Method and system for updating an AI model for classifying and determining the causes of defects.
Unsupervised learning-based periodic updates of AI models for defect classification address accuracy degradation by automatically replacing models with high-accuracy versions, maintaining consistent performance in product defect judgment.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- IBIDEN CO LTD
- Filing Date
- 2024-11-07
- Publication Date
- 2026-05-19
AI Technical Summary
Conventional AI models for classifying defect causes in products like printed circuit boards suffer from decreased accuracy over time due to specification changes, necessitating burdensome retraining by analysts, which hinders consistent high accuracy rates.
Implement unsupervised learning to periodically update the AI model by classifying data into clear and gray categories, using a predetermined percentage of clear data for training and maintaining a high accuracy rate through automatic model replacement.
Maintains consistently high accuracy in defect classification without burdening analysts by automatically updating the AI model, ensuring accurate judgment results.
Smart Images
Figure 2026082401000001_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and system for updating an AI model for classifying and determining the causes of product defects. In this specification, the term "product" includes "intermediate products" during the manufacturing process of finished products.
Background Art
[0002] When classifying and determining the causes of defects in, for example, printed circuit boards as products, conventionally, a determination method using an AI (artificial intelligence) model described in, for example, Patent Document 1 has been used. In this determination method, a database is created by collecting a plurality of sets of image information of defective locations obtained by analyzing an image of a printed circuit board in which a defect has been found by inspection using a checker head or the like on a production line, and text information added by an analyst regarding the classification information of the defect. Next, a plurality of arbitrarily selected sets of the image information and text information of the defective locations in the database are input into a classifier equipped with an AI model. Then, with supervised learning, the AI model is trained to determine and output the classification of the cause of each defect based on the combination of feature amounts (regularities) in each of the plurality of sets of the image information and text information of the defective locations.
[0003] Subsequently, from the multiple sets of image and text information of defective areas in the database, several sets that clearly indicate the classification of the defect cause are arbitrarily extracted as confirmation data and input into the trained AI model, excluding the classification information. Next, the AI model is made to determine the classification of the defect cause for each set of image and text information based on the combination of features in each set, and the determination of the correctness of the determination is made by comparing the determination result with the previously determined classification information, and the accuracy rate (the percentage of the total data that was determined correctly) is calculated. If the accuracy rate is less than a predetermined value, the image and text information of defective areas in the database are added to the AI model and the supervised learning is repeated. If the accuracy rate is above a predetermined value, the classifier equipped with the AI model is used to classify the defect cause of other printed circuit boards that were found to be defective during inspection on the manufacturing line, and the determination result is output along with the image and text information of the defective areas of those printed circuit boards. [Prior art documents] [Patent Documents]
[0004] [Patent Document 1] Japanese Patent Publication No. 2024-118176 [Overview of the project] [Problems that the invention aims to solve]
[0005] Incidentally, the specifications of the objects to be judged, such as printed circuit boards, usually change gradually over time, and the accuracy of the AI model's classification results decreases as these specifications change. For this reason, in conventional methods, once a certain amount of time has passed and pairs of image and text information of defective areas have accumulated in the database, the analyst inputs these pairs into the AI model and retrains it using supervised learning to classify the cause of the defect. The retrained AI model then replaces the original AI model to update the classifier's AI model. However, retraining the AI model using supervised learning is burdensome for the analyst, so the AI model is not retrained until a certain amount of image and text information of defective areas has accumulated in the database. Therefore, it has been difficult to consistently maintain a high accuracy rate for the judgment results.
[0006] The objective of the present invention is to maintain a consistently high accuracy rate of classification results without burdening the analyst, by having the AI model retrained at predetermined intervals to classify and judge the causes of defects, and by automatically updating the classifier's AI model by automatically replacing the original AI model with an AI model whose classification accuracy rate is above a predetermined level. [Means for solving the problem]
[0007] Figure 1 is a conceptual diagram illustrating the method for updating the AI model for determining the causes of defects according to the present invention. The method for updating the AI model for classifying and determining the causes of defects according to the present invention is as shown in Figure 1. (Step S1) Create a database that includes multiple types of information from inspections or measurements related to the product, and which has accumulated multiple result data over time, where at least one type of information indicates a product defect. From the database, extract the result data from a predetermined number of years ago to the present and input it into an AI model for feature classification. Then, train the AI model for feature classification to learn the number of classifications and the classification method for determining the classification of the result data from a predetermined number of years ago to the present using unsupervised learning (steps S2, S3). Using the aforementioned machine learning AI model for feature classification, the result data from a predetermined number of years ago to the present is broadly classified into clear data, where the features are clear and easy for the AI model to classify based on features, and gray data, where the features are unclear and difficult for the AI model to classify based on features. A predetermined proportion of the clear data is used as training data, and the remainder of the clear data is used as confirmation data. The classification of each result data in the training data and the confirmation data is then determined by features (steps S4 to S8). Step S9 involves inputting the aforementioned training data into a classification AI model and having the classification AI model perform unsupervised learning to classify the training data based on its features, using the number of classifications and the classification method. The machine learning AI model for judgment is used to determine the classification of the verification data, excluding the classification information determined by the feature classification AI model. If the accuracy rate of the determination result is equal to or higher than a predetermined accuracy rate, the machine learning AI model for judgment is used to replace the original AI model for judgment, thereby automatically updating the AI model for judgment (steps S10, S11). If the accuracy rate of the judgment result is less than the predetermined accuracy rate, the AI model for feature classification automatically updates the AI model by adding new result data from a predetermined number of years ago to the current result data, or by replacing the current result data with new result data from a predetermined number of years ago to the current result data (step S12). It includes.
[0008] According to the present invention's method for updating an AI model for determining the causes of defects, an AI model for feature classification, which has been trained using unsupervised learning, is used to train the AI model to classify the causes of defects again using unsupervised learning at predetermined intervals. The AI model with a predetermined or higher accuracy rate for the classification results is then replaced with the original AI model, and the classifier's AI model is automatically updated. This allows for maintaining a consistently high accuracy rate for the classification results without burdening the analyst.
[0009] Furthermore, in the method for updating the AI model for determining the cause of defects according to the present invention, the product may be a printed circuit board. In this way, the accuracy rate of the AI model's determination of the cause of defects in printed circuit boards can always be maintained at a high level.
[0010] Furthermore, in the method for updating the AI model for determining the causes of defects according to the present invention, the ratio of the number of data points for each classification among a predetermined percentage of the clear data may be adjusted in accordance with the ratio of the number of data points for each classification by the AI model for feature classification. In this way, it is possible to effectively prevent the classification of a small amount of training data from excessively influencing the judgment result and lowering the accuracy of the judgment result.
[0011] Furthermore, in the method for updating the AI model for determining the causes of defects according to the present invention, when updating the result data from a predetermined number of years prior to the present, a process may be performed to delete the result data from the predetermined number of years prior to the present in chronological order. In this way, it is possible to effectively prevent the accuracy of the judgment result from decreasing due to older result data having an excessive influence on the judgment result.
[0012] On the other hand, this is a conceptual diagram showing the update system for the AI model for determining the causes of defects according to the present invention. As shown in Figure 2, the update system for the AI model for classifying and determining the causes of defects according to the present invention is as follows: A database (1) which may include multiple types of information from inspections or measurements related to the product, and which has accumulated multiple result data over time in which at least one type of information indicates a product defect, A feature classification AI model machine learning unit (2) extracts result data from a predetermined number of years ago to the present from among the multiple result data in the aforementioned database and inputs it into a feature classification AI model, and uses unsupervised learning to machine-learn the number of classifications and the classification method for determining the classification of the result data from the predetermined number of years ago to the present based on its features, The aforementioned machine-learned AI model for feature classification broadly classifies the result data from a predetermined number of years ago to the present into clear data, where the features are clear and easy for the AI model to classify based on features, and gray data, where the features are unclear and difficult for the AI model to classify based on features. A predetermined percentage of the clear data is used as training data, and the remainder of the clear data is used as verification data. The training data acquisition unit (3) then determines the classification of each result data in the training data and the verification data based on features. The AI model machine learning unit (4) inputs the aforementioned training data into the AI model for classification and causes the AI model to perform unsupervised learning to classify the training data based on the number of classifications and the classification method, using its features. A judgment AI model update unit (5) automatically updates the judgment AI model by replacing the original judgment AI model with the machine learning judgment AI model if the machine learning judgment AI model is used to determine the classification of the verification data excluding the classification information determined by the feature classification AI model, and the machine learning judgment AI model is used to determine the classification of the verification data excluding the classification information determined by the feature classification AI model, and the accuracy rate is equal to or higher than a predetermined accuracy rate. If the accuracy rate of the judgment result is less than the predetermined accuracy rate, the result data update unit (6) automatically updates the result data from a predetermined number of years ago to the present, which is input to the AI model for feature classification, by adding new result data from that point onward to the result data from a predetermined number of years ago to the present, or by replacing the result data from a predetermined number of years ago to the present with new result data from a predetermined number of years ago to the present. It is equipped with.
[0013] In the AI model update system for determining the causes of defects according to the present invention, the feature classification AI model machine learning unit uses unsupervised learning to train the feature classification AI model to determine the number of classifications and the determination method for determining the classification of result data from a predetermined number of years ago to the present based on features. Next, the training data acquisition unit uses the feature classification AI model to broadly classify the result data into clear data and gray data, using a predetermined percentage of the clear data as training data and the remainder of the clear data as verification data. Furthermore, the determination AI model machine learning unit inputs the training data into the determination AI model and uses unsupervised learning to train the determination AI model to determine the classification of the training data based on features according to the number of classifications and the determination method.
[0014] Next, the judgment AI model update unit uses the machine-learned judgment AI model to determine the classification of the verification data, excluding the classification information determined by the feature classification AI model. It then compares the determination result with the classification information, and if the accuracy rate is above a predetermined accuracy rate, it automatically updates the judgment AI model by replacing the original judgment AI model with the machine-learned judgment AI model. Furthermore, if the accuracy rate of the determination result is below the predetermined accuracy rate, the result data update unit automatically updates the result data from a predetermined number of years ago to the present, which is input to the feature classification AI model, by adding subsequent new result data to the result data from a predetermined number of years ago to the present, or by replacing the result data from a predetermined number of years ago to the present with new result data from a predetermined number of years ago to the present.
[0015] Therefore, according to the AI model update system for determining the causes of defects of the present invention, it is possible to maintain a consistently high accuracy rate of the determination results without burdening the analyst.
[0016] In the update system of the AI model for determining the causes of defects in the present invention, the product may be a printed wiring board. By doing so, the correct answer rate of the determination result of the causes of defects in the printed wiring board by the determination AI model can always be maintained high.
[0017] Also, in the update system of the AI model for determining the causes of defects in the present invention, the ratio of the number of data for each classification of a predetermined proportion of the clear data may be adjusted according to the ratio of the number of data for each classification by the AI model for feature classification. By doing so, it is possible to effectively prevent the classification of learning data with a small number of data from having an excessive impact on the determination result and reducing the correct answer rate of the determination result.
[0018] Furthermore, in the update system of the AI model for determining the causes of defects in the present invention, when updating the result data from a predetermined number of data before to the present, a process of deleting the result data in descending order of age from the result data from a predetermined number of data before to the present may be performed. By doing so, it is possible to effectively prevent old result data from having an excessive impact on the determination result and reducing the correct answer rate of the determination result.
Brief Description of the Drawings
[0019] [Figure 1] It is a conceptual diagram showing a method for updating an AI model for determining the causes of defects in the present invention. [Figure 2] It is a conceptual diagram showing an update system of the AI model for determining the causes of defects in the present invention. [Figure 3] It is an explanatory diagram showing a time series of an embodiment of a method for updating an AI model for determining the causes of defects in the present invention using an embodiment of an update system of the AI model for determining the causes of defects in the present invention. [Figure 4] It is an explanatory diagram further showing the time series of the method for updating the AI model for determining the causes of defects in the above embodiment.
Modes for Carrying Out the Invention
[0020] Hereinafter, an embodiment of the method for updating the AI model for determining defect causes of the present invention, using an embodiment of the AI model updating system for determining defect causes of the present invention, will be described with reference to the drawings. Here, Figure 3 is an explanatory diagram showing in chronological order one embodiment of the method for updating the AI model for determining defect causes of the present invention, using one embodiment of the AI model updating system for determining defect causes of the present invention applied to the manufacturing process of a printed circuit board as an example of a product. Figure 4 is an explanatory diagram showing in chronological order the method for updating the AI model for determining defect causes of the above embodiment.
[0021] In the manufacturing process of printed circuit boards (PCBs), multiple pins protruding from a checker head are typically used to simultaneously contact multiple locations on the conductive circuits of the manufactured PCBs to inspect for defects such as insulation failures (circuit shorts) and contact failures (continuity failures). For PCBs with defects, the imaging device of the optical inspection system is used to image the defective areas in order to improve the manufacturing process. From this image information, an analyst determines the classification of the cause of the defect and adds this classification to the text information obtained during the inspection process, such as the detailed location and continuity status of the defective areas.
[0022] The AI model update system and update method for determining the cause of defects in this embodiment update the classification determination of the cause of defects at defective locations in order to improve the manufacturing process of printed circuit boards, based on result data in which at least one type of such information indicates a defect in the specifications of the printed circuit board.
[0023] Furthermore, "Inspection Result Information" may include image information of defective areas obtained by an analyst analyzing images of the manufactured printed circuit board captured by an imaging device, as well as text or audio information such as the classification of defect causes added by the analyst based on their judgment of the image information. In addition, "Measurement Result Information" may include numerical information such as the dimensions of each part of the manufacturing equipment, the location of defects on the manufactured printed circuit board, the defective layer number, the defective sheet number, the planar coordinates of the defective area, and the temperature and concentration of the electroplating solution used when forming the conductor layer of the printed circuit board.
[0024] To update the classification and determination of the defect occurrence factors described above, the update method of this embodiment first accumulates the result data, which is the omni-data, sequentially in a database composed of a hard disk drive (HDD) or solid-state drive (SSD), etc., as time progresses. For example, result data 1 to result data 6 in the figure, that is, at predetermined intervals such as every hour, every day, or every month. Here, "present" moves from left to right in the figure as time progresses. Therefore, the above database corresponds to database 1 in the conceptual diagram of Figure 2.
[0025] Next, as shown in the flowchart on the left half of Figure 3, if we consider the present time as the point in time when result data 2 was accumulated in the database, then multiple result data from the database, from a predetermined number of points in the past to the present, i.e., from result data 1 to result data 2, are extracted and input into a feature classification AI (artificial intelligence) model, which is a normal computer operated by a predetermined program. This feature classification AI model, using functions such as a known recursive neural network, digitizes the audio and text information of the multiple result data from a predetermined number of points in the past to the present as needed, and then uses unsupervised learning, preferably deep learning, to machine learn the number of classifications and the classification method for determining the classification of these result data based on their features, based on the commonality of the features. Therefore, the machine learning process performed by the feature classification AI model corresponds to the feature classification AI model machine learning unit 2 in the conceptual diagram of Figure 2.
[0026] Next, this feature classification AI model broadly classifies the input result data 1 and result data 2 into clear data, which has clear features and is easy to classify using the AI model's neural network, and gray data, which has unclear features and is difficult to classify using the AI model's neural network. Then, a predetermined percentage, for example 85%, of the clear data is randomly extracted and designated as training data A, and the remaining 15% of the clear data is designated as confirmation data X. The classification of each result data in training data A and confirmation data X is then determined by features. Therefore, the process in which the feature classification AI model classifies the result data into clear data and gray data, and further divides the clear data into training data and confirmation data, corresponds to the training data acquisition unit 3 in the conceptual diagram of Figure 2.
[0027] In this embodiment, the proportion of data points for each classification in 85% of the clear data is adjusted according to the proportion of data points for each classification determined by the feature classification AI model to create training data A. This effectively prevents the classification of a small amount of training data from excessively influencing the subsequent judgment results of the judgment AI model, which would otherwise lower the accuracy of the judgment results.
[0028] Next, in this embodiment, training data A is input to a judgment AI model, which is a standard computer operating under a predetermined program. The judgment AI model is then trained through unsupervised learning to classify the training data A based on the number of classifications and the classification method previously learned by the feature classification AI model. Therefore, the process of training the judgment AI model corresponds to the judgment AI model machine learning unit 4 in the conceptual diagram of Figure 2. Note that the computer constituting the feature classification AI model and the computer constituting the judgment AI model may be the same or separate.
[0029] Next, in this embodiment, the machine-learned AI model A for judgment is used to determine the classification of the verification data X (data that serves as a standard for verifying the performance of the machine-learned AI model for judgment) excluding the classification information determined by the feature classification AI model, and the judgment result is compared with the classification information that was previously excluded. If the accuracy rate is above a predetermined accuracy rate, for example 90% or higher, the machine-learned AI model A for judgment is replaced with the original AI model for judgment, thereby automatically updating the AI model for judgment installed in the classifier on the printed circuit board manufacturing line, as shown in the lower part of Figure 3. Therefore, the process of determining the classification of the verification data X using the machine-learned AI model for judgment and automatically updating the AI model for judgment by replacing the original AI model with the machine-learned AI model A if the accuracy rate is above a predetermined accuracy rate corresponds to the AI model for judgment update unit 5 in the conceptual diagram of Figure 2.
[0030] Furthermore, in this embodiment, the classification of the verification data X excluding the classification information described above is determined, and the classification result is compared with the classification information previously excluded. If the accuracy rate of the determination result is less than the predetermined accuracy rate, for example, less than 90%, it is determined that there is insufficient result data. Then, the result data to be input to the AI model for feature classification is automatically updated by adding result data 3 and / or result data 4, which are new result data extracted from the database from the previous extraction to the present, to the result data from a predetermined number of years prior to the current extraction time. Alternatively, the result data to be input to the AI model for feature classification is automatically updated by replacing at least the older of result data 1 and result data 2, result data 1, with result data 3 and result data 4, which are new result data up to the present. Therefore, the process of automatically updating the AI model for feature classification by adding new result data to the result data from a predetermined number of years prior to the present, or by replacing the result data from a predetermined number of years prior to the present with new result data, corresponds to the result data update unit 6 in the conceptual diagram of Figure 2.
[0031] In the above embodiment, when initially training the AI model for feature classification and the AI model for judgment, supervised learning may be performed based on the correct answers provided by the analyst, instead of unsupervised learning, in order to ensure initial judgment accuracy.
[0032] Next, in this embodiment, assuming that the present time is when, for example, result data 5 has been accumulated in the database over time, multiple result data from the database, from a predetermined number of years prior to the present, i.e., from result data 3 to result data 5, are extracted and input into the feature classification AI model, as shown in the flowchart on the right half of Figure 3. Then, the feature classification AI model, which has been trained in an unsupervised manner using these result data to determine the classification of these result data based on their features, is made to determine the classification of these result data. From thereafter, similar to the flowchart on the left half of Figure 3, the result data is classified into clear data and gray data, the number of data for each classification of 85% of the result data randomly extracted from the clear data is adjusted to create training data B, and the judgment AI model B is trained in an unsupervised machine learning manner using the new training data B+A, which is created by adding training data A from the flowchart on the left half to training data B.
[0033] Furthermore, the remaining 15% of the cleared data results, after removing the classification information previously determined by the feature classification AI model, are combined with the confirmation data Y from the left half of the flowchart to create new confirmation data Y+X. The classification of this confirmation data Y+X is then determined by the machine learning-developed judgment AI model B. The determination result is then compared with the classification information previously removed, and if the accuracy rate is above a predetermined accuracy rate, for example, 90% or higher, the machine learning-developed judgment AI model B is replaced with the original judgment AI model. This automatically updates the judgment AI model installed in the classifier on the printed circuit board manufacturing line, as shown in the lower part of Figure 3.
[0034] On the other hand, in this embodiment, the classification of the verification data Y+X excluding the classification information described above is determined, and the classification result is compared with the classification information previously excluded. If the accuracy rate of the determination result is less than the predetermined accuracy rate, for example, less than 90%, it is determined that there is insufficient result data. Then, the result data to be input to the AI model for feature classification is automatically updated by adding result data 6 and / or subsequent result data, which are new result data extracted from the database from a predetermined number of years prior to the current time of extraction, to result data 3 to result data 5, which are result data from the previous extraction to the present time. Alternatively, the result data to be input to the AI model for feature classification is automatically updated by replacing at least the older result data 3 of result data 3 to result data 5 with result data 6 and subsequent result data, which are new result data up to the present time.
[0035] Furthermore, in this embodiment, if we consider the point in time when result data 7 is accumulated in the database as the new present, then, as shown in the flowchart on the left half of Figure 4, result data 6 and result data 7 are used as the result data extracted from the database up to the new present, instead of result data 3 to result data 5 from the previous update of the judgment AI model. Also, as the training data for judgment AI model C, training data C+B+A is used instead of training data B+A from the previous update of the judgment AI model, and as the verification data for judgment AI model C, verification data Z+Y+X is used instead of verification data Y+X from the previous update of the judgment AI model. Then, as the new result data to be input to the feature classification AI model, at least result data 8 from result data 8 and result data 9 is used, and the new machine-learned judgment AI model C replaces the original judgment AI model B, thereby automatically updating the judgment AI model installed in the classifier on the printed circuit board manufacturing line, as shown in the lower part of Figure 4.
[0036] Furthermore, in this embodiment, if we consider the point in time when result data 10 is accumulated in the database as the new present, then, as shown in the flowchart on the right half of Figure 4, result data 8 to 10 are used as the result data up to the new present extracted from the database, instead of result data 6 to 7 from the previous update of the AI model for judgment. Also, as the training data for the AI model for judgment D, training data D+C+B is used instead of training data C+B+A from the previous update of the AI model for judgment (old training data A is deleted), and as the verification data for the AI model for judgment D, verification data α+Z+Y is used instead of verification data Y+X from the previous update of the AI model for judgment (old verification data X is deleted). Then, by replacing the original AI model C with a new AI model D for judgment that has been trained using result data 11 and at least result data 11 from the subsequent result data as new result data to be input into the AI model for feature classification, the AI model for judgment installed in the classifier installed on the printed circuit board manufacturing line is automatically updated, as shown in the lower part of Figure 4.
[0037] As described above, according to the method and system for updating the judgment AI model of this embodiment, when the accuracy rate is above a predetermined accuracy rate, for example 90% or higher, the machine-learned judgment AI model is replaced with the original judgment AI model, thereby automatically updating the judgment AI model installed in the classifier installed on the printed circuit board manufacturing line. This makes it possible to maintain a consistently high accuracy rate of judgment results without burdening the analyst. [Examples]
[0038] The applicant of this application actually used the classification method for the causes of defects in printed circuit boards according to the above embodiment, using the classification system for the causes of defects in printed circuit boards according to the above embodiment, for approximately one year up to the time of filing this application. The results showed that the accuracy rate of the judgment results of the AI model installed in the classifier installed on the printed circuit board manufacturing line could be kept consistently high without burdening the analyst.
[0039] Although the present invention has been described above based on the illustrated embodiments, it is not limited to the embodiments described above. For example, the defect classification method and classification system of the above embodiments can also be used to classify defect causes of products other than printed circuit boards. [Explanation of symbols]
[0040] 1 Database 2. AI Model for Feature Classification: Machine Learning Section 3. Training data acquisition unit 4. AI Model for Judgment: Machine Learning Section 5. AI model update section for judgment 6. Result Data Update Unit
Claims
1. This involves creating a database that includes multiple types of information related to inspections or measurements performed on a product, and which accumulates multiple result data over time, where at least one type of information indicates a product defect. From the database, extract the result data from a predetermined number of years prior to the present and input it into an AI model for feature classification. Then, use unsupervised learning to train the AI model for feature classification to determine the number of classifications and the classification method for determining the classification of the result data from the predetermined number of years prior to the present based on its features. The aforementioned machine-learned AI model for feature classification broadly classifies the result data from a predetermined number of years ago to the present into clear data, where the features are clear and easy for the AI model to classify based on those features, and gray data, where the features are unclear and difficult for the AI model to classify based on those features. A predetermined percentage of the clear data is used as training data, and the remainder of the clear data is used as confirmation data. The classification of each result data in the training data and the confirmation data is then determined based on its features. The aforementioned training data is input into a judgment AI model, and the judgment AI model is trained through unsupervised learning to classify the training data based on the number of classifications and the classification method, using its features. The machine learning-developed AI model for judgment determines the classification of the verification data, excluding the classification information determined by the feature classification AI model. If the accuracy rate of the determination result is equal to or higher than a predetermined accuracy rate, the machine learning-developed AI model for judgment replaces the original AI model for judgment, thereby automatically updating the AI model for judgment. If the accuracy rate of the judgment result is less than the predetermined accuracy rate, the AI model for feature classification will be automatically updated by adding new result data from a predetermined number of years prior to the present to the result data from a predetermined number of years prior to the present, or by replacing the result data from a predetermined number of years prior to the present with new result data from a predetermined number of years prior to the present. A method for updating an AI model for classifying and determining the causes of defects, including the following.
2. The product is a printed circuit board, and the method for updating the AI model for classifying and determining the causes of defects according to claim 1.
3. A method for updating an AI model for classifying defect causes according to claim 1, wherein the proportion of data for each classification of a predetermined proportion of the clear data is adjusted according to the proportion of data for each classification determined by the AI model for feature classification.
4. A method for updating an AI model for classifying and determining the causes of defects according to claim 1, wherein when updating the result data from a predetermined number of years prior to the present, a process is performed to delete the result data from the result data from a predetermined number of years prior to the present in chronological order.
5. A database that may include multiple types of information from inspections or measurements related to the product, and which has accumulated multiple result data over time, in which at least one type of information indicates a product defect, A feature classification AI model machine learning unit extracts result data from a predetermined number of years ago to the present from among the multiple result data in the aforementioned database and inputs it into a feature classification AI model, and uses unsupervised learning to machine-learn the number of classifications and the classification method for determining the classification of the result data from the predetermined number of years ago to the present based on its features, The aforementioned machine-learned AI model for feature classification broadly classifies the result data from a predetermined number of years ago to the present into clear data, where the features are clear and easy for the AI model to classify based on features, and gray data, where the features are unclear and difficult for the AI model to classify based on features. A predetermined percentage of the clear data is used as training data, and the remainder of the clear data is used as verification data. The training data acquisition unit determines the classification of each result data in the training data and the verification data based on features. A machine learning unit for a judgment AI model inputs the aforementioned training data into a judgment AI model and causes the judgment AI model to perform unsupervised learning to classify the aforementioned training data based on the number of classifications and the judgment method, using its features. A judgment AI model update unit automatically updates the judgment AI model by using the machine learning judgment AI model to determine the classification of the verification data excluding the classification information determined by the feature classification AI model, comparing the determination result with the classification information, and replacing the original judgment AI model with the machine learning judgment AI model if the accuracy rate is above a predetermined accuracy rate. A result data update unit automatically updates the result data input to the AI model for feature classification by adding new result data from a predetermined number of years prior to the present, or by replacing the result data from a predetermined number of years prior to the present with new result data from a predetermined number of years prior to the present, when the accuracy rate of the judgment result is less than the predetermined accuracy rate. An AI model update system for classifying and determining the causes of defects, comprising the following features.
6. The product is a printed circuit board, and the AI model update system for classifying and determining the causes of defects according to claim 5.
7. The AI model update system for classifying defect occurrence factors according to claim 5, wherein the ratio of the number of data for each classification of a predetermined proportion of the clear data is adjusted according to the ratio of the number of data for each classification by the AI model for feature classification.
8. The AI model update system for classifying and determining the causes of defects according to claim 5, which, when updating the result data from a predetermined number of years prior to the present, performs a process of deleting the result data from the result data from a predetermined number of years prior to the present in chronological order.