodeposition, ion concentration of the electrodeposition solution (MEQ), acid concentration of the electrodeposition solution, distribution of the electrodeposition solution, amount and type of solvent in the electrodeposition solution, presence or absence of foreign matter contamination in the electrodeposition solution, and cycle time.

[0191] On the other hand, the latter manufacturing process (drying process) Q includes control factors such as the amount of moisture in the drying oven, the set temperature of the drying oven, the baking time in the drying oven, and the capacity of the drying oven.

[0192] The numerical data (process data 39) quantified in relation to each manufacturing process Q may be the control targets of each control factor, such as the target temperature of the paint. If the manufacturing process Q includes multiple control factors, the numerical data obtained by quantifying each manufacturing process Q may be numerical data corresponding to each of the multiple control factors.

[0193] Process data 39, as shown in Figure 12, is defined for each of the multiple process groups G. Each piece of process data 39 can be used for machine learning of a submodel related to the process group G corresponding to that piece of process data 39.

[0194] Here, for the machine learning model 51, among the multiple process groups G, the Nth process group G is performed last in the sequence of events. N And the (Nn)th process group G, which is performed before the Nth process group.N-n The following will be explained using (n≧1). The former is an example of the "first group of processes," and the latter is an example of the "second group of processes."

[0195] Once the batch model 511 and one or more subdivided models 512 are acquired as machine learning models 51, the control process proceeds from step S4 in Figure 5 to step S5 in the same figure.

[0196] -Overview of Machine Learning Models- The machine learning model 51 uses multiple features P and the Nth process group G. N In addition to the association with the multiple manufacturing processes Q that make up the Nth process group G N Multiple manufacturing processes Q and the (Nn)th process group G constitute the process. N-n It is pre-generated to include associations with multiple manufacturing processes Q that constitute it.

[0197] More specifically, the first machine learning model 51 uses multiple features P and the Nth process group G. N A first decision tree model 511 relates multiple manufacturing processes Q that constitute the process, and the Nth process group G N Multiple manufacturing processes Q and the (N-1)th process group G constitute the process. N-1 A second decision tree model 512 relates multiple manufacturing processes Q that constitute the process, and the N-1st process group G 1N-1 Multiple manufacturing processes Q and the (N-2)th process group G constitute the process. N-2 It is composed of N-1 decision tree models arranged in the order Oc shown in Figure 1, such as a third decision tree model 513 that associates multiple manufacturing processes Q that make up the process.

[0198] For example, if we start with a feature P, then the Nth process group G will be involved in that feature P. N The manufacturing process Q that constitutes it is linked (see Figure 13). Then, the Nth process group G N Each manufacturing process Q that makes up the process includes the N-1 process group G N-1The manufacturing process Q that makes up the first process group G1 will be linked. By linking them in this way, it is possible to trace back from the analysis system 101 located at the end of the manufacturing process to each manufacturing process Q that makes up the first process group G1.

[0199] The machine learning model 51 is pre-generated through supervised learning.

[0200] More specifically, the machine learning model 51 is generated from the second training data D2. This second training data D2 consists of numerical data (process data 39) corresponding to each manufacturing process Q for each process group G, which has been acquired in advance for each of the multiple samples, and numerical data (feature data 33) corresponding to multiple features P, which has also been acquired in advance for each of the multiple samples.

[0201] Specifically, the machine learning model 51 according to this embodiment is a model based on a decision tree algorithm. An example of a decision tree-based machine learning model is an RF model. When using an RF model, the contribution of each manufacturing process Q (so-called RF importance), which is obtained for each process group G, can be used as the importance, as described later. The example in Figure 13 can be interpreted as a model in which multiple RF models are nested together.

[0202] Thus, a method for constructing a machine learning model 51 in which multiple RF models are nested together has already been systematized by the inventors of the present invention. Details of this systematization are disclosed, for example, in Japanese Patent Application Publication No. 2023-57729.

[0203] The model described in the aforementioned publication has multiple explanatory variables (Nth process group G N The process of generating a submodel using a random forest, consisting of the manufacturing process Q) that constitutes the process and one of the multiple target variables (e.g., feature P), is performed for all of the multiple target variables to generate a first decision tree model 511 composed of multiple submodels.

[0204] After generating the first decision tree model 511, multiple explanatory variables (the Nth process group G) are generated. N Another variable (the N-1st process group G) that is connected to the manufacturing process Q that constitutes it N-1 The manufacturing process Q) that constitutes the process is used as a new explanatory variable, and multiple explanatory variables (process group N G) are used. N When each of the manufacturing processes Q) that constitute the process is set as a new target variable, another decision tree model 512 is further generated by generating multiple submodels. Thus, the machine learning model (first machine learning model) 51 in this embodiment is a model constructed by sequentially generating multiple decision tree models.

[0205] Note that each decision tree model does not connect all variables. Instead, it connects variables with high importance (e.g., the top three) based on the impurity of the submodels. The impurity of the submodels represents the degree of error improvement before and after branching of each decision tree. Impurity can also be expressed as, for example, the decrease in the Gini coefficient (Gini impurity).

[0206] Here, the first decision tree model 511 is pre-machine-trained using the second training data D2 as the training data, as shown in Figure 14A, and the Nth process group G N This is a machine learning model that takes numerical data (process data 39) corresponding to each manufacturing process Q that makes up the system as input, and outputs a single feature P.

[0207] By generating the first decision tree model 511, the Nth process group G N It is possible to determine one or more manufacturing processes Q that contributed to the output of a particular feature P from among the multiple manufacturing processes Q that constitute the feature P. This selection can be made based on the importance of each of the multiple manufacturing processes Q (see the second importance I2 in Figure 13A). Specifically, the importance of each of the multiple manufacturing processes Q is determined, and one or more manufacturing processes Q are selected in order of their importance, from highest to lowest.

[0208] Here, the second decision tree model 512 is pre-machine-trained using the second training data D2 as the training data, as shown in Figure 14B, and the N-1st process group G N-1 Numerical data (process data 39) corresponding to each manufacturing process Q that constitutes the Nth process group G is taken as input. N This is a machine learning model that outputs numerical data (process data 39) corresponding to each manufacturing process Q that makes up the system.

[0209] By generating the second decision tree model 512, the N-1 step group G N-1 Of the multiple manufacturing processes Q that make up the Nth process group G N It is possible to determine one or more manufacturing processes Q that contributed to each manufacturing process Q that constitutes the N-1 process group G. N-1 This can be done based on the importance of each of the multiple manufacturing processes Q in the process (see the second importance level I2 in Figure 13B). Specifically, the N-1st process group G N-1 The importance of each of the multiple manufacturing processes Q is determined, and the N-1st process group G is selected in order of importance from highest to lowest. N-1 One or more manufacturing processes Q will be selected in this process.

[0210] As an example, in the case of Figure 13, the Nth process group G N Among the manufacturing processes Q belonging to this category, the manufacturing process Q with high importance for a single feature P is the NA process Q. N And, the NB process Q NB And, the first NC process Q NC And so it is. Process Group N-1 G N-1 Among the manufacturing processes Q belonging to this process, the NA process Q N The manufacturing process Q with high importance is the (N-1)-I process Q. (N-1)-I And, the (N-1)-J process Q (N-1)-J That is the case.

[0211] Similarly, for other features P, models consisting of multiple nested RF models are generated in advance. By doing this for all features P, a machine learning model 51 can be generated that associates each of the multiple features P with one or more of the previous manufacturing processes Q across one or more of the process groups G, as shown in Figure 13.

[0212] By following the arrows in Figure 13, paths can be formed connecting manufacturing processes Q. This allows, for example, when focusing on a specific feature Ps, to extract manufacturing processes Q with high importance for that feature Ps from each process group G. In this embodiment, each manufacturing process Q connected by such paths is included in the specific process group Gs described later.

[0213] The machine learning model 51, when given a feature P as input, outputs one or more manufacturing processes Q linked to that feature P based on the RF model, for each process group G. This output is none other than the specific process group Gs described later. The output from the machine learning model 51 corresponds to a sequence of manufacturing processes Q with high importance for the feature P input to the model 51.

[0214] In other words, the machine learning model 51 according to this embodiment can be considered as a map or model that takes a single feature P as input and outputs a sequence of manufacturing processes Q selected from each process group G that have high importance for that feature P (see the specific process group Gs in Figure 18). This property is common to both the overall model 511 and each sub-model 512.

[0215] Furthermore, as mentioned above, when the quality of product W is quantified as a dimensionless number between 0 and 2, a correlation coefficient can be obtained that shows the correlation between each manufacturing process Q and each feature P, separate from importance (see Figure 20). Note that for manufacturing processes Q that can only be represented as categorical data, the correlation coefficient is undefined, as shown for the third process Q3 in Figure 21.

[0216] In this embodiment, such correlation coefficients (hereinafter also referred to as "second correlation coefficients") are calculated in advance and stored in the data memory of the SSD9 as the second correlation coefficient list 792 shown in Figures 4 and 20.

[0217] Furthermore, as shown only in Figure 4, a correlation coefficient can be obtained that shows the correlation between each manufacturing process Q in one process group G and each manufacturing process Q in other process groups G that are connected to it.

[0218] In this embodiment, such correlation coefficients (hereinafter also referred to as "third correlation coefficients") are calculated in advance and stored in the data memory of the SSD9 as the third correlation coefficient list 793 shown in Figure 4. The third correlation coefficient list 793 can be set for each process group G. For example, the third correlation coefficient list 793 in this embodiment consists of (N-1) lists, each of which is stored in the data memory of the SSD9.

[0219] -Bulk Model- The batch model 511 is a machine learning model that generates all units of a manufacturing unit, that is, all equipment of the multiple manufacturing facilities 121 exemplified in Figure 1, all at once (see Figure 25A). The second training data D2 is handled in a special way during the generation of the batch model 511.

[0220] In generating the batch model 511, the second training data D2 consists of the sample dataset 801 exemplified in Figure 15. This sample dataset 801 is assigned to each process group G. As shown in the example figure, one sample dataset 801 can be assigned to one manufacturing system 102.

[0221] The sample dataset 801 is composed of multiple sub-datasets 802 (two in the example figure). Each of the multiple sub-datasets 802 is a set of datasets that have been acquired in advance for each of the multiple samples, quantified by manufacturing process Q, and each belongs to a different manufacturing facility 121. As illustrated in Figure 15, one sub-dataset 802 corresponds to one manufacturing facility 121. Each sub-dataset 802 is a set of numerical data (process data 39) associated with each manufacturing process Q performed at each manufacturing facility 121.

[0222] In Figure 15, one of the two rows of matrices, the upper one, shows a sub-dataset 802 corresponding to the first manufacturing facility 1211, and the lower one shows a sub-dataset 802 corresponding to the second manufacturing facility 1212.

[0223] The vertical (column) direction of each sub-dataset 802 consists of labels (sample numbers) to distinguish each sample that makes up the sub-dataset 802. On the other hand, the horizontal (horizontal) direction of each sub-dataset 802 consists of labels (manufacturing process Q numbers) to distinguish each manufacturing process Q that makes up the sub-dataset 802. Each cell that makes up each sub-dataset 802 is assigned a value of process data 39 obtained for each sample and each manufacturing process Q.

[0224] The sub-dataset 802 is a dataset that collects numerical data (process data 39) associated with each manufacturing process Q, while assigning it to each process group G and distinguishing between multiple manufacturing equipment 121 belonging to that process group G.

[0225] On the other hand, the sample dataset 801 is a dataset that collects numerical data (process data 39) associated with each manufacturing process Q, without assigning it to each process group G or distinguishing between multiple manufacturing equipment 121 belonging to that process group G.

[0226] For example, the Nth process group G NThe assigned sample dataset 801 can be used as the training data D2 on the input (explanatory variable) side to generate the first decision tree model 511 exemplified in Figure 13A.

[0227] Similarly, the Nth process group G N The assigned sample dataset 801 can be used as the training data D2 on the output (target variable) side to generate the second decision tree model 512 exemplified in Figure 13B.

[0228] Here, as shown in the nth system 102n in Figure 1, if there are multiple manufacturing equipment 121 belonging to the same process group G, it is permissible to use the process data 39 obtained from each manufacturing equipment 121 together as a sample data set 801 without distinguishing between them, as shown in Figures 15 and 25A.

[0229] In the case of the nth system 102n in Figure 1, since L manufacturing equipment 121 is provided, the number of process data 39, and consequently the number of samples (training data) in the sample dataset 801, will be approximately L times greater.

[0230] Furthermore, each sub-dataset 802 may contain data values ​​corresponding to a manufacturing process Qu specific to the manufacturing facility 121 to which the sub-dataset 802 belongs. Therefore, as illustrated in Figure 15, the sample dataset 801 according to this embodiment can be constructed by combining multiple sub-datasets 802 while excluding the data values ​​corresponding to the specific manufacturing process Qu.

[0231] During pre-training, CPU3 executes the machine learning program 217. This enables computer 1 to function as a machine learning tool. As shown in Figure 16A, CPU3 obtains the batch training data Dl from SSD9 (step S711). The batch training data Dl consists of a sample dataset 801 spanning multiple process groups G. Subsequently, CPU3 sequentially generates each decision tree model based on the sample dataset 801, and by concatenating the generated decision tree models in the order Oc, it generates the batch model 511 as shown in Figure 25A (step S712).

[0232] -Subdivision Model- On the other hand, the subdivided model 512 is one or more machine learning models generated by associating the manufacturing equipment 121 selected for each process group G across one or more process groups G (see Figure 25B). In generating the subdivided model 512, the handling of the training data D2 is improved, similar to the batch model 511.

[0233] When multiple subdivision models 512 are generated, the manufacturing equipment 121 selected for each process group G is linked across multiple process groups G, while the selection of the manufacturing equipment 121 differs (see Figure 25C).

[0234] In generating the subdivision model 512, the second training data D2 is composed of subdivision datasets 802. In constructing the second training data D2, a subdivision dataset 802 specific to the manufacturing equipment 121 selected for each process group G is used.

[0235] During pre-training, CPU3 executes the machine learning program 217. This enables computer 1 to function as a machine learning tool. Specifically, as shown in Figure 16B, CPU3 obtains fractional training data Ds from SSD9 (step S721). The fractional training data Ds consists of fractional datasets 802 selected from multiple process groups G. Subsequently, CPU3 sequentially generates decision tree models based on the fractional datasets 802 and concatenates the generated decision tree models in sequential order Oc to generate a fractional model 512 as shown in Figure 25B (step S722).

[0236] Furthermore, as shown in the comparison between Figure 25B and Figure 25C, multiple sub-models 512 can be generated by selecting different sub-datasets 802 for each process group G. Generating multiple sub-models 512 contributes to a more multifaceted analysis.

[0237] (6-5. Process for Estimating Specific Process Groups) Figure 17 is a flowchart detailing the process for estimating a specific process group. Figure 18 is a diagram illustrating the specific process group Gs. When the control process proceeds to step S5 in Figure 5, the CPU 3 executes the flow shown in Figure 17 sequentially, starting from step S501.

[0238] Hereinafter, one or more manufacturing processes Q from among multiple manufacturing processes Q that are presumed to contribute to quality determination will be referred to as specific processes Qs. Simultaneously, a sequence of one or more manufacturing processes Q in the order Oc, selected from among the one or more manufacturing processes Q constituting each of the multiple process groups G, and presumed to contribute to quality determination, will be referred to as specific process groups Gs. Specific processes Qs and specific process groups Gs are illustrated in Figure 18 below.

[0239] In the specific example shown in Figure 2, etc., the Nth process group G is presumed to have contributed to the determination of the rust prevention performance. N This is one or more manufacturing processes Q in the process. The specific processes Qs are presumed to have contributed to the determination of rust prevention performance, and are part of the Nth process group G. NIt becomes one or more manufacturing processes Q in

[0240] On the other hand, the specific process group Gs is a manufacturing process Q with high importance and / or strong dependency for a specific process Qs selected from the Nth process group G N and is a sequence of manufacturing processes Q selected from each of the first process group G1 to the Nth process group G N .

[0241] For example, as shown in FIG. 18, assume that the N-C process Q NC is selected as the specific process Qs. In that case, the specific process group Gs is, as shown in the figure, the N-C process Q NC and the (N-1)-J process Q (N-1)J and the (N-2)-M process Q (N-2)M and is a sequence including them.

[0242] More generally, the CPU 3 corresponds to each of the plurality of feature amounts P in the machine learning model 51 and the importance of each manufacturing process Q included in the first process group (the Nth process group G N ), and calculates the importance of each manufacturing process Q included in the second process group (the N-1th process group G N ) associated with each manufacturing process Q of the first process group (the Nth process group G N-1 ), respectively. Based on each importance, by determining one or more manufacturing processes Q for each process group G, the specific process group Gs can be estimated.

[0243] Note that FIG. 18 is an example based on the batch model 511 as the machine learning model 51, but the specific process group Gs can be similarly given for the case based on other types of machine learning models 51, that is, each partial model 512.

[0244] Returning to FIG. 17, in step S501, the CPU 3 reads a specific feature amount Ps preselected in the feature amount selection process from among the plurality of feature amounts P.

[0245] In the following steps S502 to S504, the CPU 3 estimates a specific process Qs and a specific process group Gs from among multiple manufacturing processes Q. This estimation is performed based on a specific feature Ps and multiple types of machine learning models 51, for each type of machine learning model 51, i.e., for the whole model 511 and the subdivided model 512.

[0246] Specifically, in step S502, the CPU3 inputs the specific feature Ps into the batch model 511. The batch model 511 outputs the manufacturing process Q that has a high contribution to the specific feature Ps, i.e., the specific process Qs, and the specific process group Gs that includes that specific process Qs.

[0247] In the following step S503, the CPU3 inputs the specific feature Ps into one or more subdivision models 512. The subdivision models 512 output the manufacturing process Q that has a high contribution to the specific feature Ps, i.e., the specific process Qs, and the specific process group Gs that includes the specific process Qs.

[0248] In step S503, the specific feature Ps may be input to all of the multiple subdivision models 512, or some of the multiple subdivision models 512 may be selected and the specific feature Ps may be input to the selected subdivision models 512.

[0249] In the subsequent step S504, the CPU3 stores the specific process Qs and specific feature Ps output for each type of machine learning model 51 into RAM7 or SSD9, respectively. After that, the control process returns from the flow in Figure 17 to the flow in Figure 5 and proceeds to step S6.

[0250] (6-6. Correlation Analysis Process) Figure 19 is a flowchart detailing the correlation analysis process. Figure 20 is an example of the first correlation coefficient list 791. Figure 21 is an example of the second correlation coefficient list 792.

[0251] When the control process proceeds to step S6 in Figure 5, the CPU 3 executes the flow shown in Figure 19 sequentially, starting from step S601. The correlation analysis process is performed after the estimation of the specific process group Gs using the batch model 511 and the subdivision model 512.

[0252] In this correlation analysis process, the CPU 3 notifies the aggregate model 511 and the subdivided model 512, respectively, of the breakdown of specific process Qs and specific process group Gs, and the control mode of the specific process group Gs for improving quality, based on the first, second, and third correlation coefficient lists 791, 792, and 793. This notification is performed, for example, by displaying the calculation results of the CPU 3 on the display 11, which is a display unit capable of displaying such results.

[0253] More specifically, the CPU 3 displays on the display 11 the breakdown of a specific process Qs and the control methods for specific process groups Gs to improve quality, in order from the specific process groups Gs that contribute to the quality. Processing regarding the display order of specific process groups Gs is performed when there are multiple specific process groups Gs.

[0254] Specifically, in step S601, the CPU3 loads one or more specific processes Qs and specific features Ps associated with each specific process Qs, according to the type of machine learning model 51.

[0255] In the subsequent step S602, the CPU3 reads the correlation coefficient (second correlation coefficient) of a specific feature Ps corresponding to each specific process group Gs, in particular the specific process Qs that constitutes the specific process group Gs, by referring to the second correlation coefficient list 792.

[0256] Also in step S602, the CPU3 selects a specific feature Ps for each specific process group Gs such that the absolute value of the second correlation coefficient is equal to or greater than a second predetermined value (for example, 0.2). If no specific feature Ps satisfying this condition exists, the CPU3 displays a message to that effect on the display 11.

[0257] In the following step S603, the CPU3 generates a union of specific feature quantities Ps selected for each specific process group Gs, and considers the specific feature quantities Ps included in that union as the feature quantities P to be controlled.

[0258] In the subsequent step S604, the CPU3 reads the correlation coefficient (first correlation coefficient) of the quality data 35 corresponding to the specific feature quantity Ps that constitute the union by referring to the first correlation coefficient list 791. This reading is performed for each specific process group Gs.

[0259] In the same step S604, the CPU3 selects a specific feature Ps from among the specific feature Ps that constitute the union such that the absolute value of the correlation coefficient (first correlation coefficient) read immediately beforehand is equal to or greater than a third predetermined value (for example, 0.2). If no specific feature Ps that satisfies this condition exists, the CPU3 displays a message to that effect on the display 11.

[0260] In the subsequent step S605, the CPU3 lists the specific feature quantities Ps selected in step S604, associating them with the corresponding specific process group Gs. This results in a list of one or more specific feature quantities Ps that are strongly correlated with the quality data 35 for each specific process group Gs.

[0261] In the subsequent step S606, the CPU3 multiplies each specific feature Ps listed in step S607 by the first correlation coefficient and the second correlation coefficient for each specific process group Gs. Hereinafter, this multiplied value will be referred to as the "effect index".

[0262] Note that this example is the Nth process group G N This is an example of controlling manufacturing process Q, which belongs to the N-1 process group G. N-1 When controlling manufacturing process Q belonging to the group, the product of the first correlation coefficient and the second correlation coefficient should be multiplied further by the third correlation coefficient read from the third correlation coefficient list 793. The appropriate effect indicator should be used depending on the group of processes G being controlled.

[0263] In the case of a specific process group Gs, for each of the specific feature quantities Ps associated with the process, an effect index will be calculated. For example, if there are two specific feature quantities Ps related to a specific process group Gs, two effect indexes will also be calculated. Further, if there are two specific process groups Gs based on the batch model 511, the calculation of the effect index will be performed for each of the two specific process groups Gs.

[0264] The larger the effect index, the greater the change in the quality data 35 when the corresponding specific process Qs is changed. By paying attention to the positive or negative sign of the multiplication value, it becomes possible to determine the control mode of the specific process Qs to improve the quality data 35.

[0265] As described above, the effect index is calculated separately for each specific feature quantity Ps for each specific process group Gs. Therefore, in step S607, the CPU 3 adds the effect indexes calculated separately for each specific feature quantity Ps. For each specific process group Gs, a total value of the effect indexes will be calculated. This total value is calculated for each manufacturing process Q that constitutes the specific process group Gs.

[0266] Thereafter, in the subsequent step S608, the CPU 3 displays the specific process group Gs and its control mode on the display 11 in order from the ones with larger absolute values of the effect index.

[0267] Note that the estimation results of the specific process group Gs by the batch model 511 and the estimation results of the specific process group Gs by the sub-model 512 may be different from each other. In this case, during the process of the flow shown in FIG. 19, the CPU 3 notifies the estimation results of the batch model 511 and the sub-model 512 respectively, and the number of samples (number of learning data) used in the generation of the batch model 511 and the sub-model 512 respectively. This notification may be performed, for example, through display on the display 11.

[0268] Here, the details of the control mode are predetermined for each manufacturing process Q of each process group G. Each manufacturing process Q may include one or more control factors, and when displaying as in step S608, the adjustment direction of the control factor may also be displayed. The control factors are as described with reference to Figure 12.

[0269] The RAM7, acting as a memory unit, pre-stores the association between increasing or decreasing each of the aforementioned control factors and increasing or decreasing the numerical data corresponding to each manufacturing process Q. Based on this stored information, increasing or decreasing one or more control factors makes it possible to control each manufacturing process Q that constitutes a specific process group Gs in a way that improves quality.

[0270] Furthermore, each control factor constituting each manufacturing process Q may be considered as an independent manufacturing process Q itself. For example, in the electrodeposition coating process shown in Figure 12, the control factor related to conductivity may be considered as one manufacturing process Q, or the control factor related to paint temperature may be considered as another manufacturing process.

[0271] In other words, instead of directly estimating each manufacturing process Q that constitutes a specific process group Gs, the CPU3 may estimate the control factors constituting each manufacturing process Q based on specific features Ps and two or more machine learning models 51, according to the type of machine learning model 51.

[0272] (6-7. Examples of displays on the display unit) Figures 22 and 23 illustrate the display screens Sc and Sc' on the display 11. For example, when the process shown in Figure 17 is completed, the CPU 3 displays on the display 11 the specific feature Ps that contributed to the selection of the specific process group Gs, the specific process group Gs selected by the batch model 511, and the specific process group Gs selected by each sub-model 512. Refer to the enclosed section C1 in Figure 22 for the specific feature Ps, the enclosed section C2 for the specific process group Gs related to the batch model 511, and the enclosed section C3 for the specific process group Gs related to the sub-model 512. In the example shown, there are two specific feature Ps, but only one specific process group Gs is represented. In a more general case, multiple specific process group Gs may be selected for each machine learning model, or multiple specific feature Ps may be selected for each specific feature Ps.

[0273] In the example shown, the specific process group Gs for the unified model 511 is the Cth manufacturing process (3rd-C process Q) in the 3rd process group G3. 3C ) and the Jth manufacturing process in the second process group G2 (second-J process Q) 2J ) and the Mth manufacturing process in the first process group G1 (1st-Mth process Q) 1M It is composed of ) and . In this example, it is set to "N=3" as shown in Figures 1 and 14, and the analysis by the analysis system 101 is performed after the third process group G3.

[0274] Furthermore, in the example shown, the specific process group Gs for the subdivision model 512 is the Cth manufacturing process (3rd-C process Q) in the 3rd process group G3. 3C ) and the G-th manufacturing process in the second process group G2 (2nd-G process Q) 2G It is composed of ) and .

[0275] Next, CPU3 generates a union of specific process groups Gs associated with the batch model 511 and specific process groups Gs associated with the subdivision model 512 (see enclosed section C4). As shown in Figure 21 below, CPU3 determines whether the union is an empty set.

[0276] CPU3 also calculates the aforementioned effectiveness indicators for each of the manufacturing processes Q that make up the specific process group Gs corresponding to the unified model 511.

[0277] In the example shown, the third-C process Q constitutes one specific process group Gs. 3C , 2nd-G process Q 2G and process 1-M Q 1M For each of these, an effectiveness index is calculated for each specific feature quantity Ps that forms the end of the specific process group Gs. If there are multiple specific process groups Gs, an effectiveness index is further calculated for each specific process group Gs.

[0278] As shown in the diagram, if one specific process group Gs is identified, the third-C process Q that constitutes the specific process group Gs is... 3C For each specific feature Ps, the first correlation coefficient and the second correlation coefficient are multiplied together. Second-G process Q 2G Regarding the third-C process Q, 3C The multiplicative value calculated for the second-G process Q is used. 2G and the 3rd-C process Q 3C The correlation coefficient between (the third correlation coefficient) is further multiplied. Process Q (Step 1-M) 1M Regarding the second-G process Q, 2G The multiplicative value calculated for the first-M process Q is used. 1M and the second-G process Q 2G The correlation coefficient between them (another third correlation coefficient) is then multiplied. After that, CPU3 aggregates the calculated effect indicators for each specific process group Gs and for each manufacturing process Q.

[0279] Subsequently, the CPU 3 displays on the display 11, in descending order of the calculated effectiveness indicators, each manufacturing process Q constituting each specific process group Gs, along with its control mode ("increase," "decrease," etc.) (see enclosed section C5). Here, if each manufacturing process Q includes multiple control factors, each control factor may be displayed separately (for example, a display indicating "increase paint temperature"). Furthermore, if the specific feature quantity Ps at the end of each specific process group Gs is, for example, a categorical value, and it is not possible to set a correlation coefficient, and therefore "increase" and "decrease," etc., then "Not estimable" may be displayed instead of "increase" and "decrease" (not shown in the diagram).

[0280] Furthermore, as shown in the enclosed sections C12 and C13 of Figure 23, if the specific process Qs differs between the two models 511 and 512, a message indicating this (for example, "mismatch") may be displayed on the screen Sc', and the number of samples (number of training data) for each model 511 and 512 may also be displayed (see enclosed section C14).

[0281] <7. Significance of the analysis method> As described above, the analysis method according to the embodiment uses a single model 511 and a sub-model 512 as machine learning models 51, which are configured to associate each of the multiple feature quantities P with one or more of the multiple manufacturing processes Q across one or more of the multiple process groups G (see Figures 13, 25A, and 25B).

[0282] Here, as an example of quality, let's consider the appearance of product W. In this example, if we can determine the features P that are estimated to have contributed to the decrease in appearance, such as the presence or absence and location of scratches, we can use the machine learning model 51 to estimate a specific manufacturing process Q in one process group G that may have caused the scratches, and a specific manufacturing process Q in another process group G that is connected to that process group G. The estimation using the machine learning model 51 can be performed with high accuracy and easily, even by someone who is not a skilled worker.

[0283] Furthermore, in the case of a product W consisting of numerous parts such as a vehicle body, it is thought that there are countless feature quantities P that can be used as candidates for analysis. In such cases, even a skilled worker would find it difficult to identify the process that caused the quality to be poor or poor. The machine learning model 51 can perform estimation with high accuracy and ease, even when there are many feature quantities.

[0284] Furthermore, instead of simply using the machine learning model 51, the specific process group Gs is estimated using both the unified model 511 and the subdivided model 512, which are part of the machine learning model 51. This allows for a more multifaceted analysis and enables more accurate analysis.

[0285] For example, by using both the batch model 511 and the subdivision model 512 in combination, a more multifaceted analysis can be achieved when estimating a specific process group Gs, such as whether or not a common trend appears across all manufacturing equipment 121.

[0286] Furthermore, as explained with reference to Figure 15, by switching between using the entire sample dataset 801 or using a smaller dataset 802 that corresponds to a part of the sample dataset 801, the batch model 511 and the smaller model 512 can be generated, respectively. This makes it easy to perform analysis using the batch model 511 and the smaller model 512 as machine learning models 51.

[0287] Furthermore, according to the above embodiment, the subdivided dataset 802 used to generate the subdivided model 512 has fewer data points than the sample dataset 801 used to generate the batch model 511. The subdivided model 512 is generated with fewer training data points than the batch model 511.

[0288] Therefore, by using both the batch model 511 and the subdivision model 512 in combination, it is possible to achieve both high-precision estimation using the batch model 511, which excels in the number of samples (number of training data), and multifaceted estimation using the subdivision model 512, which places emphasis on the differences between the manufacturing equipment 121.

[0289] Furthermore, as explained with reference to Figure 15, the sample dataset 801 used to generate the unified model 511 has fewer manufacturing process Qs than the sub-dataset 802 used to generate the sub-model 512, because the unique manufacturing process Qu is excluded. The unified model 511 is generated by fewer factors (manufacturing process Qs) than the sub-model 512.

[0290] Therefore, by using the unified model 511 and the subdivided model 512 in combination, it is possible to achieve both multifaceted estimation using the subdivided model 512, which is generated by a large number of factors, and general estimation using the unified model 511, which is generated by factors common to the manufacturing equipment 121.

[0291] Furthermore, by using a machine learning model 51 as illustrated in Figure 13, even when multiple process groups G are connected, it is possible to estimate a specific manufacturing process Q in one process group G and a specific manufacturing process Q in another process group G connected to that process group G (see Figure 18). This enables easy and highly accurate analysis.

[0292] Furthermore, the importance of each manufacturing process Q can be mechanically calculated for each process group G, depending on the selection of the machine learning model 51. For example, if random forest models, as illustrated in Figure 13, are linked, the importance specific to that model can be calculated. In this way, the scope for user input in determining specific process groups Gs can be reduced, thereby enabling more accurate analysis.

[0293] Furthermore, as illustrated in Figure 13, a specific group of process steps Gs can be determined based on a chain of interconnected decision tree models. Since the importance of each model can be defined, the specific group of process steps Gs can be easily estimated.

[0294] Furthermore, as illustrated in Figure 23, even if the estimation results differ between the batch model 511 and the subdivision model 512, the estimation results for each machine learning model 51 are notified to the user without overwriting or rejecting one's results with the other. By also notifying the sample size at that time, it becomes possible to perform a more comprehensive analysis, such as analyzing the learning status of each model.

[0295] Furthermore, as illustrated in step S608 of Figure 19, the process not only determines a specific group of processes Gs, but also proposes a quality improvement plan. This makes it possible for even unskilled workers to easily improve the quality of product W.

[0296] Furthermore, as illustrated in step S608 of Figure 19, a highly visible display can be used when proposing a quality improvement plan. This makes it possible to easily improve the quality of product W.

[0297] Furthermore, the feature quantity P used in this disclosure is not limited to electrical data. For example, the analysis method according to this disclosure can be performed based on text data written by a worker (e.g., text describing the condition of scratches, etc.). This improves the usability of the analysis method.

[0298] Furthermore, as explained in relation to step S1 in Figure 5, the specific feature Ps is determined by the quality judgment model 41, separate from the machine learning model 51 for determining the manufacturing process Q. This reduces the amount of user input involved in the analysis of the manufacturing process Q, thereby enabling more accurate analysis.

[0299] Furthermore, the quality determination results used in this disclosure are not limited to the output from the quality determination model 41. The quality determination can be performed by a worker, and the subsequent processes can be carried out by computer 1 as described above, thereby enabling a division of labor in the manufacturing process. This improves the usability of the analysis method.

[0300] Furthermore, as explained with reference to Figure 2, etc., this disclosure is particularly useful for analyzing the rust prevention performance of metal product 201.

[0301] Furthermore, the manufacturing unit relating to this disclosure is not limited to manufacturing equipment 121. The manufacturing unit may be a label for distinguishing a manufacturing line, a label for distinguishing a manufacturing plant, or a label for distinguishing a manufacturer or supplier. This disclosure is applicable to various manufacturing units configured to individually perform the same process group G.

[0302] <8. Other Embodiments> In the above embodiment, both the batch model 511 and the subdivision model 512 were configured as first machine learning models 51 based on decision tree algorithms, but the present disclosure is not limited to such configurations.

[0303] Alternatively, a second machine learning model 52 based on a non-decision tree algorithm may be used in place of, or in addition to, the first machine learning model 51, and the generation of a batch model 511 and a subdivision model 512 may be performed in the second machine learning model 52.

[0304] As illustrated in Figure 24, the second machine learning model 52 is, for example, a Bayesian network in which each of the multiple feature quantities P is a child node and each of the multiple manufacturing processes Q is a parent node.

[0305] More specifically, the second machine learning model 52 takes each of the multiple features P as a child node, and the (Nn)th process group G is the second process group. N-n The manufacturing process Q comprising one or more processes is designated as the parent node, and the Nth process group G is designated as the first process group. N This is a Bayesian network in which one or more manufacturing processes Q that make up the network are intermediate nodes interposed between child nodes and parent nodes.

[0306] This Bayesian network can be visualized, for example, as a directed graph structure as shown in Figure 24. The second machine learning model 52 is pre-generated by unsupervised learning.

[0307] The second machine learning model 52 may be, for example, a machine learning model based on a non-decision tree algorithm. The second machine learning model 52 relating to a non-decision tree may be a graph structure determined by so-called graph structured analysis (GSA).

[0308] GSA is a big data analysis method proposed by the inventors of this application, which combines probability theory (Bayesian estimation) and graph theory. In this embodiment, GSA is used to determine the graph structure. However, using GSA to determine the graph structure is not essential. For details on GSA, please refer to Japanese Patent Application Publication No. 2021-111063.

[0309] In Figure 24, boxes containing text represent nodes. Arrows connecting boxes represent edges. In Figure 24, the Nth process group G. N One or more manufacturing processes Q belonging to the group are connected to the feature P via a single edge.

[0310] The probability distribution function for Bayesian networks is typically expressed by multiplying multiple conditional probabilities. Two nodes connected by an edge have a probability distribution function that includes a conditional probability where one node is the condition and the other is the variable. This suggests that two nodes connected by an edge have a relatively stronger dependency relationship compared to other nodes.

[0311] Furthermore, in Figure 24, the N-1 process group G N-1 One or more manufacturing processes Q belonging to the Nth process group G NOne or more manufacturing processes Q belonging to this group are connected by an edge. By connecting edges in this way, the Nth process group G can be derived from the feature P. N It is possible to trace back to one or more manufacturing processes Q belonging to the first process group G1, after passing through one or more manufacturing processes Q belonging to the first process group G1.

[0312] In this way, by following the arrows in Figure 24, a path can be formed connecting the manufacturing processes Q. This allows, for example, when focusing on a specific feature Ps, to extract the manufacturing process Q that has a strong dependence on that specific feature Ps from each process group G. In this embodiment, each manufacturing process Q connected by such a path can be considered as a specific process group Gs similar to that of the first machine learning model 51.

[0313] The second machine learning model 52, when given a feature P as input, outputs one or more manufacturing processes Q connected to the feature P via one or more edges, for each process group G. This output is none other than the specific process group Gs mentioned above. The output from the second machine learning model 52 corresponds to a sequence of manufacturing processes Q that have a strong dependency on the feature P input to the model 52.

[0314] In the second machine learning model 52 configured in this way, a batch model 511 and a subdivided model 521 may be generated. More generally, the CPU 3 may generate both the batch model 511 and the subdivided model 512 in at least one of the first machine learning model 51 and the second machine learning model 52.

[0315] Furthermore, in the above embodiment, the output data (quality data 35) of the quality determination program 212 was used as the quality determination result, but this disclosure is not limited to such a configuration.

[0316] In other words, of the flow shown in Figure 5, at least the inspection data analysis process (step S1) and the quality determination process (step S2) may be omitted.

[0317] In this disclosure, the "quality judgment result" may be, in addition to or instead of, the output from a pre-generated machine learning model (quality judgment model 41), a judgment result made by a factory worker, or a judgment result obtained by rule-based judgment using multiple feature quantities P as input. For example, when using a judgment result made by a factory worker, as described above, the configuration of the machine learning model 51 can be rearranged so that the text data recorded by that factory worker is used as input.

[0318] Furthermore, the feature quantities P used in this process may include, for example, the indicators used by a worker to determine quality based on their experience (such as the color, shape, and texture of the product). By quantifying these indicators, the process described in the above embodiment can be carried out in a similar manner.

[0319] Furthermore, although the above embodiment shows an example in which the analysis device is configured by a single computer 1, this disclosure is not limited to that example. The analysis method and analysis program 21 according to this disclosure may be executed using multiple computers 1, for example, by having the processing related to the specific process estimation process executed by the first computer and the processing related to the correlation analysis process executed by the second computer. In addition, the computer 1 in this disclosure also includes parallel computers such as supercomputers and PC clusters.

[0320] Furthermore, the screen on which various information can be displayed is not limited to the display screen on computer 1's display 11. Display images as shown in Figures 22 and 23 may be shown on a screen prepared separately from computer 1. [Explanation of Symbols]

[0321] 100 Manufacturing Management Systems 101 Analysis System 102 Manufacturing Systems 121 Manufacturing equipment (manufacturing units) 122 Measuring device 1. Computer (analysis device) 3 CPU (arithmetic unit) 7 RAM (memory section) 9 SSD (storage unit) 11. Display (Display Unit) 15. Keyboard (Reception area) 17. Mouse (Reception area) 18 Storage medium 21 Analysis Program 211 Inspection Data Analysis Program 212 Quality Assessment Program 213 Feature Selection Program 214 Model Acquisition Program 215 Specific Process Group Estimation Program 216 Correlation Analysis Program 217 Machine Learning Programs 31. Test Data 33 Feature Data 35 Quality Data 39 Process Data 41 Quality Judgment Model 51 Machine Learning Models (First Machine Learning Model) 511 Batch Model (Machine Learning Model) 512 Subdivision Models (Machine Learning Models) 52. The Second Machine Learning Model 801 sample datasets 802 sub-datasets D2 Second Training Data Downloadable training data Ds Small-scale training data P features Ps specific features Q Manufacturing process Qu's unique manufacturing process G process group Gs Specific process group W Products 201 Metal products 203 Insulating layer

Claims

1. A method for analyzing a manufacturing process, which is performed using a computer equipped with a memory unit and an arithmetic unit, and the result of determining the quality of a product manufactured through a group of processes consisting of one or more manufacturing processes, The multiple aforementioned process groups are arranged in the order of the progress of the manufacturing stages of the product. At least one of the multiple aforementioned process groups is performed individually by each of the multiple manufacturing units, If, among the multiple feature quantities used for determining the quality, one or more feature quantities estimated to contribute to determining the quality are designated as specific feature quantities, and a sequence of manufacturing processes selected from one or more manufacturing processes constituting each of the multiple process groups, and each estimated to contribute to determining the quality, in the order of progression, is designated as a specific process group, The calculation unit acquires a pre-generated machine learning model that associates each of the multiple feature quantities with one or more of the multiple manufacturing processes across one or more of the multiple process groups, The calculation unit estimates the specific process group based on the specific feature and the machine learning model, according to the type of machine learning model. The aforementioned machine learning model, A batch model that generates all units of the aforementioned manufacturing unit at once, The manufacturing unit selected for each process group is associated with one or more sub-models across one or more process groups, Methods for analyzing manufacturing processes.

2. In the method for analyzing the manufacturing process described in claim 1, The aforementioned group of processes is assigned a sample dataset that constitutes the training data for the machine learning model. The aforementioned sample dataset is composed of multiple sub-datasets that are quantified according to the manufacturing process and belong to different manufacturing units, The calculation unit generates the batch model based on the sample dataset, The calculation unit generates the sub-model based on the sub-dataset. Methods for analyzing manufacturing processes.

3. In the method for analyzing the manufacturing process described in claim 2, The aforementioned sub-dataset includes data values ​​corresponding to the manufacturing process specific to the manufacturing unit to which the sub-dataset belongs, The sample dataset is formed by combining multiple sub-datasets while excluding data values ​​corresponding to the specific manufacturing process. Methods for analyzing manufacturing processes.

4. In the method for analyzing the manufacturing process described in claim 1, If, among the multiple groups of processes, the group of processes performed at the end of the sequence is designated as the first group of processes, and the other groups of processes performed before the first group of processes are designated as the second group of processes, The aforementioned batch model and the aforementioned subdivision model are, respectively, A pre-generated machine learning model that includes, in addition to the association between a plurality of features and one or more manufacturing processes constituting the first process group, an association between each manufacturing process constituting the first process group and one or more manufacturing processes constituting the second process group, The calculation unit estimates the specific process group for both the batch model and the subdivision model, based on the specific feature and the machine learning model. Methods for analyzing manufacturing processes.

5. In the method for analyzing the manufacturing process described in claim 4, The aforementioned machine learning model includes a first machine learning model based on a decision tree algorithm, The calculation unit calculates the importance of each manufacturing process included in the first process group, corresponding to each of the multiple features in the first machine learning model, and the importance of each manufacturing process included in the second process group, which is related to each manufacturing process in the first process group. Based on these importance values, it determines one or more manufacturing processes for each process group, thereby estimating the specific process group. Methods for analyzing manufacturing processes.

6. In the method for analyzing the manufacturing process described in claim 5, The first machine learning model is, By performing a process of generating submodels using a random forest, each consisting of multiple explanatory variables and one of multiple target variables, for all of the aforementioned target variables, a decision tree model composed of multiple submodels is obtained. This model is constructed by sequentially generating another decision tree model, which is formed by generating multiple submodels, each of which is newly generated when another variable connected to the multiple explanatory variables is used as a new explanatory variable, and each of the multiple explanatory variables is used as a new dependent variable. Methods for analyzing manufacturing processes.

7. In the method for analyzing the manufacturing process described in claim 5, The aforementioned machine learning model, The first machine learning model described above, This includes a second machine learning model based on a non-decision tree algorithm, The aforementioned arithmetic unit, In at least one of the first machine learning model and the second machine learning model, the generation of both the batch model and the sub-model is performed. Methods for analyzing manufacturing processes.

8. In the method for analyzing the manufacturing process described in claim 4, If the calculation unit determines that the estimation result of the specific process group by the batch model differs from the estimation result of the specific process group by the subdivision model, The estimation results of the aforementioned batch model and the aforementioned subdivision model, The number of samples used in generating the aforementioned batch model and the aforementioned subdivision model will be notified, respectively. Methods for analyzing manufacturing processes.

9. In the method for analyzing the manufacturing process described in claim 4, The aforementioned quality judgment result is quantified as quality data that increases or decreases depending on whether the quality is good or bad. The manufacturing processes that constitute the first and second process groups are each quantified in a way that characterizes the content of each manufacturing process, The aforementioned storage unit is A first correlation coefficient that shows the correlation between each of the aforementioned multiple features and the quality data, A second correlation coefficient that shows the correlation between each manufacturing process in the first group of processes and each of the multiple feature quantities, A third correlation coefficient, which shows the correlation between each manufacturing process in the second process group and each manufacturing process in the first process group, is stored in advance. After estimating the specific process group using the batch model and the subdivision model, the calculation unit displays the calculation results of the calculation unit on a display unit capable of displaying them, based on the first, second, and third correlation coefficients. The breakdown of the specific process group according to the aforementioned batch model and the aforementioned subdivision model, The improvement methods for the specific process group to improve the aforementioned quality are shown for the batch model and the sub-model, respectively. Methods for analyzing manufacturing processes.

10. In the method for analyzing the manufacturing process described in claim 1, The calculation unit reads, as the quality determination result, the determination result by a worker, or the determination result by rule-based determination using multiple features as input, or the output from a pre-generated machine learning model. Methods for analyzing manufacturing processes.

11. In the method for analyzing the manufacturing process described in claim 1, The calculation unit acquires at least one of the following as a plurality of feature quantities: image data, audio data, text data, and mechanical or electrical data that characterize the product. Methods for analyzing manufacturing processes.

12. In the method for analyzing the manufacturing process described in claim 1, The aforementioned product is a metal product having an insulating layer on its surface, The quality of the aforementioned product refers to the rust prevention performance of the insulating layer. The aforementioned arithmetic unit, By applying a voltage while a corrosive factor is in contact with the surface of the metal product, the change in current over time caused by the voltage is obtained. As multiple features, measurements that characterize the waveform of the change over time are obtained. Methods for analyzing manufacturing processes.

13. In the method for analyzing the manufacturing process described in claim 1, The multiple manufacturing units consist of different manufacturing lines, manufacturing plants, or manufacturers or suppliers, each configured to perform the process group individually. Methods for analyzing manufacturing processes.

14. A manufacturing process analysis device that uses a computer equipped with a memory unit and an arithmetic unit, and the quality determination results of products manufactured through a group of processes consisting of one or more manufacturing processes, The multiple aforementioned process groups are arranged in the order of the progress of the manufacturing stages of the product. At least one of the multiple aforementioned process groups is performed individually by each of the multiple manufacturing units, If, among the multiple feature quantities used for determining the quality, one or more feature quantities estimated to contribute to determining the quality are designated as specific feature quantities, and a sequence of manufacturing processes selected from one or more manufacturing processes constituting each of the multiple process groups, and each estimated to contribute to determining the quality, in the order of progression, is designated as a specific process group, A model acquisition means for acquiring a pre-generated machine learning model that associates each of the multiple features with one or more of the multiple manufacturing processes across one or more of the multiple process groups, The system includes estimation means for estimating the specific process group according to the type of machine learning model, based on the specific features and the machine learning model. The aforementioned machine learning model, A batch model that generates all units of the aforementioned manufacturing unit at once, The manufacturing unit selected for each process group is associated with one or more sub-models across one or more process groups, A device for analyzing manufacturing processes.

15. A computer equipped with a memory unit and an arithmetic unit, and a manufacturing process analysis program that is executed using the quality determination results of products manufactured through a group of processes consisting of one or more manufacturing processes, The multiple aforementioned process groups are arranged in the order of the progress of the manufacturing stages of the product. At least one of the multiple aforementioned process groups is performed individually by each of the multiple manufacturing units, If, among the multiple feature quantities used for determining the quality, one or more feature quantities estimated to contribute to determining the quality are designated as specific feature quantities, and a sequence of manufacturing processes selected from one or more manufacturing processes constituting each of the multiple process groups, and each estimated to contribute to determining the quality, in the order of progression, is designated as a specific process group, To the aforementioned computer, The calculation unit includes a process of acquiring a pre-generated machine learning model that associates each of the multiple feature quantities with one or more of the multiple manufacturing processes across one or more of the multiple process groups, The calculation unit executes a process of estimating the specific group of steps according to the type of machine learning model, based on the specific feature and the machine learning model. The aforementioned machine learning model, A batch model that generates all units of the aforementioned manufacturing unit at once, The manufacturing unit selected for each process group is associated with one or more sub-models across one or more process groups, A program for analyzing manufacturing processes.

16. The analysis program described in claim 15 is stored. A computer-readable storage medium.

Images