Manufacturing process optimization methods based on defect pattern mining

Abstract

This invention discloses a manufacturing process optimization method based on defect pattern mining, relating to the field of data processing. The method includes: obtaining defect indicator detection data; performing a first-level de-randomization analysis based on the detected defect indicator type data to obtain high-frequency detected defect indicator types; performing a second-level de-randomization analysis on the defect indicator deviation vector data to obtain high-frequency defect indicator deviation vectors; obtaining production control attributes from the first node to the Nth node; performing correlation analysis to obtain correlation control attributes; matching control variables on the high-frequency defect indicator deviation vectors to obtain process calibration attributes; and sending the results to a quality control terminal. This method solves the technical problem of existing production process optimization methods lacking in-depth analysis of quality-related causes, thus failing to fundamentally optimize the production process. It achieves the technical effect of precise control and optimization of the production process by uncovering the root causes of product quality problems.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a manufacturing process optimization method based on defect pattern mining. Background Technology

[0002] Intelligent manufacturing has become an important direction for the transformation and upgrading of the manufacturing industry. In the production of complex products, due to long production lines, complex processes, diverse raw materials, and fluctuations in environmental factors, product quality control becomes particularly critical and challenging. Traditional quality management models often rely on post-production inspection and feedback; that is, after product production is completed, quality problems are discovered through sampling or full inspection, and then defective products are rejected or reworked. This cannot fundamentally optimize the production process or prevent similar quality problems from recurring. Because the production of complex products involves multiple stages and multiple factors, the same quality inspection results may be caused by a variety of different production factors, making the tracing of the source of quality problems extremely difficult.

[0003] At present, the relevant technologies for production process optimization suffer from a lack of in-depth exploration and analysis of the underlying quality factors, thus failing to fundamentally optimize the production process. Summary of the Invention

[0004] This application provides a manufacturing process optimization method based on defect pattern mining. When the product quality pass rate falls below a preset threshold, it automatically acquires defect indicator detection data of abnormal products. Two-level randomness analysis is performed on the defect indicators to screen out high-frequency detected defect indicator types and their deviation vectors. Correlation analysis is conducted between these high-frequency detected defect indicators and various stages of the production process to determine key control attributes affecting product quality. Based on big data analysis, control variables are matched to the deviation vectors of high-frequency defect indicators to obtain process calibration attributes, i.e., optimized production process parameters. These optimized process parameters are then promptly fed back to the quality control terminal for real-time adjustment and optimization of the production process. These technical means accurately identify and eliminate potential quality problems in the production process, achieving the technical effect of precise control and optimization of the production process by uncovering the root causes of product quality issues.

[0005] This application provides a manufacturing process optimization method based on defect pattern mining, comprising: when the product quality pass rate is less than or equal to a quality pass rate threshold, obtaining defect index type data and defect index deviation vector data of products with quality abnormalities; performing a first-level randomness analysis based on the defect index type data to obtain high-frequency defect index types; performing a second-level randomness analysis based on the high-frequency defect index types to obtain high-frequency defect index deviation vectors; performing correlation analysis on production control attributes from the first node to the Nth node based on the high-frequency defect index types to obtain correlated control attributes, wherein the correlated control attributes have node sequence number labels; traversing the correlated control attributes based on the node sequence number labels, performing control variable matching on the high-frequency defect index deviation vectors based on big data to obtain process calibration attributes; and sending the node sequence number labels and the process calibration attributes to a quality control terminal.

[0006] In a possible implementation, a first-level randomness analysis is performed based on the detected defect index type data to obtain high-frequency detected defect index types, and the following processing is performed: based on the detected defect index type data, the number of products detected by the first defect index type is counted up to the number of products detected by the Mth defect index type; the product quantity threshold is received from the interactive quality control terminal; from the number of products detected by the first defect index type up to the number of products detected by the Mth defect index type, defect index types that are greater than or equal to the product quantity threshold are selected and added to the high-frequency detected defect index types.

[0007] In a possible implementation, a second-level randomness analysis is performed on the defect indicator deviation vector data based on the high-frequency detected defect indicator type to obtain a high-frequency defect indicator deviation vector. The following processing is then performed: obtaining the defect indicator deviation vector data of the first high-frequency detected defect indicator type; performing cluster analysis on the defect indicator deviation vector data to obtain a defect indicator deviation vector clustering result, wherein the defect indicator deviation vector clustering result has an intra-cluster vector quantity label; interacting with a quality control terminal to receive an intra-cluster vector quantity threshold; and selecting clusters greater than or equal to the intra-cluster vector quantity threshold from the defect indicator deviation vector clustering result based on the intra-cluster vector quantity label to obtain the high-frequency defect indicator deviation vector.

[0008] In a possible implementation, cluster analysis is performed on the defect indicator deviation vector data to obtain the defect indicator deviation vector clustering result. The following processing is then performed: First defect indicator deviation vector and second defect indicator deviation vector of the defect indicator deviation vector data of the first high-frequency detected defect indicator type are obtained; the cosine distance between the first defect indicator deviation vector and the second defect indicator deviation vector is calculated to generate a first clustering parameter; the Chebyshev distance between the first defect indicator deviation vector and the second defect indicator deviation vector is calculated to generate a second clustering parameter; when the first clustering parameter is less than or equal to the first clustering parameter threshold, and the second clustering parameter is less than or equal to the second clustering parameter threshold, the first defect indicator deviation vector and the second defect indicator deviation vector are clustered into one class; when the first clustering parameter is greater than the first clustering parameter threshold, and / or the second clustering parameter is greater than the second clustering parameter threshold, the first defect indicator deviation vector and the second defect indicator deviation vector are clustered into two classes; when the defect indicator deviation vector data traversal is complete, the defect indicator deviation vector clustering result is output.

[0009] In a possible implementation, a correlation analysis is performed on the production control attributes of the first node up to the Nth node based on the high-frequency detected defect index type to obtain the associated control attribute, and the following processing is performed: Based on the high-frequency detected defect index type, a first high-frequency detected defect index type is extracted; based on the first high-frequency detected defect index type, a joint negative sample retrieval of k control attributes is performed on the production control attributes of the first node up to the Nth node to obtain the abnormal frequency ratio of the defect indexes of the k control attributes, where the total number of control attributes is ≥ k ≥ 1, and the initial value of k is equal to 1; when the abnormal frequency ratio of the defect indexes of the k control attributes is greater than or equal to the abnormal frequency ratio threshold, the k control attributes are jointly stored and added to the associated control attribute; when k equals the total number of control attributes, the associated control attribute is output.

[0010] In a possible implementation, the following processing is performed: when k is less than the total number of control attributes, k is incremented by one; the k control attributes are deleted from the first node production control attribute up to the Nth node production control attribute to obtain the production control attribute update result; based on the high-frequency detected defect index type, the production control attribute update result is analyzed by adding one k to obtain the associated control attribute.

[0011] In a possible implementation, based on the node sequence label, the associated control attributes are traversed, and control variable matching is performed on the high-frequency defect index deviation vector based on big data to obtain process calibration attributes. The following processing is then performed: Based on the associated control attributes, k control attributes of the first high-frequency detected defect index type are extracted, where the total number of control attributes is ≥k≥1; based on the k control attributes, the time series information of the monitoring feature values ​​of the k control attributes and the time series information of the baseline feature values ​​of the k control attributes are extracted; through the time series information of the monitoring feature values ​​of the k control attributes and the time series information of the baseline feature values ​​of the k control attributes, anomaly analysis is performed on the k control attributes to obtain… The joint anomaly coefficient of k control attributes; when the joint anomaly coefficient of the k control attributes is greater than or equal to the joint anomaly coefficient threshold, the attribute deviation vector matrix of the joint anomaly coefficient of the k control attributes is used as the unique anomaly, and historical production backtracking is performed based on big data to obtain historical defect detection index types and historical defect detection index vectors; when the historical defect detection index type is consistent with the first high-frequency defect detection index type, and any one of the historical defect detection index vector and the high-frequency defect index deviation vector can be clustered into one class, then the k control attributes are added to the process calibration attribute of the first high-frequency defect detection index type, and added to the process calibration attribute.

[0012] In a possible implementation, anomaly analysis is performed on the k control attributes by comparing the time series information of the monitored feature values ​​of the k control attributes with the time series information of the baseline feature values ​​of the k control attributes, to obtain the joint anomaly coefficient of the k control attributes, and the following processing is performed: the proportion of the number of time series information of the monitored feature values ​​of the k control attributes that does not conform to the time series information of the baseline feature values ​​of the k control attributes is counted, and the joint anomaly coefficient of the k control attributes is obtained.

[0013] The proposed manufacturing process optimization method based on defect pattern mining involves obtaining data on detected defect index types and defect index deviation vectors for products with quality abnormalities when the product quality pass rate is less than or equal to a quality pass rate threshold. Then, a first-level randomness analysis is performed based on the detected defect index type data to obtain high-frequency detected defect index types. Next, a second-level randomness analysis is performed on the defect index deviation vector data based on the high-frequency detected defect index types to obtain high-frequency defect index deviation vectors. Then, a correlation analysis is performed on the production control attributes from the first node to the Nth node based on the high-frequency detected defect index types to obtain correlated control attributes. These correlated control attributes have node sequence labels. Based on the node sequence labels, the correlated control attributes are traversed, and control variables are matched against the high-frequency defect index deviation vectors using big data to obtain process calibration attributes. Finally, the node sequence labels and process calibration attributes are sent to the quality control terminal. This achieves the technical effect of precise control and optimization of the production process by uncovering the root causes of product quality problems. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the method according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0015] Figure 1 This is a flowchart illustrating the manufacturing process optimization method based on defect pattern mining provided in an embodiment of this application.

[0016] Figure 2 This is a flowchart illustrating the process of obtaining the clustering results of the defect index deviation vector in the manufacturing process optimization method based on defect pattern mining provided in this application embodiment. Detailed Implementation

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0020] This application provides a manufacturing process optimization method based on defect pattern mining, such as... Figure 1 As shown, the method includes:

[0021] Step S100: When the product quality pass rate is less than or equal to the quality pass rate threshold, obtain the defect index type data and defect index deviation vector data of the products with quality abnormalities. Specifically, during the production process, product quality inspection data is collected in real time or periodically. This data includes, but is not limited to, the number of qualified products, the number of unqualified products, and the total production quantity. Calculate the product quality pass rate for the current batch or time period, i.e., the number of qualified products divided by the total production quantity. A quality pass rate threshold is preset, which is determined based on product standards, customer requirements, or the company's internal quality control standards. Compare the calculated product quality pass rate with the preset threshold. If the product quality pass rate is less than or equal to the preset threshold, it indicates that there is a problem with the product quality in the current production batch or time period, requiring further analysis. At this time, the data collection mechanism is triggered to obtain detailed defect index detection data of the products with quality abnormalities from the quality inspection system. Among them, the defect index detection data is product defect information discovered during the quality inspection process through specific means (such as machine vision, sensors, etc.), including two parts: defect index type data and defect index deviation vector data. The detected defect type data refers to the specific defect type detected, such as size deviation, color mismatch, cracks, etc. The defect index deviation vector data is the result of quantifying or encoding these defects, describing the specific degree and characteristics of the defects in vector form. It includes multiple dimensions describing defect characteristics (such as size, color, shape, etc.), and the value on each dimension represents the deviation between the actual measured value and the standard value, which is used for subsequent data analysis and processing.

[0022] Step S200: Perform a first-level randomness removal analysis on the detected defect index type data to obtain high-frequency detected defect index types. Specifically, the detected defect index type data is organized to ensure its accuracy and completeness. The data is cleaned to remove duplicates, outliers, or invalid data to improve the accuracy of the analysis. Frequency statistics are performed on the organized detected defect index type data using tools such as database queries, statistical functions in programming languages, or data analysis software to calculate the frequency of each defect type. Based on the frequency statistics results, defect index types with a significantly higher frequency than a frequency threshold (e.g., the average level) are identified; these are the high-frequency detected defect index types. Low-frequency defect types caused by random factors (such as operational errors, temporary equipment malfunctions, etc.) are excluded, thus focusing on high-frequency defects caused by systemic problems in the production process. The results of the randomness removal analysis are output as a list of high-frequency detected defect index types.

[0023] In one possible implementation, a first-level randomness analysis is performed based on the detected defect indicator type data to obtain high-frequency detected defect indicator types. Step S200 further includes step S210, which involves counting the number of products detected for the first defect indicator type up to the number of products detected for the Mth defect indicator type based on the detected defect indicator type data. Specifically, the obtained detected defect indicator type data is traversed, and for each defect indicator type, the corresponding number of detected products is counted, i.e., how many products are detected with that type of defect. Step S220 involves interacting with the quality control terminal and receiving a product quantity threshold. Specifically, the system interacts with the quality control terminal, which is a system or device responsible for product quality monitoring and management, through a user interface, API interface, or other communication methods, and receives the product quantity threshold provided by the quality control terminal. This threshold is used to determine whether a certain defect indicator type is frequent enough to be considered a high-frequency defect. Step S230 involves sorting out defect indicator types that are greater than or equal to the product quantity threshold from the number of products detected for the first defect indicator type up to the number of products detected for the Mth defect indicator type, and adding them to the high-frequency detected defect indicator types. Specifically, the number of products detected for each defect indicator type, as counted in step S210, is iterated. The number of products detected for each defect indicator type is compared with the product quantity threshold received in step S220. If the number of products detected for a certain defect indicator type is greater than or equal to the product quantity threshold, it is considered a high-frequency defect indicator type and added to a list. Ultimately, this list will contain all detected high-frequency defect indicator types. This implementation method, by counting the number of products detected for each defect indicator type and comparing it with the product quantity threshold, eliminates accidental defects that only appear in a few products, thereby achieving the technical effect of improving the accuracy of high-frequency defect identification.

[0024] Step S300: Based on the high-frequency detected defect index type, perform a second-level de-randomization analysis on the defect index deviation vector data to obtain high-frequency defect index deviation vectors. Specifically, from the obtained defect index deviation vector data, select a data subset corresponding to the high-frequency detected defect index type. This data subset contains all deviation vector information related to high-frequency defects. Quantitatively evaluate the selected defect index deviation vector data across various dimensions (or "features") to determine which dimensions have the most significant deviations. Based on the deviation vector analysis, perform a second-level de-randomization analysis by setting deviation thresholds, conducting statistical tests (such as t-tests, chi-square tests, etc.), or using machine learning algorithms. This excludes deviation vectors caused by random factors (such as temporary equipment failures, operational errors, etc.) and focuses on high-frequency defect index deviation vectors that frequently occur in quality inspection and may be caused by systemic problems. Based on the results of the second-level de-randomization analysis, determine the high-frequency defect index deviation vectors. These vectors represent the specific manifestations of defects that frequently occur in the production process and may be caused by systemic problems.

[0025] In one possible implementation, a second-level randomness analysis is performed on the defect indicator deviation vector data based on the high-frequency detected defect indicator type to obtain the high-frequency defect indicator deviation vector. Step S300 further includes step S310, obtaining the defect indicator deviation vector data of the first high-frequency detected defect indicator type of the high-frequency detected defect indicator type. Specifically, from the obtained high-frequency detected defect indicator types, any high-frequency detected defect indicator type is selected as the first high-frequency detected defect indicator type, and all defect indicator deviation vector data related to this defect indicator type are extracted. Step S320, cluster analysis is performed on the defect indicator deviation vector data to obtain the defect indicator deviation vector clustering result, wherein the defect indicator deviation vector clustering result has an in-cluster vector count label. Specifically, a clustering algorithm (such as K-means, DBSCAN, etc.) is applied to group the defect indicator deviation vector data obtained in step S310, aggregating similar deviation vectors together to form different clusters, each cluster being identified by the in-cluster vector count (i.e., the number of vectors in the cluster). Step S330, the quality control terminal is interacted with to receive the in-cluster vector count threshold. Specifically, the system interacts with the quality control terminal via a user interface or API, receiving a threshold for the number of in-class vectors input by the user. This threshold is used to determine which clusters are significant (i.e., contain a sufficient number of deviation vectors, potentially caused by systemic problems). Step S340 involves selecting clusters with a number of in-class vectors greater than or equal to the threshold from the defect index deviation vector clustering results, obtaining the high-frequency defect index deviation vectors. Specifically, each cluster in the defect index deviation vector clustering results is traversed, its in-class vector number label is checked, and clusters with an in-class vector number greater than or equal to the threshold are selected. The defect index deviation vectors in these clusters are considered high-frequency defect index deviation vectors because they represent frequently occurring defects that may be caused by systemic problems. This implementation grouped similar deviation vectors together and set an in-class vector number threshold, focusing only on clusters containing a sufficient number of deviation vectors. This excludes isolated deviation vectors caused by random factors. Through cluster analysis, it is easier to identify defect patterns that may be caused by systemic problems, achieving the technical effect of improving the accuracy and effectiveness of the secondary randomness analysis.

[0026] like Figure 2As shown, in one possible implementation, cluster analysis is performed on the defect index deviation vector data to obtain the defect index deviation vector clustering result. Step S320 further includes step S321, obtaining the first defect index deviation vector and the second defect index deviation vector of the defect index deviation vector data of the first high-frequency detected defect index type. Specifically, a first high-frequency detected defect index type is selected from the high-frequency detected defect index types, and two defect index deviation vectors are randomly selected or selected in a specific order (such as time order, sample number, etc.) from the defect index deviation vector data of this type as the first defect index deviation vector and the second defect index deviation vector. Step S322, the cosine distance between the first defect index deviation vector and the second defect index deviation vector is calculated to generate the first clustering parameter. Specifically, the cosine similarity formula is used to calculate the cosine distance between the two vectors (the higher the cosine similarity, the smaller the cosine distance, indicating that the two vectors are more similar), and the calculated cosine distance is used as the first clustering parameter to measure the similarity of the two vectors in direction. Step S323: Calculate the Chebyshev distance between the first defect index deviation vector and the second defect index deviation vector to generate a second clustering parameter. Specifically, the maximum coordinate difference between the two vectors is calculated using the Chebyshev distance formula, and the calculated Chebyshev distance is used as the second clustering parameter to measure the maximum deviation of the two vectors in each dimension.

[0027] Step S324: When the first clustering parameter is less than or equal to the first clustering parameter threshold, and the second clustering parameter is less than or equal to the second clustering parameter threshold, the first defect index deviation vector and the second defect index deviation vector are clustered into one class. Specifically, it is checked whether the first clustering parameter and the second clustering parameter both satisfy their respective threshold conditions. If both are satisfied, it is considered that the deviations of these two vectors in direction and in each dimension are sufficiently small, and they can be regarded as the same class. Step S325: When the first clustering parameter is greater than the first clustering parameter threshold, and / or the second clustering parameter is greater than the second clustering parameter threshold, the first defect index deviation vector and the second defect index deviation vector are clustered into two classes. Specifically, if either clustering parameter does not satisfy its threshold condition, it is considered that the deviations of these two vectors in direction or in a certain dimension are large, and they should not be classified into the same class. Step S326: When the defect index deviation vector data traversal is completed, the defect index deviation vector clustering result is output. Specifically, steps S321 to S325 are repeated until all defect indicator deviation vectors have been processed. Based on the decision logic in the clustering process, similar vectors are grouped into the same class, forming the final clustering result. This implementation combines two measurement methods: cosine distance and Chebyshev distance. Cosine distance mainly focuses on the similarity of vectors in direction, used to capture the overall trend between vectors; Chebyshev distance focuses on the maximum deviation of vectors in each dimension, used to capture extreme cases or outliers. By considering both distance measures simultaneously, the similarity and differences between defect indicator deviation vectors are comprehensively evaluated, enabling more accurate clustering and achieving the technical effect of improving the accuracy and effectiveness of cluster analysis.

[0028] Step S400: Based on the high-frequency detected defect index type, perform correlation analysis on the production control attributes of the first node up to the Nth node to obtain associated control attributes, wherein the associated control attributes have node sequence number labels. Specifically, identify all nodes included in the production process. These nodes can be physical production procedures or equipment units, or logical operation stages or steps. Determine the total number of nodes N based on the complexity and detail of the production process. For each node (from the first node to the Nth node), collect all related production control attributes by interfacing with the production control system. These production control attributes refer to various parameters and conditions used to control and adjust the production process, directly affecting product quality, including but not limited to temperature, pressure, speed, time, raw material specifications, equipment status, operator information, etc. Production control attributes reflect the key control elements in the production process of that node. Organize the acquired data according to the production process sequence to form a list of production control attributes from the first node to the Nth node.

[0029] Based on the characteristics of the production process and business needs, a correlation analysis model is selected, including statistical methods (such as correlation coefficient, chi-square test, etc.), machine learning algorithms (such as decision trees, random forests, neural networks, etc.), or data mining techniques. The high-frequency defect detection index type is used as the target variable (or "dependent variable"), and the production control attribute of each node is used as a potential predictor variable (or "independent variable"). Using the defined correlation analysis model, the correlation degree or correlation between each production control attribute and the high-frequency defect detection index type is calculated. Based on the results of the correlation analysis, production control attributes that have a significant correlation with the high-frequency defect detection index type are selected through a significance threshold; these attributes are identified as correlated control attributes. For each selected correlated control attribute, its node number is recorded and labeled with a corresponding node number tag. This node number tag is a unique identifier used to identify different nodes in the production process, indicating which nodes' control attributes have a significant correlation with the defect index type.

[0030] In one possible implementation, a correlation analysis is performed on the production control attributes of the first node up to the Nth node based on the high-frequency detected defect index type to obtain the associated control attributes. Step S400 further includes step S410, extracting a first high-frequency detected defect index type according to the high-frequency detected defect index type; step S420, based on the first high-frequency detected defect index type, performing a joint negative sample retrieval of k control attributes on the production control attributes of the first node up to the Nth node to obtain the abnormal frequency ratio of the defect indexes of the k control attributes, where the total number of control attributes is ≥k≥1, and the initial value of k is equal to 1. Specifically, the first high-frequency detected defect index type is any one of the high-frequency detected defect index types. The initial value of k is set to 1, indicating that the analysis starts with a single control attribute. For the current combination of k control attributes (k represents the number of attributes in the combination, initially a single attribute), products with the highest frequency of detected defect index type are selected from all products as a candidate negative sample set. Within this candidate negative sample set, samples are further selected that maintain consistent values ​​or states (or meet preset conditions) on other control attributes (not the currently analyzed k attributes), but only exhibit defect-related values ​​or states on the current k control attributes. These samples constitute a refined negative sample set for the current k control attributes. In this refined negative sample set, the frequency of defect-related values ​​or states for the current k control attributes is statistically analyzed to obtain the abnormal frequency ratio of the k control attribute defect index, i.e., the proportion of samples exhibiting defect-related values ​​or states under a specific control attribute combination to the entire sample set (or reference set). This process is repeated, incrementing the value of k each time, until k equals the total number of control attributes. Step S430: When the abnormal frequency proportion of the k control attribute defect indicators is greater than or equal to the abnormal frequency proportion threshold, the k control attributes are stored together and added to the associated control attributes. Specifically, for each combination of k control attributes, it is checked whether the abnormal frequency proportion of its defect indicators is greater than or equal to the preset abnormal frequency proportion threshold. If so, this combination of k control attributes is considered to have a significant correlation with the first high-frequency detected defect indicator type, and it is stored together and added to the associated control attribute list. Step S440: When k equals the total number of control attributes, the associated control attributes are output. Specifically, when k equals the total number of control attributes, the retrieval and analysis process is stopped, and a list of all identified as associated control attributes is compiled and output. This implementation method gradually increases the number of control attributes and filters samples highly correlated with defects in each step, gradually narrowing the search range, and finally finding the control attribute or attribute combination most correlated with a specific defect type, thus achieving the technical effect of improving the accuracy of the analysis.

[0031] In one possible implementation, step S440 further includes step S441, where k is incremented by one when k is less than the total number of control attributes; step S442, where the k control attributes are deleted from the first node production control attributes up to the Nth node production control attributes to obtain the production control attribute update result; and step S443, where a k-plus-one correlation analysis is performed on the production control attribute update result based on the high-frequency detected defect index type to obtain the associated control attribute.

[0032] Specifically, check if the current value of k is less than the total number of control attributes. If so, increment k to prepare for analyzing the next larger combination of control attributes. Here, k is the number of attributes in the current combination of control attributes being analyzed; the total number of control attributes is the total number of all controllable attributes involved in the production process. From the original set of production control attributes (including production control attributes from the first node to the Nth node), delete the k combinations of control attributes that have already been analyzed. The updated production control attributes are used as input for the next iteration. Based on the updated set of production control attributes (i.e., the remaining attributes after excluding the k combinations of control attributes that have already been analyzed), perform a joint negative sample retrieval of k control attributes and calculation of the abnormal frequency ratio of defect indicators, similar to step S420. However, since k has been increased, a larger-scale combination of control attributes is being analyzed. If the abnormal frequency ratio of defect indicators in a new combination of k control attributes is greater than or equal to the abnormal frequency ratio threshold, it is considered a new associated control attribute combination and added to the associated control attribute list. Repeat this process until k equals the total number of control attributes. This implementation method ensures that all possible combinations of control attributes are taken into account by gradually increasing the value of k, thus achieving the technical effect of comprehensively exploring the correlation between production control attributes and high-frequency detected defect index types.

[0033] Step S500: Based on the node sequence label, traverse the associated control attributes and perform control variable matching on the high-frequency defect index deviation vector using big data to obtain process calibration attributes. Specifically, using a loop or iterative approach, traverse the list of associated control attributes one by one based on the node sequence label. Each associated control attribute represents one or more production control points related to the high-frequency defect index deviation vector. For the currently traversed associated control attribute, use big data analysis tools (such as data mining algorithms, machine learning models, etc.) to match the high-frequency defect index deviation vector, i.e., find the potential relationship or pattern between the control attribute and the defect index deviation. The matching process includes comparison and analysis of multiple dimensions, such as the historical data of the control attribute, its changing trend, and its correlation with other attributes. Based on the matching results, evaluate the contribution or influence of the control attribute on the high-frequency defect index deviation. Based on the control variable matching results, identify those control attributes that are highly correlated with or have a significant influence on the high-frequency defect index deviation vector as process calibration attributes. That is, process calibration attributes are key control attributes that need to be optimized or adjusted after analysis during the production process. These attributes directly affect product quality and are the focus of process optimization work. A process calibration attribute may be a single control attribute or a combination of multiple control attributes.

[0034] In one possible implementation, based on the node sequence label, the associated control attributes are traversed, and control variables are matched against the high-frequency defect index deviation vector based on big data to obtain process calibration attributes. Step S500 further includes step S510, extracting k control attributes of the first high-frequency detected defect index type based on the associated control attributes, wherein the total number of control attributes is ≥k≥1. Specifically, based on the analysis results of the associated control attributes, k control attributes are extracted, and the k control attributes are any one of the control attributes or attribute combinations most relevant to the first high-frequency detected defect index type. Step S520, based on the k control attributes, extracting the time series information of the monitoring feature values ​​and the time series information of the baseline feature values ​​of the k control attributes. Specifically, for the selected k control attributes, their monitoring feature value time series information and baseline feature value time series information are extracted respectively. The monitoring feature values ​​refer to the data collected in the actual production process, and the baseline feature values ​​are preset, ideal data or historical averages, etc. The time series information refers to the data sequence of these feature values ​​changing over time. The monitoring characteristic value time series information is the real-time or historical data sequence of k control attributes in the actual production process; the benchmark characteristic value time series information is the preset or calculated data sequence of k control attributes representing the normal or ideal state.

[0035] Step S530: By comparing the time-series information of the monitored feature values ​​of the k control attributes with the time-series information of the baseline feature values ​​of the k control attributes, anomaly analysis is performed on the k control attributes to obtain the joint anomaly coefficient of the k control attributes. Specifically, anomaly analysis is performed on the k control attributes using the time-series information of the monitored feature values ​​and the time-series information of the baseline feature values. By comparing the differences between the two, time periods or data points that deviate significantly from the baseline values ​​are identified. The joint anomaly coefficient of the k control attributes is calculated using algorithms (such as thresholding, statistical testing, etc.). This coefficient reflects the degree of anomaly of the k control attributes combined and is used to determine whether there is a significant anomaly. Step S540: When the joint anomaly coefficient of the k control attributes is greater than or equal to the joint anomaly coefficient threshold, the attribute deviation vector matrix of the joint anomaly coefficient of the k control attributes is used as the unique anomaly. Based on big data, historical production backtracking is performed to obtain the historical defect index types and historical defect index vectors. Specifically, when the joint anomaly coefficient of the k control attributes is greater than or equal to a preset joint anomaly coefficient threshold, these k attributes are considered to have a significant anomaly. In this case, the attribute deviation vector matrix of these k control attributes (i.e., their deviation from the benchmark value) is used as the unique anomaly. Big data technology is used for historical production backtracking, including querying production records in the historical database that are similar to the current anomaly, and extracting historical defect detection index types and historical defect detection index vectors (both of which are frequently occurring). Step S550: When the historical defect detection index type is consistent with the first high-frequency defect detection index type, and any one of the historical defect detection index vector and the high-frequency defect index deviation vector can be clustered into one class, then the k control attributes are added to the process calibration attribute of the first high-frequency defect detection index type. Specifically, if the historically detected defect index type is consistent with the currently analyzed first high-frequency detected defect index type, and any one of the deviation vectors of the historically detected defect index vectors and the high-frequency defect indexes corresponding to the first high-frequency detected defect index type can be clustered into one category (i.e., they have similar characteristics or patterns), then these k control attributes are considered to be the key factors causing the high-frequency defect. These k control attributes are added to the process calibration attribute list of the first high-frequency detected defect index type as the basis for subsequent process optimization and adjustment. In this implementation, the entire analysis process is based on a large amount of production data and historical data, reducing the interference of subjective judgment. By comparing the monitoring characteristic values ​​with the benchmark characteristic values, and the similarity between historical data and current data, potential problems are scientifically identified, achieving the technical effect of improving the objectivity and accuracy of the analysis.

[0036] In one possible implementation, anomaly analysis is performed on the k control attributes by comparing the time-series information of the monitored feature values ​​of the k control attributes with the time-series information of the baseline feature values ​​of the k control attributes, thereby obtaining a joint anomaly coefficient for the k control attributes. Step S530 further includes step S531, which involves statistically analyzing the percentage of the number of times the time-series information of the monitored feature values ​​of the k control attributes does not conform to the time-series information of the baseline feature values ​​of the k control attributes, thereby obtaining the joint anomaly coefficient for the k control attributes. Specifically, the time-series information of the monitored feature values ​​and the time-series information of the baseline feature values ​​are aligned in time so that they have corresponding values ​​at the same time point. For each pair of monitored feature values ​​and baseline feature values ​​(at the same time point) of the k control attributes, a comparison is made one by one. Based on a preset threshold or rule, it is determined whether the monitored feature value exceeds the normal range of the baseline feature value. If it exceeds the range, the control attribute at that time point is considered to be abnormal. The number of all control attributes with anomalies at all time points is counted, and the proportion of this number to the total number of monitoring points is calculated. This proportion is used as the joint anomaly coefficient for the k control attributes. The higher this coefficient, the more severe the overall anomaly. This implementation method uses the quantity ratio method to calculate the joint anomaly coefficient of k control attributes. The quantity ratio is a scientific, effective and easy-to-implement calculation method, achieving an intuitive and easy-to-understand technical effect.

[0037] Step S600: The node sequence number tag and the process calibration attribute are sent to the quality control terminal. Specifically, the node sequence number tag and the corresponding process calibration attribute are packaged and sent to the quality control terminal via wired or wireless means. Based on the received node sequence number tag and process calibration attribute, the quality control terminal adjusts the corresponding production control parameters or processes, sends the adjustment strategy to the corresponding production nodes, optimizes the production process, and establishes a feedback mechanism to monitor the adjusted production process and product quality, ensuring the effectiveness of the optimization measures. Based on the implemented production data and product quality feedback, the production process is continuously evaluated and optimized, forming a closed-loop quality control system. This application employs a method that automatically acquires defect index detection data of abnormal products when the product quality pass rate is lower than a preset threshold. It performs a two-level randomness analysis on the defect indexes, filters out high-frequency detected defect index types and their deviation vectors, and conducts correlation analysis between these high-frequency detected defect indexes and various stages of the production process to determine key control attributes affecting product quality. Based on big data analysis, it matches control variables to the deviation vectors of high-frequency defect indexes to obtain process calibration attributes, i.e., optimized production process parameters. These optimized process parameters are then promptly fed back to the quality control terminal for real-time adjustment and optimization of the production process. This approach accurately identifies and eliminates potential quality hazards in the production process, achieving the technical effect of precise control and optimization of the production process by uncovering the root causes of product quality problems.

[0038] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A manufacturing process tuning method based on defect pattern mining, characterized by, The method comprises the following steps: When the product quality pass rate is less than or equal to a quality pass rate threshold, obtaining detection defect index type data and defect index deviation vector data of quality abnormal products; Performing primary decoupling analysis according to the detection defect index type data to obtain high-frequency detection defect index types; Performing secondary decoupling analysis on the defect index deviation vector data based on the high-frequency detection defect index types to obtain a high-frequency defect index deviation vector; Performing correlation analysis on first node production control attributes to Nth node production control attributes based on the high-frequency detection defect index types to obtain associated control attributes, wherein the associated control attributes have node serial number labels; According to the node serial number labels, traversing the associated control attributes, and performing control variable matching on the high-frequency defect index deviation vector based on big data to obtain process calibration attributes; Sending the node serial number labels and the process calibration attributes to a quality management terminal; Performing correlation analysis on first node production control attributes to Nth node production control attributes based on the high-frequency detection defect index types to obtain associated control attributes, comprising: Extracting first high-frequency detection defect index types according to the high-frequency detection defect index types; Performing k-item control attribute joint negative sample retrieval on the first node production control attributes to the Nth node production control attributes based on the first high-frequency detection defect index types to obtain k-item control attribute defect index abnormal frequency proportions, control attribute total number≥k≥1, and k initial value equals to 1; When the k-item control attribute defect index abnormal frequency proportions are greater than or equal to an abnormal frequency proportion threshold, storing the k-item control attributes jointly and adding them to the associated control attributes; When k equals to the control attribute total number, outputting the associated control attributes; According to the node serial number labels, traversing the associated control attributes, and performing control variable matching on the high-frequency defect index deviation vector based on big data to obtain process calibration attributes, comprising: Extracting k-item control attributes of the first high-frequency detection defect index types according to the associated control attributes, wherein control attribute total number≥k≥1; Extracting k-item control attribute monitoring characteristic value time sequence information and k-item control attribute reference characteristic value time sequence information according to the k-item control attributes; Performing abnormality analysis on the k-item control attributes through the k-item control attribute monitoring characteristic value time sequence information and the k-item control attribute reference characteristic value time sequence information to obtain k-item control attribute joint abnormality coefficients; When the k-item control attribute joint abnormality coefficients are greater than or equal to a joint abnormality coefficient threshold, taking an attribute deviation vector matrix of the k-item control attribute joint abnormality coefficients as a unique abnormality quantity, and performing historical production backtracking based on big data to obtain historical detection defect index types and historical detection defect index vectors. When the historical detection defect indicator type is consistent with the first high-frequency detection defect indicator type, and any one of the historical detection defect indicator vector and the high-frequency defect indicator bias vector can be clustered into a class, the k control attributes are added to the first high-frequency detection defect indicator type process calibration attribute and the process calibration attribute.

2. The defect pattern mining based manufacturing process tuning method of claim 1, wherein, According to the detection defect indicator type data, a first decoupling analysis is performed to obtain a high-frequency detection defect indicator type, including: According to the detection defect indicator type data, the number of products detected by the first defect indicator type is counted to the number of products detected by the Mth defect indicator type. An interactive quality control terminal receives a product quantity threshold value. From the number of products detected by the first defect indicator type to the number of products detected by the Mth defect indicator type, defect indicator types greater than or equal to the product quantity threshold value are sorted and added to the high-frequency detection defect indicator type.

3. The defect mode mining based manufacturing process tuning method of claim 1, wherein, Based on the high-frequency detection defect indicator type, a secondary decoupling analysis is performed on the defect indicator bias vector data to obtain a high-frequency defect indicator bias vector, including: Obtaining the defect indicator bias vector data of the first high-frequency detection defect indicator type of the high-frequency detection defect indicator type; Performing clustering analysis on the defect indicator bias vector data to obtain defect indicator bias vector clustering results, wherein the defect indicator bias vector clustering results have a within-class vector quantity label; An interactive quality control terminal receives a within-class vector quantity threshold value. According to the within-class vector quantity label, clusters greater than or equal to the within-class vector quantity threshold value are sorted from the defect indicator bias vector clustering results to obtain the high-frequency defect indicator bias vector.

4. The defect mode mining based manufacturing process tuning method of claim 3, wherein, Performing clustering analysis on the defect indicator bias vector data to obtain defect indicator bias vector clustering results, including: Obtaining the first defect indicator bias vector and the second defect indicator bias vector of the defect indicator bias vector data of the first high-frequency detection defect indicator type of the high-frequency detection defect indicator type; Calculate the cosine distance of the first defect indicator bias vector and the second defect indicator bias vector to generate a first clustering parameter; Calculate the Chebyshev distance of the first defect indicator bias vector and the second defect indicator bias vector to generate a second clustering parameter; When the first clustering parameter is less than or equal to the first clustering parameter threshold value, and the second clustering parameter is less than or equal to the second clustering parameter threshold value, the first defect indicator bias vector and the second defect indicator bias vector are clustered into a class; When the first clustering parameter is greater than the first clustering parameter threshold value, or / and the second clustering parameter is greater than the second clustering parameter threshold value, the first defect indicator bias vector and the second defect indicator bias vector are clustered into two classes; When the defect indicator bias vector data is traversed, the defect indicator bias vector clustering results are output.

5. The defect mode mining based manufacturing process tuning method of claim 1, wherein, Also includes: When k is less than the total number of control attributes, k is incremented by one; The k control attributes are deleted from the first node production control attribute to the Nth node production control attribute to obtain a production control attribute update result; Perform k+1 correlation analysis on the production control attribute update result based on the high-frequency detected defect index type to obtain the correlation control attribute.

6. The defect mode mining based manufacturing process tuning method of claim 1, wherein, Perform abnormality analysis on the k control attributes by comparing the k control attribute monitoring characteristic value time sequence information with the k control attribute reference characteristic value time sequence information to obtain a k control attribute joint abnormality coefficient, including: Obtain the k control attribute joint abnormality coefficient by counting the proportion of the number of k control attribute monitoring characteristic value time sequence information that does not conform to the k control attribute reference characteristic value time sequence information.

Images