KR20230016343A - Automatic Analysis System for Quality Data Based on Machine Learning - Google Patents

Abstract

In this embodiment, provided are a quality data analysis system and a method thereof. In order to reduce the quality cost due to the reduction of the time required for product quality analysis and the occurrence of defects, a machine learning-based inference model is trained based on product quality data, analysis reports are provided by analyzing the quality data based on the inference model, and process features are adjusted using the inference model as a simulator.

Description

Automatic Analysis System for Quality Data Based on Machine Learning}

The present disclosure relates to a system for automatically analyzing quality data based on machine learning. More specifically, an AI-based inference model is trained based on the accumulated quality data for the product, an analysis report is provided by analyzing the quality data based on the inference model, and a process factor (process factor) is provided by using the inference model as a simulator. It relates to a quality data analysis system and method for adjusting feature).

The information described below merely provides background information related to the present invention and does not constitute prior art.

Although the conventional quality system accumulates quality data on process features (process features or process parameters) and field claim data generated during the sales process, they are rarely utilized. at an insignificant level. In terms of creating quality cost reduction, it is necessary to select process factors that cause defects by analyzing the correlation between accumulated field claim data and process factors, and improve defects by adjusting the values of the process factors.

Recently, quality data analysis using machine learning has been sporadically attempted, but there is a problem that it takes an average of 2 to 3 months per process and there is a shortage of data analysis experts to expand and develop the analyzed results. In addition, since quality data analysis is rarely completed once, there is a problem that the quality data must be reanalyzed and the results applied to the field when product specifications or production conditions are changed.

On the other hand, quality data collected in the production process is a valuable asset in terms of defect cause analysis, process improvement, and consequent quality cost reduction. However, in the collected quality data, there are many cases in which process factor values are biased, and the biased process factor can make the quality analysis process based on the quality data difficult. One of the causes of such process factor bias is unsystematic management of process factors.

In general, process parameters can be adjusted within the range of quality control standards. However, due to the nature that the person in charge of the field must directly change or manage the process factor, a case in which a process factor value is fixed and managed can often occur. For example, since the process factor value is changed according to the judgment of the person in charge of the field, there is a case where a specific process factor value is managed as a single value without being changed. In particular, in this case, there is a problem in that analysis of the quality data itself is impossible.

Therefore, it is possible to accumulate quality data that is easy to analyze by solving process factor bias, analyze the accumulated quality data to select process factors that cause defects, and adjust the value of the process factor to reduce defects. options should be considered.

The present disclosure trains a machine learning-based inference model based on product quality data in order to reduce the quality cost due to the reduction in the time required for product quality analysis and the occurrence of defects, and based on the inference model. The purpose is to provide an analysis report by analyzing quality data, and to provide a quality data analysis system and method that adjusts process features using an inference model as a simulator.

According to an embodiment of the present disclosure, in a method for improving quality control criteria for a product, performed by a computing device, a process of acquiring quality data and field claims for the product, wherein the quality data includes: Collected for a plurality of process features, applied or occurring in the production process of, the field claim indicates the quality or defect of the product; analyzing a first influence between the plurality of process factors and the field claim; expanding an existing management range of the biased process factor in the case of a biased process factor whose first influence does not fall within a predetermined upper ratio among the plurality of process factors; recollecting quality data for the plurality of process factors based on the expanded management scope; analyzing a second influence between the plurality of process factors and the field frame with respect to the recollected quality data; and, when the first influence or the second influence falls within the predetermined upper ratio, subdividing and recollecting the data of the corresponding process factor.

According to another embodiment of the present disclosure, an input unit for acquiring quality data and field claims for the product, wherein the quality data is applied or generated in the production process of the product, a plurality of process factors (process feature), and the field claim indicates a good or bad product; an impact analysis unit analyzing a first influence between the plurality of process factors and the field claim; a management range adjusting unit expanding an existing management range of the biased process factor in the case of a biased process factor whose first influence does not fall within a predetermined upper ratio among the plurality of process factors; Based on the expanded management scope, a data re-collecting unit for re-collecting quality data for the plurality of process factors; and a data segmentation collection unit that subdivides and recollects the data of the corresponding process factor when the first influence falls within the predetermined upper ratio.

According to another embodiment of the present disclosure, a computer program stored in a computer-readable recording medium is provided to execute each step included in a method for improving quality control standards.

As described above, according to the present embodiment, quality data analysis that trains a machine learning-based inference model based on the accumulated quality data for products, analyzes the quality data collected based on the inference model, and provides an analysis report By providing the system and method, there is an effect of reducing the quality cost according to the reduction of the time required for product quality analysis.

In addition, according to the present embodiment, by providing a quality data analysis system and method for adjusting process factors using an inference model as a simulator, it is possible to reduce product defects.

In addition, according to this embodiment, by providing an analysis report by analyzing the quality data collected based on the inference model, and by using the inference model as a simulator to provide a quality data analysis system and method for adjusting process factors, data analysis non-majors It has the effect of establishing an MLaaS (Machine Learning as a Service) environment in which field personnel can perform quality analysis on products, and enabling field-led quality data management and analysis.

In addition, according to the present embodiment, by providing a quality data analysis system and method for accumulating quality data by resolving process factor bias based on improvement of quality control standards, the imbalance of quality data is reduced and the efficiency of quality data analysis is improved. There is an effect that makes it possible to increase.

1 is a schematic illustration of a quality data analysis system according to an embodiment of the present disclosure.
2 is an exemplary diagram illustrating components of an analysis report according to an embodiment of the present disclosure.
3 schematically illustrates additional components of a simulator according to one embodiment of the present disclosure.
4 is an exemplary view of a UI for selecting a process factor according to an embodiment of the present disclosure.
5 is an exemplary diagram of a UI for indicating importance of process factors according to an embodiment of the present disclosure.
6 is an exemplary diagram of a UI for displaying an analysis result according to an embodiment of the present disclosure.
7 is an exemplary view of a UI for adjusting a process factor according to an embodiment of the present disclosure.
8 schematically illustrates additional components used for training an inference model according to an embodiment of the present disclosure.
9 is a flowchart of a pre-processing process of quality data according to an embodiment of the present disclosure.
10 is a flowchart of a process factor selection process according to an embodiment of the present disclosure.
11 is a flowchart illustrating a training process for a machine learning model according to another embodiment of the present disclosure.
12 is a flowchart of a quality data analysis method according to an embodiment of the present disclosure.
13 is a flowchart of a method of changing a quality control criterion based on a simulator according to an embodiment of the present disclosure.
14 is a flowchart of a method for training an inference model according to an embodiment of the present disclosure.
15 is a schematic configuration diagram of an apparatus for improving quality control standards for process factors according to an embodiment of the present disclosure.
16 is a flowchart of a method for improving quality control standards of process factors according to an embodiment of the present disclosure.
17 is a flowchart of a process of applying an analysis system according to an embodiment of the present disclosure to a gearbox.
18 is an exemplary diagram illustrating importance of characteristics of process factors of a gearbox according to an embodiment of the present disclosure.
19 is an exemplary diagram illustrating a T-test according to an embodiment of the present disclosure.

Hereinafter, embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to components of each drawing, it should be noted that the same components have the same numerals as much as possible even if they are displayed on different drawings. In addition, in describing the present embodiments, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present embodiments, the detailed description will be omitted.

In addition, in describing the components of the present embodiments, terms such as first, second, A, B, (a), and (b) may be used. These terms are only used to distinguish the component from other components, and the nature, sequence, or order of the corresponding component is not limited by the term. Throughout the specification, when a part 'includes' or 'includes' a certain component, it means that it may further include other components without excluding other components unless otherwise stated. . In addition, the '... Terms such as 'unit' and 'module' refer to a unit that processes at least one function or operation, and may be implemented as hardware, software, or a combination of hardware and software.

The detailed description set forth below in conjunction with the accompanying drawings is intended to describe exemplary embodiments of the present invention and is not intended to represent the only embodiments in which the present invention may be practiced.

This embodiment discloses a system for automatically analyzing quality data based on machine learning. More specifically, in order to reduce the quality cost due to the decrease in the time required for product quality analysis and the occurrence of defects, a machine learning-based inference model is trained based on the accumulated quality data, and based on the inference model Provides an analysis report by analyzing quality data, and a quality data analysis system and method for adjusting process features using an inference model as a simulator.

In the following description, since a machine learning-based quality analysis service can be provided to a user (eg, a field or field manager), the service that the quality data analysis system according to this embodiment can provide to the user It is referred to as Machine Learning as a Service (MLaaS).

1 is a schematic illustration of a quality data analysis system according to an embodiment of the present disclosure.

The quality data analysis system (100, hereinafter referred to as 'analysis system') according to this embodiment trains a machine learning-based inference model based on accumulated quality data for a product, and determines quality based on the inference model. It analyzes data to provide an analysis report and adjusts process factors by using an inference model as a simulator. The analysis system 100 includes all or part of an input unit 102 , a data pre-processing unit 104 , a determination unit 106 , and a data visualization unit 108 .

Here, components included in the analysis system 100 according to the present embodiment are not necessarily limited thereto. For example, the analysis system 100 may further include a UI unit 110 to provide convenience when a user uses MLaaS. In addition, the analysis system 100 may further include a training unit 112 for training of the reasoning model included in the determination unit 106, or may be implemented in a form interlocking with an external training unit.

1 is an exemplary configuration according to the present embodiment, the shape of the input unit, the operation of the data pre-processing unit, the structure and operation of the inference model included in the determination unit, the operation of the quality data analysis unit, the structure and operation of the training unit, and the UI Depending on the configuration of the unit, various implementations including other components or other connections between components are possible.

The input unit 102 acquires quality data about the product. Here, the product may be a part included in a vehicle, such as a gearbox. Quality data can be collected for a plurality of process factors applied or occurring in the production process of a product.

Process factors may include all or part of input factors for adjusting the production process of a product, intermediate output factors formed in the middle of the production process, or output factors generated as a result of the production process.

Meanwhile, process factors input for quality data analysis may be key process factors selected in a pre-training process for an inference model. The selection process of these key process factors will be explained in the training process for the inference model.

The input unit 102 may set a data type for a process factor used as an input of an inference model. Here, the data type of the process factor may include a numeric type represented by numbers and a category type represented by characters. As another data type, there is a time type including information on the time the data was collected, but it can be removed in the process of selecting key process factors during the training process.

On the other hand, the quality data may include a factor (eg, whether a field claim has occurred for a product) that can be used as a target output (ie, a label for analysis) in order to analyze the performance of an inference model. The input unit 102 sets a factor used as a target output as a target factor.

The data pre-processing unit 104 performs an appropriate encoding process for each data type of process factor, and sets missing data generated in the collection process to an appropriate value.

The data preprocessing unit 104 may perform an encoding process of converting the categorical process factor into an embedding value suitable for an inference model.

An example of categorical data is a target factor indicating whether a field claim has occurred for a product. In the encoding process for the target factor, for example, a case where no field claim for a product has occurred is represented by 0, and a case where a field claim has occurred is represented by 1. Accordingly, encoding of the target factor may be a process of generating an analysis label for quality analysis based on an inference model.

In addition, the data pre-processing unit 104 may set values of process factors omitted in the collection process. For example, the numerical process factor may be set to a median value, and the categorical process factor may be set to a mode value.

The determination unit 106 includes an inference model, and generates a determination result on whether the product is good (OK) or bad (No Good: NG) by using the inference model based on a plurality of preprocessed process factors. Here, the determination result may be a probability value for the quantity or defect of the product.

The result of the product defect determination may indicate a case in which a field claim has occurred for the product. Therefore, the result of the judgment on the quantity of the product indicates a case where no field claim is made for the product.

Inference models are implemented in the form of machine learning models, such as tree-based decision trees, random forests, XGBoost (Extreme Gradient Boosting), or LightGBM (Light Gradient Boosting), which show good performance for quality data. Boosting Model) can be an implemented model. Using the training process, the training unit 112 may select, as an inference model, a model with the best performance among models to which each of the four machine learning algorithms is applied. A training process for selecting an inference model will be described later.

The data visualization unit 108 generates an analysis report for product quality analysis or inference model learning results based on a plurality of process factors, analysis labels, and judgment results.

2 is an exemplary diagram illustrating components of an analysis report according to an embodiment of the present disclosure.

The analysis report provided by the data visualization unit 108 includes a summary of analysis data 202, importance of process factors 204, It may include all or part of the data distribution 206 for each process factor and the analysis result 208 .

The analysis data summary 202 represents overall information of process factors constituting the quality data. Here, the overall information may include a data type, a mode value, a minimum value, a maximum value, an average, a standard deviation, and the like. The analysis data summary 202 may be provided as a result of product quality analysis or training of an inference model.

The process factor importance 204 indicates the feature importance of the process factor, so that the effect of each process factor on the judgment result can be confirmed. Process factor importance 204 may be provided as a result of training of an inference model. Feature importance is the result of a tree-based machine learning algorithm, which will be described in detail later.

The data distribution 206 for each process factor represents the distribution of the relationship between each process factor and the judgment result or between each process factor and the label for analysis.

The analysis result 208 represents a performance analysis for the inference model based on the decision result and the label for analysis. The analysis result 208 may be provided as a result of product quality analysis or learning of an inference model. The analysis result 208 will be described later.

The analysis report can be used in the process of calculating a new quality control standard by changing the quality control standard. In addition, an analysis report can be created to confirm the characteristics of the quality data collected during the production process to which the new quality control standards are applied.

Meanwhile, the determination unit 106 may use an inference model as a simulator for calculating a new quality control criterion.

After setting the adjusted factor value for a specific process factor, the determination unit 106 inputs the adjusted process factor into a simulator to generate a simulated decision result. A new quality control criterion for the process factor may be created in the direction of reducing the occurrence of defects in the product by using the adjusted process factor and the corresponding judgment result.

Meanwhile, the simulator may include additional components to provide convenience to the user. Thus, in the description below, the simulator represents a system that includes an inference model and additional components.

3 schematically illustrates additional components of a simulator according to one embodiment of the present disclosure.

The simulator is all or all of the process factor adjustment unit 302, the decision result output unit 304, the important factor output unit 306, and the standard application unit 308 for selection and adjustment of process factors and provision of judgment results. includes some

The process factor adjustment unit 302 selects an adjusted process factor from main process factors and adjusts the value of the adjusted process factor. As described above, the data distribution 206 for each process factor and the importance of the characteristics provided by the analysis report can be used to select the adjustment process factor.

Meanwhile, the process factor adjustment unit 302 may select the input factor as the adjustment process factor, as described above.

If the selected adjustment process factor is a categorical type, a user-desired category may be selected using a check box. In the case of numeric type, the value of the process factor can be adjusted using a slider. Since the process factor can be excluded from the simulation by unchecking the checkbox, simulation for a single process factor can also be performed. At this time, when the XGBoost-based model is adopted as the inference model, the excluded process factor is set to a preset value, and when another algorithm-based model is adopted, the mode or median value may be set according to the data type of the process factor.

Meanwhile, when adjusting the process factor value, the process factor value may be adjusted to minimize the distribution of defects in the product by referring to the T-test result for the corresponding process factor.

The decision result output unit 304 provides a decision result when the adjustment process factor is input to the simulator. As described above, the judgment result is a probability value for the quantity or defect of the product.

On the other hand, the user can check the effect of the adjusted process factor on the occurrence of defects by changing and inputting the value of the adjusted process factor and then checking the judgment result.

The important factor output unit 306 provides the importance of the characteristics of the process factor being used by the simulator. Here, as the feature importance, the feature importance generated in the learning process for the inference model used as the simulator is reused.

The criterion application unit 308 selects an optimal factor value for the adjusted process factor based on the result of determining the value of the adjusted process factor, and changes the quality control standard for the adjusted process factor based on this.

As described above, according to the present embodiment, by providing an analysis system for adjusting process factors using an inference model as a simulator, it is possible to reduce product defects.

The UI unit 110 connects the MLaaS provided by the analysis system 100 to the user by obtaining an input related to the analysis system 100 from the user or providing an output generated by the analysis system 100 on a display. fulfill the role of Based on the UI unit 110, a user input may be provided to the analysis system 100 using means such as a mouse and a keyboard.

4 is an exemplary view of a UI for selecting a process factor according to an embodiment of the present disclosure.

As illustrated in FIG. 4 , the UI unit 110 includes a check box for selecting a process factor from quality data applied to data analysis. For the process factor whose checkbox is selected, the type of process factor can be additionally input, and the description indicates constraints according to the type of process factor.

In addition, the bar illustrated in FIG. 4 may also be used as a check box for selecting a process factor used for training of an inference model.

Meanwhile, the UI unit 110 includes an input interface for setting a target factor among process factors.

The UI unit 110 provides an analysis report including an analysis data summary 202, process factor importance 204, data distribution by process factor 206, or analysis result 208 on a display. For example, as illustrated in FIG. 5 , the UI unit 110 may provide feature importance for process factors.

In addition, the UI unit 110 may provide an analysis result 208 based on the decision result of the reasoning model.

As illustrated in FIG. 6 , the analysis results 208 may include accuracy, precision, recall, and F1 scores for good or bad decisions for the product.

Here, accuracy is the rate at which predictions for good or bad match GT (Ground Truth, ie correct answer or label). Accuracy is the proportion of GTs predicted as defective that are also defective, and recall is the proportion of predicted defective GTs as defective. The F1 score is the harmonic mean value of precision and recall.

Meanwhile, in the example of FIG. 6 , the machine learning model identifier represents one of algorithms implemented by the machine learning model, that is, decision tree, random forest, XGBoost, or LightGBM.

The UI unit 110 provides training results for models to which each of the four machine learning algorithms are applied on a display. The analysis result 208 as illustrated in FIG. 6 may also be used as a training result for each algorithm-based model. In addition, the training result may include runtime, which is the time required for learning the inference model.

As illustrated in FIG. 7 , the UI unit 110 includes a check box for obtaining an input related to the process factor adjustment unit 302 when using a simulator. For the process factor whose checkbox is selected, the process factor value can be adjusted according to the data type. In addition, the UI unit 110 provides results related to the decision result output unit 304 and the significant factor output unit 306 on a display.

The UI unit 100 may provide a matrix-type heatmap used for correlation analysis between process factors.

In addition, the UI unit 100 provides an input interface for obtaining preset values (eg, preset values representing the number of major process factors) that are grounds for various judgments on MLaaS.

The interface supported by the UI unit 100 is not limited to the above, and an interface for connecting MLaaS to a user may be further added as needed.

The training unit 112 performs training on an inference model using quality data for learning and corresponding labels.

As described above, the inference model is implemented in the form of a machine learning model, and may be a model implemented with one of four machine learning algorithms such as decision tree, random forest, XGBoost, or LightGBM.

A decision tree is a model that classifies data according to a specific criterion (eg, a specific value of a numerical process factor or a category of a categorical process factor). Branching in a decision tree is performed in a direction in which information gain by a process factor used for branching is maximized, and this is called training for a decision tree.

When two leaf nodes are generated by branching the root node based on one process factor, the information gain is calculated by subtracting the information of the two leaf nodes from the information of the root node. can do. At this time, the label is used in the process of calculating the information gain. Since branched leaf nodes are more ordered, the information of two leaf nodes cannot be greater than that of the root node. Therefore, the information gain always has a value greater than zero. Meanwhile, as information, entropy or Gini impurity may be used.

A random forest is an ensemble model based on multiple decision trees, which combines decisions made by multiple decision trees (e.g., in the case of a classification model, a majority vote is taken, and in the case of a regression model, the average is taken). ) to produce the final output. Training for each decision tree included in the random forest may be performed in the same way as learning for one decision tree. A feature of random forest is that it allows boosttrap between training data sets used for learning each decision tree. This is called bagging (boostrap + aggregation), which encompasses extraction with restoration of random forest and combination of decisions by multiple decision trees.

Both XGBoost and LightGBM are GBM (Gradient Boosting Model) family algorithms. GBM is an ensemble algorithm based on boosting. Here, boosting is a process of sequentially generating (ie, training) a plurality of weak classifiers and then combining them to generate a strong classifier. For example, for three weak classifiers A, B, and C, classifier A is created, classifier B is created based on that information, classifier C is created again based on that information, and finally all classifiers are combined. Thus, a strong classifier can be created. In this boosting process, GBM creates a weak model of the next stage based on the negative gradient calculated from the weak model of the front stage.

The XGBoost algorithm is a GBM-based algorithm for learning an ensemble model in which a weak classifier is implemented as a decision tree. The XGBoost algorithm has the advantage of being useful in preventing overfitting, which is a disadvantage of GBM, by including a regulation term in the loss function for training.

The LightGBM algorithm is also a GBM-based algorithm for learning an ensemble model in which a weak classifier is implemented as a decision tree. The LightGBM algorithm performs tree branching leaf-wise, not level-wise, in order to improve the slow learning speed of GBM-based algorithms. The LightGBM algorithm is known to be suitable for large-volume data processing because it causes an overfitting problem when too few data are used.

Since all four machine learning algorithms operate based on decision trees, they can generate feature importance for process factors used for branching as a result of learning.

The feature importance for one process factor is the ratio of the total information gain generated by one process factor to the total information gain by (multiple) decision trees. That is, it represents the contribution of all branches according to one process factor to the total information gain generated by the learned decision tree. The higher the importance of the feature, the higher the contribution of the corresponding process factor when the inference model generates the decision result.

When adjusting the quality control criteria for a specific process factor, this feature importance can be utilized, so in this embodiment, one of the above-described decision trees, random forests, XGBoost, or LightGBM is used as a machine learning algorithm for an inference model. use

Using the training process, the training unit 110 may select a model with the best performance among models to which each of the four machine learning algorithms are applied as an inference model. After selecting an algorithm for an inference model, the training unit 110 presents training results for each of the models implementing the four machine learning algorithms as a basis for determination.

Hereinafter, the training process of the inference model performed by the training unit 110 will be described using the examples of FIGS. 8 to 11 .

8 schematically illustrates additional components used for training of an inference model according to an embodiment of the present disclosure.

In order to train the inference model, the training unit 110, in addition to the input unit 102, additionally includes a data preprocessing unit 104, a process factor selection unit 806, a data balancing unit 808, and four machine learning models 810, Hereinafter, all or part of the 'four models' may be used. Here, the four models 810 represent models to which each of the four machine learning algorithms, as described above, is applied.

The input unit 102 obtains quality data about the product for use in training. Quality data can be collected for a plurality of process factors applied or occurring in the production process of a product.

Process factors may include all or part of input factors for adjusting the production process of a product, intermediate output factors formed in the middle of the production process, or output factors generated as a result of the production process.

The input unit 102 may set data types for process factors used for training. Here, the data type of the process factor may include a numeric type expressed as a numerical value, a categorical type expressed as a character, and a time type including time information at which data was collected.

On the other hand, the quality data may include a factor that can be used as a target output (ie, a label for learning) in the training process of an inference model (eg, whether a field claim has occurred for a product). The input unit 102 sets a factor used as a target output as a target factor.

The category and target factor for the process factor may be set using the UI unit 110 as described above.

The data pre-processing unit 104 performs an appropriate encoding process for each data type of process factor, and sets missing data generated in the collection process to an appropriate value.

9 is a flowchart of a pre-processing process of quality data according to an embodiment of the present disclosure.

The data pre-processing unit 104 checks the data type of the process factor (S900).

The data pre-processing unit 104 checks whether the data type is numeric data (S902), and if not, checks whether it is categorical data (S904).

The data pre-processing unit 104 removes time-type data, not numeric/categorical data (S906). It is determined that the time for which the quality data is collected has little correlation with the quantity or defect of the product, and the time-type process factor is removed from the quality data for training.

In the case of categorical data, the data pre-processing unit 104 performs an encoding process of converting into an embedding value suitable for the inference model (S908).

An example of categorical data is a target factor indicating whether a field claim has occurred for a product. In the encoding process for the target factor, for example, a case where no field claim for a product has occurred is represented by 0, and a case where a field claim has occurred is represented by 1. Accordingly, encoding of the target factor may be a process of generating a learning label for training an inference model.

For numeric data and encoded categorical data, the data pre-processing unit 104 processes missing data generated in the collection process (S910). In this case, the categorical data may be set as the most frequent value, and the numeric data may be set as the median value. On the other hand, in the case of a process factor with severe omission, it may interfere with the training of the inference model. Accordingly, the data pre-processing unit 104 may remove a process factor having a higher omission rate than a predetermined rate (eg, 80%) in the training process.

The number of process factors included in the quality data may range from tens to hundreds depending on the target product. The process factor selection unit 806 selects a major process factor having a high influence on the target factor from a plurality of process factors included in the quality data. By using the selected key process factors, the complexity and learning time of the inference model can be reduced.

10 is a flowchart of a process factor selection process according to an embodiment of the present disclosure.

The process factor selection unit 806 obtains the quality data preprocessed by the data preprocessing unit 104 (S1000).

The process factor selection unit 806 checks whether the number of process factors is less than or equal to a predetermined number (20 in the example of FIG. 10) (S1002). If the number of process factors is less than or equal to the predetermined number, the process factor selection unit ( 806) can omit the process factor selection process.

If the number of process factors is greater than the preset number, the process factor selection unit 806 may perform processes (S1004 to S1008) for calculating the main process factors to select the main process factors to be less than or equal to the preset number. there is.

First, the process factor selector 806 performs a T-test on process factors included in the quality data (S1004).

Here, the T-test is a method of confirming statistical significance by comparing two distributions for product quantity and defect for each process factor. When the difference between the two distributions is significant, the process factor selector 806 determines that the corresponding process factor can affect the occurrence of defects, and selects the process factor as a major process factor.

The process factor selection unit 806, when the number of process factors that have passed the T-test is less than or equal to the preset number, skips the remaining processes (S1006 and S1008), and selects the process factors that have passed the T-test as the final main process. It can be selected as a factor.

The process factor selector 806 compares information gains between process factors that have passed the T-test (S1006). A predetermined number (eg, 20) of process factors may be selected in order of information gain. As described above, the information gain may be generated by subtracting the information on the normality or defect after branching by one process factor from the information on the normality or defect of the product.

The process factor selection unit 806 analyzes the correlation between the process factors selected in the order of information gain (S1008). As described above, since the process factor may be an input factor, an intermediate output factor, or an output factor of the product production process, a correlation may exist between the process factors selected in the order of information gain. At this time, for a plurality of process factors, the correlation between the two process factors is represented by a correlation coefficient, which is a value obtained by dividing the covariance of the two process factors by the product of the standard deviations of the two process factors. Meanwhile, the correlation coefficient may be expressed on a heatmap in the form of a matrix.

The process factor selection unit 806 analyzes the correlation between the selected process factors and identifies a case where the correlation coefficient is greater than a preset reference value. The process factor selection unit 806 removes, in the order of output factors, intermediate output factors, and input factors, two process factors having a correlation coefficient greater than a predetermined reference value. For example, when two process factors having a correlation are an output factor and an input factor, respectively, the output factor is removed. Meanwhile, when two process factors having a higher correlation coefficient than a predetermined reference value are of the same type, a process factor having a high information gain is selected.

Using process factor selection based on correlation, the process factor selector 806 may remove multicollinearity existing between process factors.

On the other hand, when the number of selected process factors is smaller than the preset number due to the removal of process factors according to the correlation analysis, the process factor selector 806 may additionally select process factors according to the order of information gain. can

Based on the T-test, information gain comparison, and correlation analysis as described above, the process factor selector 806 may select a final major process factor.

On the other hand, the quality data may have an imbalance state in which the defective product data is generally very small compared to the good product. For example, depending on the product, there are cases where a serious ratio of thousands to one is displayed. Since such an imbalanced state may induce biased learning for a machine learning algorithm-based model, data balancing based on augmentation of defective data may be required.

The data balancing unit 808 performs data balancing on bad data. The data balancing unit 808 upsamples bad data to increase the number of bad data, thereby achieving a balance between bad data and good data. For example, the data balancing unit 808 may generate similar data within a data distribution using a k Nearest Neighbors (kNN) model technique.

Here, the kNN model technique is a method of classifying into a category that includes more data after examining k data around (neighbors) when new data is given. Accordingly, the data balancing unit 808 may increase the number of defective data by generating new data in the vicinity including more than half of the defective data among the k pieces.

After performing preprocessing of quality data, selection of key process factors, and balancing, the training unit 112, as described above, uses four machine learning models based on the decision tree, random forest, XGBoost, and LightGBM algorithms ( 810) is trained, and one with the best performance is selected as an inference model.

First, the training unit 112 divides the balanced quality data into learning data and verification data. For example, 80% of quality data may be used as training data, and the remaining 20% of quality data may be used as verification data.

The training unit 112 performs training on four machine learning models 810 based on training data and training labels. Since each model is implemented based on a decision tree, training can be performed in a direction that maximizes the information gain at each branch in the tree.

The training unit 112 performs cross-validation on the four machine learning models 810 based on data for verification, and stores training performance of the four machine learning models 810 .

As hyperparameters for training, for example, max-depth, leaf-limit, etc. are used. The maximum depth represents the maximum value of a tree branch, and the leaf limit represents the limit value for a leaf. indicate

In particular, the training unit 112 focuses on preventing overfitting by appropriately adjusting the maximum depth in the training process for the four models 810 .

After learning of the four models 810 is completed, the training unit 112 compares performances of the four models 810 and selects an inference model. As illustrated in FIG. 6 , the performance of the learned model includes a training label and accuracy, precision, recall, and F1 score based on the decision result generated by each machine learning model. In addition, the performance of the learned model may include runtime, which is the time required for learning.

The training unit 112 selects a model having the highest F1 score as a final inference model. However, when selecting a final model, the user may use recall as a selection criterion if the purpose is to reduce defects or precision if the purpose is to reduce false defects.

11 is a flowchart illustrating a training process for a machine learning model according to another embodiment of the present disclosure.

The training unit 112 divides the balanced quality data into training data and verification data (S1100).

The training unit 112 performs training on one machine learning model based on the learning data and the learning label (S1102). Since each model is implemented based on a decision tree, training can be performed in a direction that maximizes the information gain at each branch in the tree.

The training unit 112 performs cross-validation on the learned machine learning model based on data for verification (S1104), and then stores the performance of the model as a training result (S1106).

In the training of machine learning models, one of the important issues is the trade-off between training time and model performance achieved. In order for field personnel who are not proficient in data analysis to utilize the analysis system 100, it may be appropriate to manage the learning time within 2 to 3 hours, so this amount of learning time can be used as a criterion in the negotiation process. can In order to satisfy this criterion of learning time, training is performed once for each machine learning model based on hyperparameters (hyperparameters) preset to be suitable for quality data, and four machine learning models (810) A method of selecting an optimal model by comparing cross validation performance may be used.

In terms of optimization, after adjusting the hyperparameters for each model, the performance of the models is compared to select the optimal model. However, in this embodiment, since the hyperparameters are empirically adjusted to appropriate values to match the imbalance characteristics of the quality data, the learning time can be minimized by comparing the performance of the model by performing cross-validation after one round of learning. there is.

In particular, the training unit 112 focuses on preventing overfitting by appropriately setting the maximum depth in the training process for the four models 810 .

The training unit 112 checks whether training for all four models 810 has been performed (S1110), and if unlearned models remain, continuing to learn and verify them (S1102 to S1106) carry out

After learning of the four models 810 is completed, the training unit 112 compares performance between the four models and selects an inference model (S1112).

The training unit 112 selects a model having the highest F1 score as a final inference model. However, when selecting a final model, the user may use recall as a selection criterion if the purpose is to reduce defects or precision if the purpose is to reduce false defects.

The training unit 112 performs hyperparameter optimization on the selected inference model (S1114).

As described above, for an inference model trained using the method for reducing a required learning time, the training unit 112 adjusts hyperparameters within an appropriate range to improve performance. A grid search can be used as a representative method, but it has a disadvantage in that it takes a long time because performance is checked for all hyperparameter settings.

To improve this, the training unit 112 according to the present embodiment may adjust hyperparameters based on a random search method. In the random search, hyperparameters are arbitrarily set and performance of the inference model is checked, but random settings and performance checks may be performed a preset number of times. The training unit 112 may optimize the hyperparameters by finding the hyperparameters for the case showing the best performance.

In another embodiment of the present disclosure, the training unit 112, when performing a random search for a preset number of times, with respect to an arbitrary hyperparameter, when an inference model satisfies a preset performance, selects an arbitrary hyperparameter. After selecting the optimal hyperparameter, the random search can be terminated.

As described above, according to the present embodiment, it is possible to reduce the learning time of the inference model to a minimum by applying hyperparameter optimization only to the inference model and performing optimization based on random search.

A device (not shown) on which the analysis system 100 according to this embodiment is mounted may be a programmable computer and includes at least one communication interface capable of connecting to a server (not shown).

Training for the inference model as described above may be performed in a device on which the analysis system 100 is mounted using computing power of a device on which the analysis system 100 is mounted.

Training for the inference model as described above may be performed in the server. The training unit of the server may perform training on a machine learning model having the same structure as an inference model, which is a component of the analysis system 100 mounted on the device. The server may transmit parameters of the trained machine learning model to the device using a communication interface connected to the device, and the analysis system 100 may set parameters of the inference model using the received parameters. In addition, parameters of an inference model may be set at a point in time when the analysis system 100 is loaded into a device.

12 is a flowchart of a quality data analysis method according to an embodiment of the present disclosure.

The analysis system 100 acquires quality data for the product (S1200). Quality data can be collected for a plurality of process factors applied or occurring in the production process of a product. A process factor input for quality data analysis may be a major process factor selected in a pre-training process for an inference model.

The analysis system 100 may set a data type such as a categorical type or a numeric type for a process factor used as an input of an inference model. In this case, the analysis system 100 may obtain input necessary for analysis (eg, quality data, type of process factor, etc.) from a user by using the UI unit 110 .

Meanwhile, the analysis system 100 sets a process factor used as a target output as a target factor.

The analysis system 100 performs a pre-processing process on the quality data (S1202).

For the categorical process factor, an encoding process may be performed to convert it into an embedding value suitable for the inference model.

An example of categorical data is a target factor indicating whether a field claim has occurred for a product. Accordingly, encoding of the target factor may be a process of generating an analysis label for quality analysis based on an inference model.

In addition, the analysis system 100 may set values of process factors that are omitted in the collection process. For example, the numerical process factor may be set to a median value, and the categorical process factor may be set to a mode value.

The analysis system 100 uses an inference model based on a plurality of preprocessed process factors to generate a determination result on whether the product is good or bad (S1204). Here, the determination result may be a probability value for the quantity or defect of the product.

The inference model is implemented in the form of a machine learning model, and may be a model implemented with one of four machine learning algorithms such as a tree-based decision tree, random forest, XGBoost, or LightGBM. Using the training process, the training unit 112 may select, as an inference model, a model with the best performance among models to which each of the four machine learning algorithms is applied.

The analysis system 100 generates an analysis report on product quality based on a plurality of process factors, analysis labels, and judgment results (S1206). In order to comprehensively/microscopically represent the effect of each process factor on the judgment result (amount or defect of the product), the analysis report includes analysis data summary (202), process factor importance (204), data distribution by process factor (206), and all or part of the analysis result 208 .

The analysis system 100 provides an analysis report to the user using the UI unit 110 (S1208).

13 is a flowchart of a method of changing a quality control criterion based on a simulator according to an embodiment of the present disclosure.

The simulator selects an adjustment process factor for adjusting the factor value and obtains an adjustment factor value for the adjustment process factor (S1300). After selecting the adjusted process factors using the UI unit 110 as illustrated in FIG. 7 , the simulator may acquire adjusted process factor values for these values from the user.

A check box can be used to select the adjustment process factor. For the process factor whose checkbox is selected, the process factor value can be adjusted according to the data type.

If the selected process factor is a categorical type, a user-desired process factor value category can be selected using a check box. In the case of numeric type, the value of the process factor can be adjusted using the slider.

In addition, with reference to the T-test result for the process factor, the corresponding process factor value may be adjusted to minimize the distribution of defects in the product.

The adjusted process factor may be all or part of the main process factors selected in the pre-training process for the inference model.

In addition, as the adjustment process factor, the above-described input factor may be selected.

The simulator generates a probability of whether the product is good or bad based on the adjusted process factor using the inference model (S1302). As described above, the decision result generated by the inference model may be a probability value for the quantity or defect of the product.

The inference model is implemented in the form of a machine learning model, and may be a model implemented with one of four machine learning algorithms such as a tree-based decision tree, random forest, XGBoost, or LightGBM. Using the training process, the training unit 112 may select, as an inference model, a model with the best performance among models to which each of the four machine learning algorithms is applied.

The simulator checks whether the probability of failure is less than a predetermined standard probability (S1304). If the probability of failure is greater than or equal to the reference probability, the simulator newly acquires the adjusted process factor and repeats the simulation process (S1300 to S1304).

If the probability of failure is less than the reference probability, the simulator selects the adjustment factor value as the optimal factor value for the adjustment process factor (S1306).

The simulator changes the quality control criteria for the adjusted process factor based on the optimal factor value (S1308). The changed quality control standards can be applied to the production process for future products.

14 is a flowchart of a method for training an inference model according to an embodiment of the present disclosure.

The training unit 112 acquires quality data about the product to use in training of the reasoning model (S1410). Quality data can be collected for a plurality of process factors applied or occurring in the production process of a product.

The training unit 112 may set data types for process factors used for training.

On the other hand, the quality data may include a factor that can be used as a target output (ie, a label for learning) in the training process of an inference model (eg, whether a field claim has occurred for a product). The training unit 112 sets a factor used as a target output as a target factor.

The category and target factor for the process factor may be set using the UI unit 110 as described above.

The training unit 112 performs a pre-processing process on the quality data (S1402).

The training unit 112 may perform an encoding process of converting the categorical process factor into an embedding value suitable for the inference model. In addition, the training unit 112 may set values of process factors omitted in the collection process. For example, a numerical process factor may be set as a median value, and a categorical process factor may be set as a mode value.

Encoding of the target factor, which is categorical data, may be a process of generating a learning label for training an inference model.

The training unit 112 selects a major process factor having a high influence on the target factor from a plurality of process factors included in the quality data (S1404).

If the number of process factors is greater than the preset number, the training unit 112 may select the major process factors to be less than or equal to the preset number by performing the process of calculating the major process factors as described above. All or part of a T-test, comparison of information gains, and correlation analysis may be performed as a process of calculating the main process factor.

The training unit 112 performs data balancing on bad data with respect to the main process factor (S1406). The training unit 112 may increase the number of bad data by upsampling bad data, thereby achieving a balance between bad data and good data.

The training unit 112 performs training on the four machine learning models 810 (S1408).

The training unit 112 divides the balanced quality data into training data and verification data, and then performs training on each of the four machine learning models 810 based on the training data and training labels. In addition, the training unit 112 performs cross-validation on the learned machine learning models based on data for verification, and then stores the performance of each model.

The training unit 112 may focus on preventing overfitting by appropriately adjusting the maximum depth in the training process for the four models 810 .

After learning of the four models 810 is completed, the training unit 112 selects an optimal model as an inference model by comparing performances of the four models 810 (S1410). The training unit 112 may select a model having the highest F1 score as a final inference model.

The analysis system 100 according to the present embodiment uses a method of improving quality control standards of process factors in order to solve the bias problem of process factors included in quality data.

Hereinafter, a method of improving quality control standards of process factors performed by the analysis system 100 will be described using FIGS. 15 and 16 .

15 is a schematic configuration diagram of an apparatus for improving quality control standards for process factors according to an embodiment of the present disclosure.

The quality control standard improvement apparatus according to the present embodiment is included in the analysis system 100, and the process factor having a low influence is based on the degree of influence between the process factor and the field claim for the product (hereinafter referred to as 'influence'). Adjust quality control standards for The quality control standard improvement device may include all or part of the input unit 102, the influence analysis unit 1504, the management range adjustment unit 1506, the data re-collection unit 1508, and the segmentation collection unit 1510.

The input unit 102 obtains quality data and field claims about the product. Quality data can be collected for a plurality of process factors applied or occurring in the production process of a product. On the other hand, the field claim may indicate the quantity or defect of the product, and may be set as a target factor for future training.

The influence analysis unit 1504 analyzes the degree of influence between process factors included in the quality data and field claims. As a method of analyzing the degree of influence, methods (eg, T-test, calculation of information gain, correlation analysis, etc.) used in the process of selecting key process factors as described above may be used. Based on this impact analysis, the impact analyzer 1204 may arrange the process factors in order of impact.

That is, the influence analysis unit 1504 selects a process factor having statistical significance with the quantity or defect of the product by using the T-test. The influence analyzer 1504 may compare information gains of the selected process factors and generate an array in the order of highest information gain. The influence analysis unit 1504 analyzes the correlation between the arrayed process factors, and removes one of the two process factors (eg, a process factor having a low information gain) from the array in the case of two process factors having a correlation coefficient greater than a predetermined reference value. This is because conflicting adjustment results may occur when the management scope is adjusted for both process factors with high correlation. Accordingly, the order of influence may be a higher order of information gain in which statistical significance and correlation are reflected.

The management range adjustment unit 1506 expands the management range of biased process factors for biased process factors whose influence does not fall within the top 20%. In order to expand the management range of the biased process factor, the analysis system 100 may expand the range of data collected in the production process by lowering the lower limit value of the existing management range or increasing the upper limit value of the existing management range.

The data recollecting unit 1508 recollects quality data based on the expanded management range. Using the storage device included in the analysis system 100 or the server, the data recollector 1208 may recollect and store the quality data. Depending on the nature of the production process or process parameters, this recollection process may take days, weeks, or months or longer.

Meanwhile, the influence analysis unit 1504 may analyze the degree of influence between process factors and field claims included in the recollected quality data after the management scope is adjusted. Based on the impact analysis, process factors can be rearranged in order of impact. In the recollected quality data, the influence analysis unit 1204 may maintain the existing management range for biased process factors whose influence does not fall within the top 20%.

The data segmentation collection unit 1510 subdivides and recollects the data within the management range for the process factors whose influence is within the top 20% in the input quality data or the recollected quality data. By subdividing and recollecting data within the management scope, quality data can exist evenly within the management scope.

16 is a flowchart of a method for improving quality control standards of process factors according to an embodiment of the present disclosure.

The analysis system 100 analyzes the degree of influence between process factors included in the quality data and field claims (S1600). As a method of analyzing the degree of influence, methods (eg, T-test, calculation of information gain, correlation analysis, etc.) used in the process of selecting key process factors as described above may be used. Based on the impact analysis, the analysis system 100 may arrange the process factors in order of impact. Here, the order of influence may be a higher order of information gain in which statistical significance and correlation are reflected.

The analysis system 100 checks whether the influence of the process factor is within the top 20% (S1602).

The analysis system 100 expands the management range of the biased process factor with respect to the biased process factor whose influence does not fall within 20% (S1604). In order to expand the management range of the biased process factor, the analysis system 100 may expand the range of data collected in the production process by further lowering the lower limit value or increasing the upper limit value of the existing management range.

The analysis system 100 re-collects quality data based on the expanded management scope (S1606). Using a storage device included in the analysis system 100 or the server, the analysis system 100 may recollect and store quality data.

The analysis system 100 analyzes the influence between the process factor and the field claim for the process factor recollected with the management scope expanded (S1608). Based on the impact analysis, the analysis system 100 may rearrange the process factors in order of impact.

The analysis system 100 checks whether the influence of the process factor is within the top 20% (S1610).

The analysis system 100 maintains the existing management range for biased process factors that do not have an influence within the top 20% (S1612).

On the other hand, the analysis system 100 subdivides and recollects data within the management range for process factors whose influence is within the top 20%, including biased process factors whose management range is expanded (S1602 and S1610) (S1602 and S1610). S1614). By subdividing and recollecting data within the management scope, quality data can exist evenly within the management scope.

As described above, according to the present embodiment, by providing an analysis system that subdivides and recollects data within the management scope, it is possible to reduce the influence of bias of process factors and increase the efficiency of quality data analysis. there is

In this embodiment, as described above, a product to be analyzed for quality may be a part included in a vehicle, such as a gearbox. Compared to the entire vehicle, which is a complex system, the gearbox is a system of an appropriate scale to perform quality analysis based on the analysis system 100 according to the present embodiment. That is, the machine learning-based reasoning model may model a causal relationship between a plurality of process factors and field claims (ie, product quantity or defect) for the gearbox by using a training process. In addition, by using the trained inference model, by adjusting the quality control criteria for specific process factors, it is possible to reduce the rate of defective gearboxes.

On the other hand, the process factors constituting the quality data for the gearbox include pinion plug fastening torque, lock ring press-in depth, pinion grease application amount, lock ring caulking amount, pinion plug LDVT height, caulking load, 4-point bearing press-in depth, rack These include bar load (LH direction), rack bar load (RH direction), and yoke press-in load. Since this embodiment is an invention for quality analysis for a product such as a gearbox, a detailed description of the process factors of the gearbox will be omitted.

Hereinafter, a case in which the analysis system 100 according to the present embodiment is applied to quality analysis of a gearbox will be described.

17 is a flowchart of a process of applying an analysis system according to an embodiment of the present disclosure to a gearbox.

First, the process of selecting an inference model used for quality analysis of a gearbox will be described.

The analysis system 100 acquires quality data for the gearbox (S1700).

In general, unlike field claim data managed in the quality system, process factor data is separately accumulated and managed in the manufacturing system (Manufacture Executive System: MES), so the two data must be integrated for quality analysis. It is possible to integrate the two data based on the product identifier (ID) of the gearbox, and two methods can be used.

The first is a method of classifying and integrating process factor data according to the type of field claim. In the case of a gearbox, since there are various field claims such as vibration, noise, and breakage, data analysis can be performed by classifying each of them into classes. The first method has the advantage of enabling detailed cause analysis for each type of field claim, but has a problem that the analysis result is biased when there is little data for each type of field claim.

Second, regardless of the type of field claim, it is a method of integrating process factor data by dividing it into positive or defective data depending on whether or not a field claim has occurred. This method has the advantage that the task of classifying data is simple and time consuming, and even if there is little data on some field claims, universal analysis is possible. In this embodiment, the analysis system 100 acquires integrated quality data according to the second method, but is not necessarily limited thereto. In another embodiment according to the present disclosure, an inference model may be trained to infer whether a field claim has occurred according to a class.

Meanwhile, as described above, the integrated quality data may be used for training of an inference model.

The analysis system 100 may set data types for process factors used for training. Here, the data type of the process factor may include a numeric type expressed as a numerical value, a categorical type expressed as a character, and a time type including time information at which data was collected.

On the other hand, the integrated quality data includes a factor that can be used as a target output (ie, label for learning) in the training process of the inference model (eg, whether a field claim has occurred for the gearbox). The analysis system 100 sets a factor used as a target output as a target factor.

The category and target factor for the process factor may be set using the UI unit 110 as described above.

As illustrated in FIG. 9 , the analysis system 100 performs pre-processing on quality data (S1702).

It is determined that the time for which the quality data is collected has little correlation with the quantity or defect of the product, and the time-type process factor is removed from the quality data for training.

In the case of categorical data, the analysis system 100 performs an encoding process of converting into an embedding value suitable for an inference model.

An example of categorical data is a target factor indicating whether a field claim has occurred for a product. Encoding of these target factors is a process of generating learning labels for training an inference model.

For numeric data and encoded categorical data, the analysis system 100 processes missing data generated in the collection process. In this case, the categorical data may be set as the most frequent value, and the numeric data may be set as the median value. On the other hand, in the case of a process factor with severe omission, it may interfere with the training of the inference model. Accordingly, the analysis system 100 may remove a process factor having a higher omission rate than a predetermined rate in the training process.

Meanwhile, when the number of process factors is greater than 20, the analysis system 100 may perform a process of selecting main process factors for training, as illustrated in FIG. 10 . In the example of FIG. 17 related to the gearbox, the process of selecting key process factors is omitted, because the number of process factors constituting the quality data of the gearbox was 20 or less.

The analysis system 100 performs data balancing on process factors (S1704).

The analysis system 100 upsamples bad data to augment the number of bad data, thereby achieving a balance between bad and good data. For example, the analysis system 100 may generate similar data within a data distribution using kNN model techniques.

The analysis system 100 performs learning on the four machine learning models 810, and then selects an optimal model and determines it as an inference model (S1706).

As illustrated in FIG. 11 , the analysis system 100 performs training on four machine learning models 810 using quality data related to the gearbox and labels for learning, so that the four machine learning models 810 An inference model can be selected based on the learning performance for .

In this embodiment, as a result of performing training on the four machine learning models 810, a model implementing a random forest algorithm is selected as an inference model. In the training process, we optimized the height parameters around the maximum depth to achieve optimal performance for the inference model.

Meanwhile, a commonly used algorithm is a boosting algorithm such as XGBoost or LightGBM. However, due to the nature of quality data with severe data imbalance, since a method of preventing overfitting by adjusting the maximum depth is used, bagging algorithms such as random forest can produce better results.

Hereinafter, a process of adjusting a process factor of a gearbox using an inference model will be described.

The analysis system 100 selects process factors for adjusting quality control standards (S1708).

The analysis system 100 analyzes the influence of the process factor on whether or not a field claim occurs (S1730 to S1734), and selects a process factor having a large influence.

The analysis system 100 compares the feature importance for the process factor (S1730).

The analysis system 100 generates an inference model implementing a random forest algorithm in a training process. Process factors are primarily selected by comparing the importance of characteristics.

18 is an exemplary diagram illustrating importance of characteristics of process factors of a gearbox according to an embodiment of the present disclosure.

In the example of FIG. 18, the meaning of 'worst' indicates that the characteristic importance of the process factor is greater than the predetermined reference value, and thus has a great effect on the occurrence of field claims. On the other hand, as described above, this feature importance may be provided to the user using the UI unit 110 as a part of the analysis report.

The analysis system 100 performs a T-test on process factors (S1732). The analysis system 100 secondarily selects the process factor by performing a T-test to verify the significance of the gearbox amount or defect distribution for the process factor having a high importance of the characteristic.

19 is an exemplary diagram illustrating a T-test according to an embodiment of the present disclosure.

Here, the process factor of the gearbox used as an example of the T-test is 'lock ring press-in depth', and the test result shows that it is significant. On the other hand, as described above, the data distribution for each process factor, which is the basis of the T-test, may be provided to the user using the UI unit 110 as a part of the analysis report.

The analysis system 100 checks the correlation with the process factor (S1734).

The analysis system 100 checks the correlation between the process factors that have passed the T-test, and then selects a process factor having a high characteristic importance when the correlation coefficient between the two process factors is greater than a predetermined reference value.

In conclusion, the process factors illustrated in FIG. 18 represent process factors determined to have a large influence on the occurrence of defects in the gearbox according to the above-described effect analysis. In addition, these are all process factors (ie, input factors) that can be adjusted.

The analysis system 100 changes the quality control criteria for the selected process factor (S1710)

Based on the distribution of process factors, quality control standards can be changed to minimize the distribution of gearbox defects within an adjustable management range. At this time, the distribution of the process factor may be used to change the quality control standard after parameters defining the distribution are estimated in advance.

Table 1 shows the process parameters of the gearbox before and after the change in quality control standards.

For example, as illustrated in FIG. 19 , the quality control criterion for the lock ring press-in depth may be changed to minimize the distribution of defects in the gearbox based on the T-test result.

The analysis system 100 may check whether the quality control criteria have been appropriately changed by using the inference model as a simulator and generating a probability value for the amount or defect of the gearbox according to the changed quality control criteria. For example, when the probability of defective gearbox is greater than or equal to the reference probability, it is possible to repeatedly check whether the quality control standard has been appropriately changed by obtaining the quality control standard and regenerating the determination result.

Regarding the process factors shown in Table 1, if the quality control standards after the change are applied to the gearbox production process, it is expected that the defect rate due to the process factor can be reduced from a minimum of 10% to a maximum of 90%. . In fact, as a result of applying the changed quality control standards to the gearbox production process for process factors such as lock ring caulking amount, pinion plug LVDT height, and 4-point bearing press-in depth, it was confirmed that the decrease in defect occurrence rate due to the corresponding process factor was confirmed. did

In the above description, the inference model is used for changing the quality control standard, but the inference model can also be used for quality analysis. For example, an inference model may be used to apply a new quality control standard to a gearbox production process and to confirm characteristics of quality data collected in a later production process.

In each flowchart according to the present embodiment, it is described that each process is sequentially executed, but is not necessarily limited thereto. In other words, since it will be applicable to change and execute the process described in the flowchart or to execute one or more processes in parallel, the flowchart is not limited to a time-series sequence.

Various implementations of the systems and techniques described herein may include digital electronic circuits, integrated circuits, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), computer hardware, firmware, software, and/or their can be realized in combination. These various implementations may include being implemented as one or more computer programs executable on a programmable system. A programmable system includes at least one programmable processor (which may be a special purpose processor) coupled to receive data and instructions from and transmit data and instructions to a storage system, at least one input device, and at least one output device. or may be a general-purpose processor). Computer programs (also known as programs, software, software applications or code) contain instructions for a programmable processor and are stored on a “computer readable medium”.

A computer-readable recording medium includes all kinds of recording devices that store data that can be read by a computer system. These computer-readable 　recording media include non-volatile or non-transitory media such as ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, and storage device. can be a medium. In addition, the computer-readable 　recording medium may be distributed in computer systems connected through a network, and computer-readable codes may be stored and executed in a distributed manner.

The above description is merely an example of the technical idea of the present embodiment, and various modifications and variations can be made to those skilled in the art without departing from the essential characteristics of the present embodiment. Therefore, the present embodiments are not intended to limit the technical idea of the present embodiment, but to explain, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The scope of protection of this embodiment should be construed according to the claims below, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of rights of this embodiment.

100: quality data analysis system
102: input unit 104: data pre-processing unit
106: determination unit 108: data visualization unit
110: UI unit 112: training unit
202: Summary of analysis data 204: Importance of process factors
206: Data distribution by process factor
208: Analysis result
302: process factor adjustment unit 304: judgment result output unit
306: important factor output unit 308: standard application unit
806: process factor selection unit 808: data balancing unit
1504: Impact Analysis Unit 1506: Management Scope Adjustment Unit
1508: data re-collection unit 1510: data segmentation collection unit

Claims (10)

A method for improving quality control criteria for a product, performed by a computing device, comprising:
A process of acquiring quality data and field claims for the product, wherein the quality data is collected for a plurality of process features applied or occurring in the production process of the product, and the field Claims indicate quantity or defect in respect of said product;
analyzing a first influence between the plurality of process factors and the field claim;
expanding an existing management range of the biased process factor in the case of a biased process factor whose first influence does not fall within a predetermined upper ratio among the plurality of process factors;
recollecting quality data for the plurality of process factors based on the expanded management scope;
analyzing a second influence between the plurality of process factors and the field frame with respect to the recollected quality data; and
Process of subdividing and recollecting the data of the corresponding process factor when the first or second influence falls within the predetermined upper ratio
Including, how to improve quality control standards. According to claim 1,
In the case of a biased process factor whose second influence does not fall within the predetermined upper ratio, a step of maintaining the management range of the biased process factor within the existing management range. According to claim 1,
The process of analyzing the first influence,
performing a T-test for verifying significance with respect to the distribution of field claims for the plurality of process factors;
A process of comparing information gains for process factors that have passed the T-test and generating an array in an order of high information gain, wherein the information gain is information about the quantity or defect of the product generating by subtracting information on quantity or defect after branching by the process factor from ; and
A process of analyzing the correlation between the process factors having high information gain and, in the case of two process factors having a correlation coefficient greater than a predetermined reference value, removing one of them from the sequence.
Including, how to improve quality control standards. According to claim 1,
The process of analyzing the first influence and the process of analyzing the second influence are mutually the same, a method for improving quality control standards. According to claim 1,
The process of expanding the existing management scope,
A method for improving quality control standards by further lowering the lower limit value of the existing management range or further increasing the upper limit value. An input unit for obtaining quality data and field claims for the product, wherein the quality data is collected for a plurality of process features applied or occurring in the production process of the product, and the field Claims indicate good or bad quality of the product;
an impact analysis unit analyzing a first influence between the plurality of process factors and the field claim;
a management range adjusting unit expanding an existing management range of the biased process factor in the case of a biased process factor whose first influence does not fall within a predetermined upper ratio among the plurality of process factors;
Based on the expanded management scope, a data re-collecting unit for re-collecting quality data for the plurality of process factors; and
If the first influence falls within the predetermined upper ratio, a data segmentation collection unit that subdivides and recollects the data of the corresponding process factor
Including, quality control standard improvement device. According to claim 6,
The impact analysis unit,
For the re-collected quality data, a second influence between the plurality of process factors and the field frame is analyzed. According to claim 6,
The impact analysis unit,
In the case of a biased process factor whose second influence does not fall within the predetermined upper ratio, the management range of the biased process factor is maintained within the existing management range. According to claim 6,
The impact analysis unit,
For the plurality of process factors, a T-test is performed to verify the significance of the distribution of field claims, and information gains are compared for process factors that pass the T-test to determine whether the information gain is An array is created in order of high information gain, and a correlation between process factors having a high information gain is analyzed, and in the case of two process factors having a correlation coefficient greater than a predetermined reference value, one of the two process factors is removed from the sequence, thereby causing the first effect. , Quality control standard improvement apparatus for analyzing the degree or the second influence. A computer program stored in a computer-readable recording medium to execute each step included in the method for improving quality control standards according to any one of claims 1 to 5.

Images