A data-driven SMT solder paste printing quality prediction and analysis method and system

Abstract

The present invention provides a data-driven SMT solder paste printing quality prediction method and system, the method includes: obtaining the printer data on the SMT production line and the historical defect data of solder paste printing quality, oversampling and preprocessing the above data; reading the preprocessed data, using mutual information to screen out the printing process parameters, printing data, performance indicators and printing quality from the perspective of correlation and redundancy, and the strong correlation and low redundancy of the features, mining the printing defect data association rules to supplement the above features, and forming a key feature subset; constructing a solder paste printing quality prediction model, and inputting the key feature subset into the prediction model, predicting and classifying solder paste printing good products and various defect categories; using good product prediction accuracy, defect prediction accuracy, good product prediction error rate, defect prediction error rate, good product precision rate, and prediction accuracy index to evaluate the accuracy of the classification results. The present invention solves the technical problems of relying on manual operation, low prediction accuracy and single prediction index.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic assembly, and in particular to a data-driven SMT solder paste printing quality prediction and analysis method and system. Background Art

[0002] With the widespread application of intelligent electronic devices in the military and civilian electronics fields, high-density, high-integration, lightweight, multifunctional products and their core components have put forward higher requirements on the manufacturing efficiency, reliability and quality of electronic equipment manufacturing companies.

[0003] The SMT manufacturing process has a large amount of data, a wide variety of data, complex relationships between data, various forms of stored data, and chaotic quality data management. At present, poor solder paste printing quality is one of the important quality issues faced by the SMT manufacturing process. Printing quality problems caused by the shape and size of solder paste printing usually need to wait for subsequent inspection processes to be discovered, which may cause problems such as low printing efficiency, rework of work-in-progress, and waste of human resources. The existing SMT process quality influencing factors are complexly coupled, and printing quality factors are affected by the superposition of multiple parameters. Solder paste quality analysis mostly uses statistical process control (SPC) statistical process analysis, which has low efficiency in quality abnormality tracing and many lag control processes. It is difficult to timely avoid quality defects from flowing into subsequent processes based on the monitored manufacturing process data, and it is impossible to accurately predict the yield rate of SMT products. In addition, the solder paste inspection (SPI) link relies on manual experience to set the upper and lower thresholds of solder paste area, volume, and height, which makes it easy to have false alarms and missed alarms in actual production.

[0004] In recent years, the intelligence level of each process equipment of SMT production line has been improved, and various process, performance index data, printing defect data, etc. in the manufacturing process are collected and transmitted through sensors and data communication networks. Data-driven methods have inspired new ideas for solder paste quality prediction. Solder paste quality prediction helps to deeply explore the correlation between multiple parameters of each process in the SMT manufacturing process, dynamically adjust threshold parameters, and timely reduce, avoid and control quality problems that flow into the subsequent process. The existing invention patent document "A method for optimizing solder paste detection threshold based on SMT big data" with publication number CN108960306B builds a threshold estimation data packet; estimates the SPI parameter state; optimizes the SPI threshold, takes the SPI detection value as the operating variable, and makes the value corresponding to the minimum sum of the false positive rate and the missed positive rate reach the target function, uses the genetic algorithm to optimize the target function, and obtains the optimal threshold, and sets the obtained optimal threshold as the SPI threshold of the surface mounting technology. The existing patent document does not fully disclose the technical solution of the present application, nor can it achieve the technical effect of the present application. The existing patent application document with publication number CN109657718A is "A data-driven intelligent identification method for SPI defect categories on SMT production lines". In the first stage of its technical solution: clustering of SPI historical quality inspection data sets, using the Bagging algorithm to independently sample the obtained K-class training data sets 20 times, using the BP neural network model to train independent defect classifiers for a total of K×20 training sets of K categories, and obtaining K×20 independent defect classifiers to form a classifier set; the second stage: SPI online detection of 6 solder paste printing quality parameters, comparing the historical training data sets to classify the detection record "T", determine which category of the K-class training data sets the real-time detection point belongs to, when T is exactly on the boundary of two or more types of training data sets, then simultaneously select a total of 20 independent defect classifiers from multiple types of K-class training data sets in approximately equal amounts to determine the category of the detection record T; input T into each independent defect classifier and perform integrated prediction on the output results according to the integration rules to determine the defect category. This existing patent application document does not disclose the technical solution and specific technical features of the present application, and it also cannot achieve the technical effect of the present application.

[0005] At present, the quality prediction of SMT solder paste generally only analyzes the printing performance indicators of a single device. There is a lack of in-depth exploration of the correlation between the good / defective printing process parameters of the front-end process of the SMT production line, the printing performance indicators and the coupling of SMT solder paste printing quality, the extraction of data features that are strongly correlated with printing quality and have low redundancy, the construction of a solder paste printing quality prediction model, and the prediction of good and defective solder paste printing categories.

[0006] In summary, the existing technology has technical problems such as reliance on manual operation, low prediction accuracy and single prediction indicator. Summary of the invention

[0007] The technical problem to be solved by the present invention is how to solve the technical problems of reliance on manual operation, low prediction accuracy and single prediction index.

[0008] The present invention adopts the following technical solutions to solve the above technical problems: A data-driven SMT solder paste printing quality prediction and analysis method comprises:

[0009] S1. Read the process parameters and printing defect data of the printer SPI detection equipment on the SMI production line, use adasyn to oversample the historical defect data in the printing defect data to obtain oversampled defect data, and then remove duplicates and normalize them to obtain SMT quality preprocessing data;

[0010] S2. Based on mutual information, perform at least 2 rounds of feature selection from redundancy and correlation to obtain redundant correlation features, wherein the feature selection includes: obtaining the correlation between the process parameters, printing data, printing performance indicators and printing quality of the printer and SPI detection equipment on the SMT production line through the first round of correlation screening; obtaining the correlation between the process parameters, printing data and printing performance indicators of the printer and SPI detection equipment through the second round of redundancy screening, mining the historical defect data using the Apriori algorithm to obtain a set of k frequent items of historical defects in solder paste printing, and constructing a key feature data subset based on the redundant correlation features and the set of k frequent items of historical defects in solder paste printing;

[0011] S3. Build an SMT solder paste quality prediction model based on LightGBM to perform classification prediction according to the key feature data subset, thereby obtaining defect feature classification prediction data, and use the gridsearchCV function to adjust and optimize the defect feature classification prediction data to obtain the prediction false alarm rate, good product precision rate and prediction accuracy;

[0012] S4. Evaluate the prediction false alarm rate of the defect feature classification prediction data according to the prediction false alarm rate, the good product precision rate and the prediction accuracy.

[0013] The present invention combines the adasyn resampling method to solve the imbalance problem of good and defective samples, integrates the mutual information-apriori method to select features that affect the quality of solder paste printing, and improves data quality and effectiveness; the lightGBM method is used to construct an SMT solder paste printing quality classification model to improve the efficiency and accuracy of the prediction classification algorithm, provide a scientific decision-making basis for online monitoring of abnormal SMT solder paste printing volume, optimize printing process parameters, and improve the qualified rate of solder paste printing, and efficiently ensure the reliability of SMT products.

[0014] In a more specific technical solution, step S1 includes:

[0015] S11. Obtain the process parameters of the printer SPI inspection equipment, AOI inspection process parameters, good printing data, printing defect data set and printing performance index data in the SMT production line through a parameter acquisition tool, and transmit the aforementioned data through the OPC UA protocol;

[0016] S12, importing the Adasys from the unbalanced data processing library imlearn in Python to oversample the historical defect data to obtain the oversampled defect data;

[0017] S13, deleting NAN null values ​​in the oversampled defect data, importing a variance filtering method from the Python sklearn tool, and setting a deduplication threshold to dedupe the oversampled defect data;

[0018] S14. In order to unify the data dimension, a maximum-minimum normalization process is performed according to the following logic, so that the oversampled defect data is between (0, 1):

[0019]

[0020] The present invention collects process parameters of printers and SPI detection equipment on the SMT production line, printing data, printing performance index data and historical defect data of solder paste printing quality as original data sets, and uses the Adasys method to oversample the historical defect data to ensure a balance between defect data and good product data.

[0021] The present invention performs feature selection on the data set from the perspectives of relevance, redundancy and comprehensiveness, constructs an SMT solder paste printing quality prediction model, takes the key feature subset as the input of the prediction model, optimizes the prediction model parameters, predicts and classifies the solder paste printing good products and defect categories, and realizes real-time monitoring and prediction of solder paste printing quality.

[0022] In a more specific technical solution, the parameter acquisition tool in step S11 includes: semiconductor equipment SECS / GEM protocol and MES system.

[0023] In a more specific technical solution, step S2 includes:

[0024] S21, reading the SMT quality preprocessing data, and calculating the mutual information value of solder paste printing quality index between the SMT quality preprocessing data;

[0025] S22, sorting the mutual information values ​​of the SMT quality preprocessing data and the solder paste printing quality index in descending order, setting a threshold, and deleting features with a low correlation with the solder paste printing quality index;

[0026] S23. Obtain the mutual information correlation between the data, and delete the historical defect data with low correlation with the quality index based on the correlation.

[0027] S24, using the Apriori algorithm to mine the printing data related to the historical defect type in the historical defect data to supplement the selected associated features, thereby obtaining the key feature data subset.

[0028] The present invention reads processed SMT printing quality-related data, calculates mutual information between data, sorts the mutual information between all data and quality data, deletes features with low correlation with quality data according to the mutual information value, deletes features below a correlation threshold for mutual information values ​​between all data, adopts an Apriori method to mine association rules of historical defect data, complements the formed features, constructs a key feature subset, eliminates redundancy between data, and ensures that the association rules obtained in the present invention are strongly correlated with printing quality and have low-redundancy data features.

[0029] Since the SMT manufacturing process generates a massive amount of data, in order to avoid the "dimensionality disaster" problem, the present invention uses a "filter-complement" hybrid method to obtain data features with high correlation and low redundancy to quality indicators through two rounds of mutual information screening from the perspectives of correlation and redundancy. Combined with historical defect mining, supplementation and correction, a key feature subset that affects the production cycle is formed to achieve data dimensionality reduction, improve the quality and effectiveness of the data, and at the same time, the key feature subset is used as the input of the prediction model to further improve the accuracy of the prediction model.

[0030] In a more specific technical solution, in step S21, metrics are imported from the sklearn tool to call mutual_info_score to calculate the mutual information value of the solder paste printing quality index.

[0031] In a more specific technical solution, step S24 includes:

[0032] S241, create an initial frequent item set F with a length of 1 1 ;

[0033] S242, using the initial frequent item set F 1 The superset (k+1) of forms the candidate set;

[0034] S243, analogically checking whether the item set is a frequent item set, and deleting items in the candidate set that are less than a preset minimum support threshold;

[0035] S244, determining whether a new frequent item set is formed;

[0036] S245, if yes, then reconstruct the candidate item set;

[0037] S246: If not, then a set of k frequent items of the solder paste printing history defects is formed.

[0038] In a more specific technical solution, step S3 includes:

[0039] S31, taking the key feature data subset as a sample data set, and dividing the sample data set into a training data set and a test data set according to a preset ratio;

[0040] S32, using the sklearn tool and lightgbm module in Python to import solder paste quality prediction model data, based on which the SMT solder paste quality prediction model is constructed, and the training data set and the test data set are input into the SMT solder paste quality prediction model;

[0041] S33. Use the gridsearchCV function to optimize the maximum depth max_depth, the number of leaves num_leaves, and the minimum number of leaf records min_data_in_leaf of the Lightgbm tree.

[0042] The present invention uses SMT printing good / defect data, printing process parameters, printing performance indicators and other data as the driving force to build an SMT printing quality prediction model, providing a scientific decision-making basis for online quality monitoring and real-time control. The present invention reduces the reliance on statistical analysis, avoids the lag in quality abnormality processing, and considers the impact of the SMT production line printer, SPI detection equipment process parameters and printing data on printing quality, thereby improving the accuracy of solder paste printing quality prediction.

[0043] The present invention reduces the dependence of the adjustment operation of printing process parameters (scraper pressure, speed, solder paste area, volume, upper and lower thresholds of height, etc.) on manual repeated debugging. The present invention is suitable for the military electronic production mode of multiple varieties, small batches, and batch research and mixed production. The present invention fully extracts the complex, coupled, nonlinear, and uncertain relationship between process parameters and quality through association mining, thereby improving prediction analysis and production efficiency.

[0044] In a more specific technical solution, step S4 includes:

[0045] S41: Record the number of real good products T and the number of defects FR in the data labels in the test data set, and the number of good products TP and the number of defects FP detected by the prediction model;

[0046] S42: Calculate the predicted true yield rate TG, false yield rate FG, true defect rate TF and false defect rate FG in the SMT solder paste quality prediction model using preset logic;

[0047] S33: According to the prediction false alarm rate, the good product precision rate, and the prediction accuracy, the accuracy of the classification result is evaluated using the following logic to obtain the prediction false alarm rate:

[0048]

[0049] In a more specific technical solution, in step S42, the predicted true yield rate TG, false yield rate FG, true defect rate TF and false defect rate FG in the SMT solder paste quality prediction model are calculated using the following logic:

[0050]

[0051]

[0052]

[0053]

[0054] The present invention uses good product prediction accuracy, defect prediction accuracy, good product prediction error rate, defect prediction error rate, good product precision rate, and prediction accuracy index to evaluate the accuracy of classification results, thereby improving the accuracy of solder paste printing quality prediction analysis.

[0055] In a more specific technical solution, a data-driven SMT solder paste printing quality prediction and analysis system includes:

[0056] The SMI quality data acquisition module is used to read the process parameters and printing defect data of the printer SPI detection equipment on the SMI production line, and use adasyn to oversample the historical defect data in the printing defect data to obtain oversampled defect data, and then remove duplicates and normalize them to obtain SMT quality preprocessing data;

[0057] A key feature acquisition module is used to perform at least 2 rounds of feature selection from redundancy and correlation based on mutual information to obtain redundant correlation features, wherein the feature selection includes: a first round of correlation screening: correlation between the process parameters of the printer and SPI detection equipment on the SMT production line, printing data, printing performance indicators and printing quality; a second round of redundancy screening: correlation between the process parameters of the printer and SPI detection equipment, printing data, and printing performance indicators, using the Apriori algorithm to mine the historical defect data to obtain a k-item frequent item set of historical defects in solder paste printing, and construct a key feature data subset based on the redundant correlation features and the k-item frequent item set of historical defects in solder paste printing. The key feature acquisition module is connected to the SMI quality data acquisition module;

[0058] A classification prediction module is used to build an SMT solder paste quality prediction model based on LightGBM, perform classification prediction according to the key feature data subset, and obtain defect feature classification prediction data, and use the gridsearchCV function to adjust and optimize the defect feature classification prediction data to obtain the prediction false alarm rate, good product precision rate and prediction accuracy. The classification prediction module is connected to the key feature acquisition module;

[0059] A prediction evaluation module is used to evaluate the prediction false alarm rate of the defect feature classification prediction data according to the prediction false alarm rate, the good product precision rate and the prediction accuracy, and the prediction evaluation module is connected to the classification prediction module.

[0060] Compared with the prior art, the present invention has the following advantages: the present invention combines the adasyn resampling method to solve the imbalance problem of good and defective samples, integrates the mutual information-apriori method to select features that affect the quality of solder paste printing, and improves data quality and effectiveness; the lightGBM method is used to construct an SMT solder paste printing quality classification model to improve the efficiency and accuracy of the prediction classification algorithm, provide a scientific decision-making basis for online monitoring of abnormal SMT solder paste printing volume, optimize printing process parameters, and improve the qualified rate of solder paste printing, and efficiently ensure the reliability of SMT products.

[0061] The present invention collects process parameters of printers and SPI detection equipment on the SMT production line, printing data, printing performance index data and historical defect data of solder paste printing quality as original data sets, and uses the Adasys method to oversample the historical defect data to ensure a balance between defect data and good product data.

[0062] The present invention performs feature selection on the data set from the perspectives of relevance, redundancy and comprehensiveness, constructs an SMT solder paste printing quality prediction model, takes the key feature subset as the input of the prediction model, optimizes the prediction model parameters, predicts and classifies the solder paste printing good products and defect categories, and realizes real-time monitoring and prediction of solder paste printing quality.

[0063] The present invention reads processed SMT printing quality-related data, calculates mutual information between data, sorts the mutual information between all data and quality data, deletes features with low correlation with quality data according to the mutual information value, deletes features below a correlation threshold for mutual information values ​​between all data, adopts an Apriori method to mine association rules of historical defect data, complements the formed features, constructs a key feature subset, eliminates redundancy between data, and ensures that the association rules obtained in the present invention are strongly correlated with printing quality and have low-redundancy data features.

[0064] Since the SMT manufacturing process generates a massive amount of data, in order to avoid the "dimensionality disaster" problem, the present invention uses a "filter-complement" hybrid method to obtain data features with high correlation and low redundancy to quality indicators through two rounds of mutual information screening from the perspectives of correlation and redundancy. Combined with historical defect mining, supplementation and correction, a key feature subset that affects the production cycle is formed to achieve data dimensionality reduction, improve the quality and effectiveness of the data, and at the same time, the key feature subset is used as the input of the prediction model to further improve the accuracy of the prediction model.

[0065] The present invention uses SMT printing good / defect data, printing process parameters, printing performance indicators and other data as the driving force to build an SMT printing quality prediction model, providing a scientific decision-making basis for online quality monitoring and real-time control. The present invention reduces the reliance on statistical analysis, avoids the lag in quality abnormality processing, and considers the impact of the SMT production line printer, SPI detection equipment process parameters and printing data on printing quality, thereby improving the accuracy of solder paste printing quality prediction.

[0066] The present invention reduces the reliance on manual repeated debugging of the adjustment operation of printing process parameters (scraper pressure, speed, solder paste area, volume, upper and lower thresholds of height, etc.). The present invention is suitable for military electronic production modes with multiple varieties, small batches, and batch-research mixed production. The present invention fully extracts the complex, coupled, nonlinear, and uncertain relationship between process parameters and quality through association mining, thereby improving prediction analysis and production efficiency. The present invention solves the technical problems of reliance on manual operation, low prediction accuracy, and single prediction index in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a schematic diagram of the steps of a data-driven SMT solder paste printing quality prediction and analysis method according to Example 1 of the present invention;

[0068] Figure 2 This is a schematic diagram of specific steps of data preprocessing in Example 1 of the present invention;

[0069] Figure 3 This is a schematic diagram of specific steps for feature selection in Example 1 of the present invention;

[0070] Figure 4 This is a schematic diagram of specific steps for constructing a quality prediction model according to Embodiment 1 of the present invention;

[0071] Figure 5 Schematic diagram of specific steps for predicting and classifying solder paste quality according to Example 1 of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] Example 1

[0074] This embodiment provides a technical solution: a data-driven SMT solder paste quality prediction and analysis method. First, obtain SMT printing good / defect data, printer, SPI detection equipment process parameters, and historical printing defect data. Secondly, a two-round mutual information method is used to screen from the perspectives of correlation and redundancy, and to mine the complementary features of historical defect data and the aforementioned screened data to form a key feature subset. Then, an SMT solder paste quality prediction model is constructed, and the above feature subset is used as the prediction model input. Finally, the accuracy of the prediction classification results is evaluated by using the indicators of good product prediction accuracy, defect prediction accuracy, good product prediction error rate, defect prediction error rate, good product precision rate, and prediction accuracy.

[0075] In this embodiment, Python software tools are used to write programs, and tools such as pandas, numpy, and sklearn libraries are combined to perform data preprocessing, production cycle prediction, and correction.

[0076] like Figure 1 As shown, a data-driven SMT production line solder paste printing quality prediction and analysis method includes the following steps:

[0077] S1: Preprocessing of SMT quality-related data. In this embodiment, process parameters, printing data, printing performance index data and historical defect data of solder paste printing quality of the printer and SPI detection equipment on the SMT production line are obtained, and the adaptive comprehensive oversampling (adasyn) method is used to oversample the historical printing defect data and preprocess the above data;

[0078] S2: Constructing a key feature data subset; in this embodiment, the pre-processed data is read, and the printing process parameters, printing data, performance indicators and printing quality are screened out from the perspective of correlation and redundancy using mutual information to mine the printing defect data association rules, and the above features are supplemented to form a key feature subset;

[0079] S3: Constructing an SMT solder paste printing quality prediction model; in this embodiment, a solder paste printing quality prediction model is constructed, and a key feature subset is used as an input of the prediction model to predict and classify solder paste printing good products and various defect categories;

[0080] S4: Quality prediction classification evaluation; In this embodiment, the accuracy of good product prediction, defect prediction accuracy, good product prediction error rate, defect prediction error rate, good product precision rate, and prediction accuracy index are used to evaluate the accuracy of the classification results.

[0081] The present invention collects process parameters of printers and SPI detection equipment on the SMT production line, printing data, printing performance index data and historical defect data of solder paste printing quality as original data sets, performs feature selection on the data sets from the perspectives of relevance, redundancy and comprehensiveness, constructs an SMT solder paste printing quality prediction model, takes a key feature subset as the input of the prediction model, optimizes the prediction model parameters, predicts and classifies solder paste printing good products and defect categories, and realizes real-time monitoring and prediction of solder paste printing quality.

[0082] Furthermore, if Figure 2 As shown, step S1 includes:

[0083] S11: Obtain printing process parameters, printing performance index data, and historical printing good / defect data of printers and SPI inspection equipment in the SMT manufacturing process through the SECS / GEM protocol of semiconductor equipment or the MES system. In this embodiment, the SPI inspection of printers and AOI inspection process parameters, printing good / defect data, and printing performance index data of the SMT manufacturing process are obtained through the SECS / GEM protocol of semiconductor equipment or the MES system. The above data can be transmitted through the OPC UA protocol;

[0084] S12: To ensure the balance between defect data and good product data, the Adasys method is used to oversample the historical defect data. In this embodiment, to ensure the balance between defect data and good product data, Adasys is imported from the unbalanced data processing library imlearn in Python to oversample the historical defect data.

[0085] S13: The NAN values ​​of the above data are deleted, and variance filtering is used and the threshold is set to 0 to achieve data deduplication. In this embodiment, the NAN values ​​of the above data are deleted, and the variance filtering (variance threshold) method is imported from the Python sklearn tool, and the threshold is set to 0 to achieve data deduplication (the threshold can be adjusted according to the specific situation).

[0086] S14: To ensure that the above data have uniform dimensions and are between 0 and 1, the formula Perform maximum-minimum normalization processing. In this embodiment, in order to unify the data dimension, perform maximum-minimum normalization processing so that the above-mentioned various types of data are between (0, 1).

[0087] Furthermore, if Figure 3 The specific process of step S2 is as follows:

[0088] S21: Read the printing quality related data processed in step S13, and calculate the correlation between the data through mutual information. In this embodiment, read the data processed in step S1, import metrics from the sklearn tool, and call mutual_info_score to calculate the mutual information between all the data;

[0089] S22: setting a threshold based on mutual information to filter out features that are strongly correlated with printing process parameters, printing data, performance indicators and printing quality judgment results (good, passed, defective, etc.). In this embodiment, all data and the mutual information values ​​of solder paste printing quality indicators (good, passed, defective, etc.) are sorted from large to small, and a threshold is set to delete features with low correlation with solder paste printing quality indicators;

[0090] S23: setting a threshold based on mutual information to screen out the mutual information values ​​between printing process, printing data, and performance indicators from the perspective of redundancy. In this embodiment, based on step S22, if the mutual information correlation between data is other than the quality indicator, the data with low correlation with the quality indicator is deleted;

[0091] S24: The Apriori method is used to mine the association rules of the historical defect data, which is complementary to the features formed in step S23, and the adasyn method is used to oversample the historical defect data to construct a key feature subset; in this embodiment, the Apriori algorithm is used to mine the printing data related to the historical defect types (such as large volume, small volume, high height, low height, bridging, etc.) to form a supplement to the features screened in step S23. The specific steps are as follows:

[0092] The historical defect data of solder paste printing forms a set of k frequent items:

[0093] First, create an initial frequent item set F of length 1 1 ;

[0094] Secondly, the superset (k+1) of frequent itemsets forms the candidate set;

[0095] Then, by analogy, check whether the item set is a frequent item set, and delete the items in the candidate set that are less than the set minimum support threshold;

[0096] Finally, determine whether to form a new frequent item set. If found, reconstruct the candidate item set, otherwise form all frequent item sets.

[0097] Furthermore, if Figure 4 As shown, the specific process of step S3 is as follows:

[0098] S31: taking the key feature subset data obtained in step S24 as a sample data set, and dividing it into a training data set and a test data set according to a certain ratio;

[0099] S32: Use lightGBM to build an SMT solder paste printing quality prediction model, and use the key feature subset data of step S24 as the prediction model input. In this embodiment, the sklearn tool in python and the lightgbm module are imported to build an SMT solder paste quality prediction model, and the data of step S31 is used as the prediction model input;

[0100] S33: Calculate the true yield rate, false yield rate, true defect rate, and false defect rate in the predicted output results, and use the prediction false alarm rate, yield precision rate, and prediction accuracy indicators to evaluate the accuracy of the classification results; in this embodiment, the gridsearchCV function is used for the maximum depth (max_depth) of the Lightgbm tree, the number of leaves (num_leaves), the minimum number of records that a leaf may have (min_data_in_leaf) and other parameters.

[0101] like Figure 5 As shown, the specific implementation process of step S4 is as follows:

[0102] S41: Record the actual number of good products (TR) and the number of defects (FR) in the data labels, and the number of good products (TP) and the number of defects (FP) detected by the prediction model;

[0103] S42: Calculate the predicted true yield rate (TG), false yield rate (FG), true defect rate (TF), and false defect rate (FG) in the prediction model;

[0104] True yield rate:

[0105] False yield rate:

[0106] True defect rate:

[0107] False defect rate:

[0108] S43: The accuracy of the classification results is evaluated using three indicators: prediction false alarm rate, good product precision rate, and prediction accuracy.

[0109] Prediction false alarm rate:

[0110] Good product precision rate:

[0111] Prediction Accuracy:

[0112] In summary, the present invention combines the adasyn resampling method to solve the imbalance problem of good and defective samples, integrates the mutual information-apriori method to select features that affect the quality of solder paste printing, and improves data quality and effectiveness; the lightGBM method is used to construct an SMT solder paste printing quality classification model to improve the efficiency and accuracy of the prediction classification algorithm, and provide a scientific decision-making basis for online monitoring of abnormal SMT solder paste printing volume, optimization of printing process parameters, and improvement of solder paste printing good product qualification rate, thereby efficiently ensuring the reliability of SMT products.

[0113] The present invention collects process parameters of printers and SPI detection equipment on the SMT production line, printing data, printing performance index data and historical defect data of solder paste printing quality as original data sets, and uses the Adasys method to oversample the historical defect data to ensure a balance between defect data and good product data.

[0114] The present invention performs feature selection on the data set from the perspectives of relevance, redundancy and comprehensiveness, constructs an SMT solder paste printing quality prediction model, takes the key feature subset as the input of the prediction model, optimizes the prediction model parameters, predicts and classifies the solder paste printing good products and defect categories, and realizes real-time monitoring and prediction of solder paste printing quality.

[0115] The present invention reads processed SMT printing quality-related data, calculates mutual information between data, sorts the mutual information between all data and quality data, deletes features with low correlation with quality data according to the mutual information value, deletes features below a correlation threshold for mutual information values ​​between all data, adopts an Apriori method to mine association rules of historical defect data, complements the formed features, constructs a key feature subset, eliminates redundancy between data, and ensures that the association rules obtained in the present invention are strongly correlated with printing quality and have low-redundancy data features.

[0116] Since the SMT manufacturing process generates a massive amount of data, in order to avoid the "dimensionality disaster" problem, the present invention uses a "filter-complement" hybrid method to obtain data features with high correlation and low redundancy to quality indicators through two rounds of mutual information screening from the perspectives of correlation and redundancy. Combined with historical defect mining, supplementation and correction, a key feature subset that affects the production cycle is formed to achieve data dimensionality reduction, improve the quality and effectiveness of the data, and at the same time, the key feature subset is used as the input of the prediction model to further improve the accuracy of the prediction model.

[0117] The present invention uses SMT printing good / defect data, printing process parameters, printing performance indicators and other data as the driving force to build an SMT printing quality prediction model, providing a scientific decision-making basis for online quality monitoring and real-time control. The present invention reduces the reliance on statistical analysis, avoids the lag in quality abnormality processing, and considers the impact of the SMT production line printer, SPI detection equipment process parameters and printing data on printing quality, thereby improving the accuracy of solder paste printing quality prediction.

[0118] The present invention reduces the reliance on manual repeated debugging of the adjustment operation of printing process parameters (scraper pressure, speed, solder paste area, volume, upper and lower thresholds of height, etc.). The present invention is suitable for military electronic production modes with multiple varieties, small batches, and batch-research mixed production. The present invention fully extracts the complex, coupled, nonlinear, and uncertain relationship between process parameters and quality through association mining, thereby improving prediction analysis and production efficiency. The present invention solves the technical problems of reliance on manual operation, low prediction accuracy, and single prediction index in the prior art.

[0119] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven SMT solder paste printing quality prediction and analysis method, characterized in that: The method comprises: S1. Read the process parameters and printing defect data of the printer SPI detection equipment on the SMI production line, use adasyn to oversample the historical defect data in the printing defect data to obtain oversampled defect data, and obtain SMT quality preprocessing data by deduplication and normalization; S2. Based on mutual information, at least 2 rounds of feature selection are performed from redundancy and correlation to obtain redundant correlation features. The feature selection includes: obtaining the correlation between the process parameters, printing data, printing performance indicators and printing quality of the printer and SPI detection equipment on the SMT production line through the first round of correlation screening; obtaining the correlation between the process parameters, printing data and printing performance indicators of the printer and SPI detection equipment through the second round of redundancy screening, and using the Apriori algorithm to mine the historical defect data to obtain the historical defects of solder paste printing. The frequent item set is based on the redundant association features and the solder paste printing history defects. Build key feature data subsets from frequent item sets; S3. Build an SMT solder paste quality prediction model based on LightGBM, perform classification prediction based on the key feature data subset to obtain defect feature classification prediction data, and use the gridsearchCV function to adjust and optimize the defect feature classification prediction data to obtain the prediction false alarm rate, good product precision rate and prediction accuracy; S4. Evaluate the prediction false alarm rate of the defect feature classification prediction data according to the prediction false alarm rate, the good product precision rate and the prediction accuracy.

2. A data-driven SMT solder paste printing quality prediction and analysis method according to claim 1, characterized in that: The step S1 comprises: S11. Obtain the process parameters of the printer SPI inspection equipment, AOI inspection process parameters, good printing data, printing defect data set and printing performance index data in the SMT production line through a parameter acquisition tool, and transmit the aforementioned data through the OPC UA protocol; S12, importing Adasys from the imbalanced data processing library imlearn in Python to oversample the historical defect data to obtain the oversampled defect data; S13, deleting NAN null values ​​in the oversampled defect data, importing a variance filtering method from the Python sklearn tool, and setting a deduplication threshold to dedupe the oversampled defect data; S14. In order to unify the data dimension, a maximum-minimum normalization process is performed according to the following logic, so that the oversampled defect data is between (0, 1): 。 3. A data-driven SMT solder paste printing quality prediction and analysis method according to claim 2, characterized in that: The parameter acquisition tool in step S11 includes: semiconductor equipment SECS / GEM protocol and MES system.

4. A data-driven SMT solder paste printing quality prediction and analysis method according to claim 1, characterized in that: The step S2 comprises: S21, reading the SMT quality preprocessing data to calculate the mutual information value of solder paste printing quality index between the SMT quality preprocessing data; S22, sorting the mutual information values ​​of the SMT quality preprocessing data and the solder paste printing quality index in descending order, setting a threshold, and deleting features with a low correlation with the solder paste printing quality index; S23, obtaining the mutual information correlation between the data to delete the historical defect data with low correlation with the quality index; S24. Using the Apriori algorithm to mine the printing data related to the historical defect type in the historical defect data to supplement the filtered associated features, so as to obtain the key feature data subset.

5. A data-driven SMT solder paste printing quality prediction and analysis method according to claim 4, characterized in that: In the step S21, metrics is imported from the sklearn tool to call mutual_info_score to calculate the mutual information value of the solder paste printing quality index.

6. A data-driven SMT solder paste printing quality prediction and analysis method according to claim 4, characterized in that: The step S24 comprises: S241, create an initial frequent itemset of length 1 ; S242, using the initial frequent item set The superset (k+1) of forms the candidate set; S243, analogically checking whether the item set is a frequent item set, and deleting items in the candidate set that are less than a preset minimum support threshold; S244, determining whether a new frequent item set is formed; S245, if yes, then reconstruct the candidate item set; S246, if not, the solder paste printing history defect is formed Frequent item set.

7. The data-driven SMT solder paste printing quality prediction and analysis method according to claim 1, characterized in that: The step S3 comprises: S31, taking the key feature data subset as a sample data set, and dividing the sample data set into a training data set and a test data set according to a preset ratio; S32, using the sklearn tool and lightgbm module in Python to import solder paste quality prediction model data to build an SMT solder paste quality prediction model, and input the training data set and the test data set into the SMT solder paste quality prediction model; S33. Use the gridsearchCV function to optimize the maximum depth max_depth, the number of leaves num_leaves, and the minimum number of leaf records min_data_in_leaf of the Lightgbm tree.

8. The data-driven SMT solder paste printing quality prediction and analysis method according to claim 1, characterized in that: The step S4 comprises: S41: Record the number of true good products in the data labels in the test data set , number of defects , the number of good products detected by the prediction model And the number of defects ; S42: Calculate the predicted true yield rate in the SMT solder paste quality prediction model using preset logic , Fake yield rate , True Defect Rate and false defect rate ; S43: According to the prediction false alarm rate, the good product precision rate, and the prediction accuracy, the accuracy of the classification result is evaluated by the following logic to obtain the prediction false alarm rate: 。 9. A data-driven SMT solder paste printing quality prediction and analysis method according to claim 8, characterized in that: In the step S42, the predicted true yield rate TG, false yield rate FG, true defect rate TF and false defect rate FG in the SMT solder paste quality prediction model are calculated using the following logic: 。 10. A data-driven SMT solder paste printing quality prediction and analysis system, characterized in that: The system comprises: The SMI quality data acquisition module is used to read the process parameters and printing defect data of the printer SPI detection equipment on the SMI production line, and use adasyn to oversample the historical defect data in the printing defect data to obtain oversampled defect data, and then remove and normalize the data to obtain SMT quality preprocessing data; The key feature acquisition module is used to perform at least two rounds of feature selection from redundancy and correlation based on mutual information to obtain redundant correlation features. The feature selection includes: the first round of correlation screening: the correlation between the process parameters of the printer and SPI detection equipment on the SMT production line, printing data, printing performance indicators and printing quality; the second round of redundancy screening: the correlation between the process parameters of the printer and SPI detection equipment, printing data, and printing performance indicators, and the Apriori algorithm is used to mine the historical defect data to obtain the historical defects of solder paste printing. The frequent item set is based on the redundant association features and the solder paste printing history defects. The frequent item set constructs a key feature data subset, and the key feature acquisition module is connected with the SMI quality data acquisition module; A classification prediction module is used to build an SMT solder paste quality prediction model based on LightGBM, perform classification prediction according to the key feature data subset to obtain defect feature classification prediction data, and use the gridsearchCV function to adjust and optimize the defect feature classification prediction data to obtain the prediction false alarm rate, good product precision rate and prediction accuracy. The classification prediction module is connected to the key feature acquisition module; A prediction evaluation module is used to evaluate the prediction false alarm rate of the defect feature classification prediction data according to the prediction false alarm rate, the good product precision rate and the prediction accuracy, and the prediction evaluation module is connected to the classification prediction module.

Images