A machine learning-based device-level bad root cause localization method and system
By employing a machine learning-based device-level defect root cause localization method, utilizing one-hot encoding and feature selection rules, and combining multiple basic model fusion rules, the root causes of defects in display panel production can be quickly and accurately located. This solves the problems of long processing time, high cost, and low automation in existing technologies, achieving efficient and accurate root cause analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 合肥欣奕华智能机器股份有限公司
- Filing Date
- 2023-05-15
- Publication Date
- 2026-06-23
AI Technical Summary
In display panel production, existing technologies struggle to quickly and accurately pinpoint the root cause of defects. Traditional methods are time-consuming, costly, inefficient, and lack automation.
A machine learning-based device-level defect root cause localization method is adopted. By acquiring defect history information and defect rate information, performing one-hot encoding and feature screening, and combining the fusion rules of multiple basic models, the device defect grouping is quickly output, and the root cause of the anomaly is automatically located.
It improves the speed and accuracy of root cause localization, reduces the running time of learning algorithms, simplifies manual operations, and increases the degree of automation and return on investment.
Smart Images
Figure CN116738275B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of production control technology, and in particular to a machine learning-based method and system for locating the root causes of equipment-level defects. Background Technology
[0002] The display panel industry has high technological barriers, with complex processes, numerous key technologies in production procedures, and high difficulty, resulting in a long industrial chain. The display panel production process involves hundreds of stations, each containing many pieces of equipment, making it difficult to pinpoint the root cause when a product malfunctions.
[0003] Currently, the traditional method for locating root causes relies on manual inspection of defects. Manually locating defects requires relying on experience to dig out the root causes from thousands of devices. This method suffers from problems such as long analysis time, high cost, low inspection efficiency, and high requirements for personnel's professional skills.
[0004] For panel defects, existing technologies employ two analytical methods. The first uses the parameters of the equipment at each station as analytical factors, referred to as the equipment parameter analysis method. The second uses the parallel equipment through which the defective product flows at different process stations as analytical factors, referred to as the parallel equipment analysis method. The equipment parameter analysis method has the following shortcomings:
[0005] 1. Large data volume: A panel process typically involves thousands of devices, each with numerous parameters that are updated rapidly, resulting in a long time consumption in actual analysis.
[0006] 2. Equipment parameter analysis methods involve extracting defective data from the parameters of numerous devices in a factory. Currently, the industry generally believes that extracting equipment parameters is not very accurate and is difficult.
[0007] 3. Compared to other equipment-level analysis methods, the return on investment for equipment parameter-level analysis is low. In the actual analysis of defective panels, equipment parameter analysis increases the investment and is slow. Furthermore, once the problem is identified at the equipment level, most issues are already resolved, allowing process engineers to quickly pinpoint the defective parameters from the faulty equipment.
[0008] Parallel device analysis refers to the correlation analysis of defective data corresponding to different combinations of parallel devices flowing through a site, which improves accuracy compared to device parameter-level analysis. However, parallel device analysis also has the following drawbacks:
[0009] 1. Parallel equipment includes multiple devices. After the root cause is identified using parallel equipment analysis methods, process engineers still need to sequentially check which parameters of the specific devices in the parallel equipment are causing the malfunction, resulting in a low degree of automation.
[0010] 2. When a specific device malfunctions, the parallel device analysis method treats the parallel devices of the root cause device as malfunction data for correlation analysis, which weakens the contribution of the root cause to the results, resulting in lower detection accuracy. Summary of the Invention
[0011] Based on the technical problems existing in the background technology, this invention proposes a machine learning-based method and system for locating the root causes of equipment failures, which improves the speed and accuracy of root cause discovery.
[0012] This invention proposes a machine learning-based method for locating the root causes of device-level defects, comprising the following steps:
[0013] S1: Obtain the defect history information and defect rate information of the production panel, and fill the newly created empty list of site equipment columns based on the defect history information to obtain the site equipment column of the batch product. The defect history information includes the site information and equipment history information through which the production panel flows.
[0014] S2: Filter the list of station equipment and send the filtered list of station equipment to the trained fusion model to output the group of equipment failures, and determine the root cause of the anomaly based on the group of equipment failures.
[0015] Furthermore, the history information includes information on different stations the production panel process passes through, equipment information at each station, production time, and sample ID.
[0016] Furthermore, in step S1, after obtaining the defect history information and defect rate information of the production panel, the defect history information is preprocessed, specifically by deleting the rows containing outliers in the defect history information.
[0017] Furthermore, in step S1, after obtaining the station equipment list of the batch products, the station equipment list is subjected to one-hot encoding processing to convert the category information of the equipment information flowing through the station equipment list into numerical information.
[0018] Furthermore, in step S2, the specific formula for filtering the station equipment column is as follows:
[0019]
[0020] Among them, V i F represents the variance of the i-th feature column. i f represents the F-value of the i-th feature. i This indicates the score under the filtering rules.
[0021] Furthermore, in step S2, the fusion model includes several basic models, and the fusion process of the fusion model for the input site device list is as follows:
[0022] Using the filtered site equipment column as input to the basic model and the defect rate information as the label of the basic model, the equipment defect rate is predicted, and the accuracy of each basic model is given based on the prediction results.
[0023] Calculate the feature importance of each base model and rank the feature importance.
[0024] A fusion rule is set for the importance of the ranked features to obtain the root cause importance ranking. The root cause importance ranking is normalized to obtain the feature importance score. The feature importance scores are sorted from high to low and output to obtain the equipment defect group.
[0025] Furthermore, the specific formula for the fusion rule is as follows:
[0026]
[0027] Among them, R j Acc represents the score of the j-th feature after fusion. i R represents the accuracy score of the i-th base model, Acc represents the sum of the accuracy scores of all base models, and Ri represents the accuracy score of the i-th base model. (i,j) This represents the order of the i-th basic model and the j-th feature.
[0028] A machine learning-based device-level root cause localization system includes an acquisition and construction module and a screening and fusion module;
[0029] The acquisition and construction module is used to acquire the defect history information and defect rate information of the production panel, and fill the newly created empty list of site equipment columns based on the defect history information to obtain the site equipment column of the batch product. The defect history information includes the site information and equipment history information through which the production panel flows.
[0030] The filtering and fusion module is used to filter the list of station equipment and send the filtered list of station equipment to the trained fusion model to output the group of equipment failures, and determine the root cause of the anomaly based on the group of equipment failures.
[0031] Furthermore, the history information includes information on different stations the production panel process passes through, equipment information at each station, production time, and sample ID;
[0032] The specific formula for filtering the station equipment column is as follows:
[0033]
[0034] Among them, V i F represents the variance of the i-th feature column. i f represents the F-value of the i-th feature. iThis indicates the score under the filtering rules.
[0035] The advantages of the machine learning-based device-level defect root cause localization method and system provided by this invention are as follows: This invention analyzes all different devices flowing through different sites as defect data, processes this data to construct single-device statistical features, and combines machine learning knowledge to design feature selection rules, significantly reducing the running time of the learning algorithm. It also designs algorithm model fusion rules to integrate the advantages of various basic models, achieving a stronger ability to solve unknown problems. Compared with device parameter-level analysis methods and device-level analysis methods, it can more quickly locate specific defective devices, improving the speed and accuracy of root cause discovery. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation
[0037] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0038] like Figure 1 As shown, the present invention proposes a machine learning-based method for locating the root causes of device-level defects, comprising the following steps:
[0039] S1: Obtain the defect history information and defect rate information of the production panel, and fill the newly created empty list of site equipment columns based on the defect history information to obtain the site equipment column of the batch product. The defect history information includes the site information and equipment history information through which the production panel flows.
[0040] The history information includes information about the different stations the panel production process passes through, the equipment information under each station, the production time, and the sample ID. Specifically, the station equipment column refers to the fact that panel production involves multiple stations, and each station has multiple pieces of equipment. The station equipment column is obtained by linking each station with its various equipment.
[0041] Specifically, the defect rate information is the ratio of the number of defects in a product's finer-grained structure observed by factory workers using display instruments to the total number of defects (in a specific process, a panel is divided into many small areas at the PNL level, and the defect rate is calculated as the number of defective PNL panels out of all PNL panels). When a specific defect occurs on a panel, the time period in which the defect is concentrated is determined, and the station, equipment, product ID, and production time that the panel passed through during this time period are extracted as historical information. The historical information is in a large table format, which lists product ID, production time, station, and equipment information, recording the information passed through by all product IDs.
[0042] Since the defective data is a large table that requires storage in a large database, a large data storage database, such as Hadoop or HBase, is first established. Then, the factory's data is extracted into the established database using an ETL big data component. In this embodiment, the defect root cause localization method (data processing and algorithm modeling) is completed in the established database. This approach will not affect the factory's actual production.
[0043] It should be noted that actual factory data may contain missing or incorrect records. The panel production process is complex and diverse, and the defect history information constructed from actual factory data will inevitably contain missing or outlier values. This embodiment preprocesses the defect history information by deleting the rows containing outlier values. This avoids the defects caused by performing other operations on missing or outlier values, which could lead to a decrease in accuracy or discrepancies between the defect history information and the actual factory data, resulting in deviations in the identified root causes of the anomalies.
[0044] Understandably, the newly created site equipment column extracts the history information corresponding to a certain defect category from the database to record the site and equipment information through which the product ID flows. The newly created site equipment column narrows the scope of root cause analysis. One-hot encoding is used to process the specific equipment column in the site equipment column. One-hot encoding expands the number of features in the data in the site equipment column, transforming the category information of the flowing equipment information into numerical information, which is convenient for subsequent filtering and provides feature data for the fusion model in step S2.
[0045] S2: Filter the list of station equipment and send the filtered list of station equipment to the trained fusion model to output the group of equipment failures, and determine the root cause of the anomaly based on the group of equipment failures.
[0046] Set feature filtering rules to remove irrelevant and redundant features. The specific feature filtering rule formula is as follows:
[0047]
[0048] Among them, Vi F represents the variance of the i-th feature column. i f represents the F-value of the i-th feature. i This represents the score under the filtering rules, based on f. i Irrelevant features with values of 0 or close to 0 are removed.
[0049] This embodiment uses variance and F-test to filter irrelevant features, but is not limited to other feature screening methods.
[0050] Specifically, the fusion model includes several basic models, and the fusion process of the input site device list is as follows (a) to (c):
[0051] (a) Using the filtered site equipment column as the input of the basic model and the defect rate information as the label of the basic model, the equipment defect rate is predicted, and the accuracy of each basic model is given based on the prediction results.
[0052] Accuracy is a metric used to evaluate classification models; it is the proportion of correct predictions made by the basic model out of the total number of predictions.
[0053] The defect rate information is used as a threshold setting. When the equipment defect rate predicted by a certain basic model is greater than the defect rate information, the output of the basic model is 1. When the equipment defect rate predicted by a certain basic model is less than or equal to the defect rate information, the output of the basic model is 0.
[0054] (b) Calculate the feature importance of each base model and rank the feature importance;
[0055] Feature importance refers to the degree of influence of a feature on the target variable, that is, the importance of the feature in the basic model. It can quickly screen out features that have a greater impact on model performance and reduce computational costs. Feature importance is a common feature of the basic model.
[0056] (c) Set fusion rules for the ranked feature importance to obtain the root cause importance ranking, normalize the root cause importance ranking to obtain the feature importance score, and output the feature importance score from high to low to obtain the equipment defect group.
[0057] The higher the equipment ranking, the greater the likelihood that the equipment is defective.
[0058] The specific formula for the fusion rule is as follows:
[0059]
[0060] Among them, R j Acc represents the score of the j-th feature after fusion. iR represents the accuracy of the i-th base model, Acc represents the sum of the accuracy scores of all base models, and Ri represents the accuracy of the i-th base model. (i,j) This represents the order of the i-th basic model and the j-th feature.
[0061] Based on steps S1 to S2, by analyzing all different devices flowing through different sites as bad data, these data are processed to construct single-device statistical features. Combining machine learning knowledge, feature selection rules are designed, which greatly reduces the running time of the learning algorithm. Algorithm model fusion rules are designed to integrate the advantages of various basic models and obtain a stronger ability to solve unknown problems. The speed and accuracy of root cause mining are improved by comparing device parameter-level analysis methods and device-level analysis methods.
[0062] Therefore, compared to root cause analysis methods based on equipment parameters, the technical solution in this embodiment offers higher accuracy and speed. Compared to parallel equipment analysis methods, the technical solution in this embodiment saves manual judgment on which equipment the root cause of the defect occurs in, resulting in a higher degree of automation. Consequently, the technical solution in this embodiment can locate the specific defective equipment more quickly than industry solutions, offering higher accuracy and speed.
[0063] It should be noted that the columns in the "Site Equipment" column represent site equipment information, and the rows represent the ID number of each record information. The empty list of "Site Equipment" columns is populated by iterating through the record information to obtain the input data for training the fusion model. This input data serves as the training set, and the data format used in the fusion model process is consistent with the input data format used in its training. During the training of the fusion model, the training set needs to be processed in step S1 and then filtered before being fed into the already constructed fusion model for training. The fusion model is trained based on the existing objective function to achieve convergence.
[0064] In summary, the device-level root cause localization method and system based on machine learning has the following advantages:
[0065] (1) Compared with the traditional method of relying on manual search for the root cause of the problem, the technical solution disclosed in this embodiment greatly improves work efficiency and saves costs. Compared with the existing equipment root cause solutions, this patent has higher accuracy and greater diversity.
[0066] (2) Existing patent CN202011336989.5 discloses a method, system, device and medium for locating the root cause of equipment parameters, and CN202011088294.X discloses a method for analyzing the root cause of yield loss based on information fusion. Both patents are equipment analysis methods. When performing defect analysis, the analysis is time-consuming, the accuracy is low and the input-output ratio is low. Compared with the above existing patents, the technical solution of this embodiment has the advantages of short analysis time, high accuracy and high input-output ratio.
[0067] (3) Existing patent CN202111070227.X provides a root cause analysis method, device, electronic device and medium. This patent discloses a method for analyzing defects of parallel devices, which performs correlation analysis on defect data corresponding to different combinations of devices flowing through a station. However, when performing defect analysis on a panel, the degree of automation is not high and many processes require manual operation. The technical solution of this embodiment analyzes all devices flowing through different stations as defect data, integrates machine learning algorithms to model defect data and mine defect root causes, simplifies manual operation, improves the speed of defect root cause mining, and thus has a high input-output ratio.
[0068] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A machine learning-based method for locating the root causes of device-level defects, comprising the following steps: S1: Obtain the defect history information and defect rate information of the production panel, and fill the newly created empty list of site equipment columns based on the defect history information to obtain the site equipment column of the batch product. The defect history information includes the site information and equipment history information through which the production panel flows. S2: Filter the list of station equipment and send the filtered list of station equipment to the trained fusion model to output the group of equipment failures, and determine the root cause of the anomaly based on the group of equipment failures; wherein The fusion model includes several basic models, and the fusion process of the input site device column is as follows: Using the filtered site equipment column as input to the basic model and the defect rate information as the label of the basic model, the equipment defect rate is predicted, and the accuracy of each basic model is given based on the prediction results. Calculate the feature importance of each base model and rank the feature importance. Set fusion rules for the ranked feature importance to obtain the root cause importance ranking. Normalize the root cause importance ranking to obtain the feature importance score. Sort the feature importance scores from high to low and output the equipment defect group. The specific formula for the fusion rule is as follows: wherein, denotes the score of the th feature, denotes the accuracy of the th base model, denotes the sum of the accuracy scores of all base models, denotes the ranking of the th base model, the th feature. 2.The machine learning based device level bad root cause localization method of claim 1, wherein, The resume information includes information on different stations the production panel process passes through, equipment information at each station, production time, and sample ID. 3.The machine learning based device level bad root cause localization method of claim 1, wherein, In step S1, after obtaining the defect history information and defect rate information of the production panel, the defect history information is preprocessed, specifically by deleting the rows containing outliers in the defect history information. 4.The machine learning based device level bad root cause localization method of claim 1, wherein, In step S1, after obtaining the station equipment list of the batch products, the station equipment list is subjected to one-hot encoding processing to convert the category information of the equipment information flowing through the station equipment list into numerical information. 5.The machine learning based device level bad root cause localization method of claim 1, wherein, In step S2, the specific formula for filtering the station equipment column is as follows: in, Indicates the first The variance of the column features, Indicates the first F-value of column features This indicates the score under the filtering rules.
6. A machine learning-based device-level root cause localization system, characterized in that, This includes acquiring the building module and filtering / merging the module; The acquisition and construction module is used to acquire the defect history information and defect rate information of the production panel, and fill the newly created empty list of site equipment columns based on the defect history information to obtain the site equipment column of the batch product. The defect history information includes the site information and equipment history information through which the production panel flows. The filtering and fusion module is used to filter the station equipment list and send the filtered station equipment list to the trained fusion model to output the equipment failure grouping, and determine the root cause of the anomaly based on the equipment failure grouping; The fusion model includes several basic models, and the fusion process of the input site device list is as follows: Using the filtered site equipment column as input to the basic model and the defect rate information as the label of the basic model, the equipment defect rate is predicted, and the accuracy of each basic model is given based on the prediction results. Calculate the feature importance of each base model and rank the feature importance. Set fusion rules for the ranked feature importance to obtain the root cause importance ranking. Normalize the root cause importance ranking to obtain the feature importance score. Sort the feature importance scores from high to low and output the equipment defect group. The specific formula for the fusion rule is as follows: in, Indicates the fusion of the first The score of each feature, Indicates the first The accuracy of the basic model, This represents the sum of the accuracy scores of all basic models. Indicates the first The first basic model, the first The ranking of features.
7. The machine learning-based device-level root cause localization system according to claim 6, characterized in that, The history information includes information on different stations the production panel process passes through, equipment information at each station, production time, and sample ID; The specific formula for filtering the station equipment column is as follows: in, Indicates the first The variance of the column features, Indicates the first F-value of column features This indicates the score under the filtering rules.