CN108960306B - Solder paste detection threshold value optimization method based on SMT big data - Google Patents
Solder paste detection threshold value optimization method based on SMT big data Download PDFInfo
- Publication number
- CN108960306B CN108960306B CN201810650120.4A CN201810650120A CN108960306B CN 108960306 B CN108960306 B CN 108960306B CN 201810650120 A CN201810650120 A CN 201810650120A CN 108960306 B CN108960306 B CN 108960306B
- Authority
- CN
- China
- Prior art keywords
- data
- spi
- value
- rate
- threshold value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
- G06F18/24155—Bayesian classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
Abstract
The invention provides a solder paste detection threshold optimization method based on SMT big data, which mainly solves the problems that potential defects cannot be found effectively in the PCB printing process and the defects of insufficient soldering and tin connection caused by solder paste amount cannot be reduced by a traditional SPI threshold setting method based on manual experience. The implementation scheme is as follows: constructing a threshold estimation data packet; estimating the state of the SPI parameter; and optimizing the SPI threshold value, taking the SPI detection value as an operation variable, taking a value corresponding to the condition that the sum of the misjudgment rate and the missed judgment rate is the lowest as a target function, optimizing the target function by using a genetic algorithm, solving the optimal threshold value, and setting the solved optimal threshold value as the SPI threshold value of the surface mounting technology. The scheme design of the invention is rigorous and complete, the threshold setting method has theoretical and feasible properties, the defects of insufficient solder, continuous solder and the like caused by the solder paste amount can be effectively controlled to flow into the SMT subsequent process in the surface mounting process, and the overall yield of the PCB is improved.
Description
Technical Field
The invention belongs to the technical field of industrial big data, mainly relates to optimization of a detection threshold, and particularly relates to a solder paste detection threshold optimization method based on SMT big data, which is used for optimizing a solder paste detection SPI threshold of SMT.
Background
The SMT production error is distributed in each link of each assembly process. Solder paste printing is used as the first process of the SMT process flow, and the printing process is a key link of the SMT process quality control because the printing process is influenced by various uncertain process parameters (such as a scraper, printing speed, a steel mesh, the width-thickness ratio and the area ratio of open holes and the like) and parameter setting has a large relation with professional knowledge and experience of operators. The SPI for detecting the solder paste can measure the characteristic quantities of the thickness, the area, the volume and the like of the solder paste and detect the defects of more tin, less tin, solder paste bridging and the like. If the defects of the solder paste on the PCB occur, the PCB needs to be cleaned and the solder paste needs to be printed again, so that the production efficiency is greatly influenced. The amount of solder paste is greatly related to defects such as cold solder joint and continuous solder joint. Through carrying out reasonable optimization to the SPI threshold value, monitor the state of tin cream on the PCB circuit board betterly, to unusual tin cream point in time reworking, can greatly reduce the cost of reprocessing, improved the qualification rate of PCB circuit board.
Currently, the SPI threshold is set mainly by manual experience. Firstly, an approximate range is set according to manual experience, then a PCB is printed in a trial mode, and a threshold interval is adjusted according to whether the alarm rate and the false alarm rate exceed expected set values or not. Traditional artificial experience lacks theoretical guidance, makes SPI equipment often misjudge and the emergence of failing to judge, and then real defect is put by the people, flows into follow-up station, has reduced the final yield of product. In addition, the IPC-7527 standard for solder paste printing also has the defect of the standard for detecting the solder paste amount.
The traditional SPI threshold setting method based on manual experience lacks theoretical guidance, has certain blindness, and is difficult to obtain a reasonable detection threshold through a data analysis means. If the detection threshold value is set unreasonably, potential defects cannot be found effectively in the printing process, after the PCB flows into subsequent processes such as reflow soldering and the like, defects such as insufficient solder, continuous tin and the like can occur, the defects such as insufficient solder, continuous tin and the like need manual maintenance, the production efficiency and the overall yield of SMT products are greatly reduced, and the requirements for production efficiency and product quality are difficult to meet.
Disclosure of Invention
The invention aims to provide an SPI threshold optimization method based on SMT big data, which can improve the quality of solder paste printing products and reduce the false alarm rate, aiming at the problems in the prior art.
The invention relates to an SPI threshold optimization method based on SMT big data, which is characterized by comprising the following steps:
(1) constructing a threshold estimation data packet:
(1a) data collection: collecting relevant data of three main stations of an SMT production line, namely SPI detection data, AOI automatic optical detection data and Repair data aiming at a bonding pad printed by the surface mounting technology of the same packaging type;
(1b) data association: the data are related into pad data marked by PCB serial numbers for PCB serial numbers common to SPI detection data, AOI detection data and Repair data;
(1c) data classification: dividing the associated pad SPI detection value into normal data and abnormal data according to whether a defect occurs or not;
(1d) data preprocessing: performing outlier elimination and data deduplication on the normal data and the abnormal data to form threshold value estimation data packets of the normal data and the abnormal data;
(2) and (3) estimating the state of the SPI parameter:
(2a) data sampling: selecting a proper sampling method to solve the problem of data type imbalance according to the specific conditions of normal data and abnormal data;
(2b) data normalization: normalizing the data of the normal data and abnormal data threshold estimation data packets;
(2c) determining the optimal window width: respectively calculating the optimal window widths of the normal data and the abnormal data according to a window width optimal formula;
(2d) estimation of SPI parameter probability density: selecting a Gaussian kernel function, and performing probability density estimation on normal data and abnormal data by using a kernel density estimation method to obtain probability density expressions of the normal data and the abnormal data;
(3) estimation of SPI threshold:
(3a) establishing an objective function: according to a Bayesian decision minimum error rate criterion, calculating a probability density expression of normal data to obtain an SPI misjudgment rate, and calculating a probability density expression of abnormal data to obtain an SPI misjudgment rate; taking an SPI detection value as an operation variable, and taking a value corresponding to the condition that the sum of the misjudgment rate and the missed judgment rate is the lowest as a target function;
(3b) solving an optimal threshold value: and optimizing the objective function by using a genetic algorithm, solving an optimal threshold value, and setting the optimal threshold value as an SPI threshold value of the surface mounting technology.
According to the invention, through reasonably setting the SPI threshold value, the number of the insufficient soldering and continuous soldering of the bonding pads is reduced, and the overall yield of the PCB is improved.
Compared with the prior art, the invention has the following advantages:
1) at present, the SMT technology is more and more widely applied, and the traditional manual experience setting threshold value is difficult to meet the requirements on production efficiency and product quality. The invention uses a data-driven process modeling technology to link a kernel density estimation method with a minimum error Bayesian decision theory in pattern recognition, and obtains a reasonable SPI threshold value through big data analysis to control defects of pad cold joint, solder joint and the like, thereby reducing blindness of traditional manual experience setting and improving production efficiency and product yield.
2) The kernel density estimation method introduced by the invention belongs to non-parameter estimation, the non-parameter estimation has a wide application range compared with parameter estimation, can process large sample data, has better robustness, does not need to assume the distribution type of the sample in advance, and does not depend on a specific mathematical model.
3) In the optimization of the target function, the genetic algorithm is selected to optimize the target function, and compared with other algorithms, the method has better global optimization and parallelism.
Drawings
FIG. 1 is a general flow chart of an implementation of the present invention;
FIG. 2 is a graph of the probability density of the SPI parameter in the present invention;
FIG. 3 is a flow chart of genetic algorithm optimization in the present invention;
FIG. 4 is a graph of probability density for abnormal data I (i.e., cold solder data) and normal data in accordance with the present invention;
FIG. 5 is a graph of probability density of abnormal data II (i.e., wicking data) and normal data according to the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
Example 1
The traditional SPI threshold value setting mainly depends on professional knowledge and experience of operators, a PCB is printed in a trial mode firstly, the experience threshold value can be obtained, the obtained threshold value lacks theoretical guidance, and the production efficiency is greatly reduced. The method is also contrary to the intellectualization and high efficiency of the modern manufacturing industry, and aiming at the problem, the invention provides an SPI threshold optimization method based on SMT big data through research, and the SPI threshold optimization method is shown in figure 1 and comprises the following steps:
(1) constructing a threshold estimation data packet:
(1a) data collection: the method comprises the steps of collecting historical data generated in the SMT production process aiming at the welding pads printed by the surface mount technology of the same packaging type, wherein the collected data comprises historical data of three main stations, namely SPI detection data, AOI automatic optical detection data and Repair data. The SPI detection data mainly comprise PCB serial numbers, pad position numbers, solder paste volumes, solder paste areas and solder paste height data. The AOI automatic optical detection data mainly comprises PCB serial numbers, pad position numbers, manual detection and machine detection. The Repair data mainly comprises PCB board serial numbers, pad bit numbers and defect types.
(1b) Data association: and for three data of SPI detection data, AOI detection data and Repair data, associating the data according to a PCB serial number field shared by the three data to form pad data marked by the PCB serial number, wherein the defect type of each pad can be known by the associated data.
(1c) Data classification: and dividing the associated pad data into normal data and abnormal data according to the defect type field. If the defect type is not defective, the data is normal data; if the defect type is a continuous tin or a false solder defect, the data is abnormal data.
(1d) Data preprocessing: detecting outliers in the normal data and the abnormal data respectively by using SPI detection values (solder paste volume, solder paste area and solder paste height), and removing the outliers; and respectively detecting the repeated data of the normal data and the abnormal data, deleting the repeated data, and only keeping one data. Thereby forming threshold estimation data packets of normal data and abnormal data.
(2) SPI parameter state estimation
(2a) Data sampling: selecting a proper sampling method to solve the problem of data type imbalance according to the specific conditions of normal data and abnormal data; sometimes, the data volume of normal data in the selected data is far larger than that of abnormal data, which is a phenomenon of data imbalance, and the phenomenon can cause larger errors of analysis results.
(2b) Data normalization: the SPI detection value data in the threshold value estimation packet of each of the normal data and the abnormal data is normalized. The normalization processing can improve the convergence speed of a threshold estimation program and improve the accuracy of the model.
(2c) Determining the optimal window width: and respectively calculating the optimal window width of the SPI detection value in the normal data threshold value estimation data packet and the optimal window width of the SPI detection value in the abnormal data threshold value estimation data packet according to a window width optimal formula. The core density estimation can be accurately carried out on the data by obtaining the optimal window width.
(2d) Estimation of SPI parameter probability density: selecting a Gaussian kernel function, and performing kernel density estimation on an SPI detection value in a normal data threshold estimation data packet by using a kernel density estimation method to obtain a kernel density estimation expression of normal data; and carrying out probability density estimation on the SPI detection value in the abnormal data threshold value estimation data packet to obtain a probability density expression of the abnormal data.
(3) SPI threshold estimation
(3a) Establishing an objective function: according to a Bayesian decision minimum error rate criterion, integral calculation is carried out on the probability density expression of the normal data to obtain an SPI misjudgment rate, integral calculation is carried out on the probability density expression of the abnormal data to obtain an SPI misjudgment rate; and taking the SPI detection value as an operation variable, and taking a value corresponding to the condition that the sum of the misjudgment rate and the missed judgment rate is the lowest as an objective function.
(3b) Solving an optimal threshold value: and optimizing the objective function by using a genetic algorithm, solving an optimal threshold value, and setting the optimal threshold value as an SPI threshold value of the surface mounting technology.
The invention uses a data-driven process modeling technology to link a kernel density estimation method with a minimum error Bayesian decision theory in pattern recognition, and obtains a reasonable SPI threshold value through big data analysis to control defects of pad cold joint, solder joint and the like, thereby improving the production efficiency and the product quality.
Example 2
The SPI threshold optimization method based on SMT big data is the same as that in embodiment 1, and the establishing of the objective function in step (3a) includes the following steps:
(3a1) establishing an objective function: and calculating the SPI misjudgment rate and the SPI missing judgment rate according to the minimum error Bayesian decision theory, wherein the expressions of the SPI misjudgment rate and the SPI missing judgment rate are shown as the following formulas.
The SPI false positive rate is:
the SPI miss rate is:
wherein f (x | w)1) The probability density function expression of the normal data obtained in the step (2 d); f (x | w)2) A probability density function expression of the abnormal data obtained in the step (2 d); t is a solder paste SPI detection value, the solder paste SPI detection value comprises a solder paste volume, a solder paste area and a solder paste height, the solder paste volume, the solder paste area and the solder paste height all have respective unit systems, and respective misjudgment rate and misjudgment rate of the volume, the area and the height need to be calculated respectively; w is a1Represents normal data; w is a2Representing anomalous data; x isThe variables are integrated.
(3a2) Using the SPI detection value as an operation variable, and setting a value corresponding to the case where the sum of the SPI false rate and the SPI false rate is minimized as an objective function minf (x):
in the formula, P1The SPI misjudgment rate is obtained; p2The SPI miss rate is obtained; MinF (x) is the minimum value of the sum of the false rate and the missed rate.
The non-parameter estimation method does not need to assume the distribution type of the sample in advance and does not depend on a specific mathematical model, but starts from the sample data, and compared with parameter estimation, the non-parameter estimation method has the characteristics of wide application range, capability of processing large sample data and good robustness. Therefore, the invention introduces a kernel density estimation method in nonparametric estimation, combines a minimum error rate Bayes decision criterion, establishes an SPI detection threshold value target optimization function, and solves the target function by using a genetic algorithm to obtain the optimal value of the detection threshold value.
Example 3
The SPI threshold optimization method based on SMT big data is the same as the embodiment 1-2, and in the step (3b), a genetic algorithm is used for optimizing the objective function to obtain the SPI detection optimal threshold.
The optimization process involved in the genetic algorithm flow for detection threshold optimization in the present invention is described as follows:
(3b1) and (4) self-defining the size of the population, the cross probability and the mutation probability. The population size is the number of the initialization generation threshold values; the cross probability is the probability of the cross of the individuals in various groups; the mutation probability is the probability of genetic mutation from the present generation population to the next generation.
(3b2) Initial population: and randomly generating an initial population according to the custom population size, wherein each individual represents the genotype of the chromosome. For example, if the custom population is 100, then 100 thresholds are randomly generated.
(3b3) And (3) optimization criterion: and stopping iteration when one of the four options of the maximum fitness, the average fitness, the evolution algebra and the ratio of the maximum fitness to the average fitness exceeds a preset value.
(3b4) Calculating the fitness: and calculating the fitness value of each individual, judging whether the fitness value meets the optimization criterion or not, if so, obtaining the best individual and the optimal solution of the characterization of the best individual, stopping the algorithm, executing (3b9), and outputting an SPI detection optimal threshold value. If the optimization criterion is not met, the next step is carried out to continue the optimization process of the genetic algorithm. For example, the 100 threshold values are respectively substituted into an objective function, 100 objective function values are calculated, the respective fitness values corresponding to the 100 threshold values are calculated according to the principle that the smaller the objective function value is, the larger the individual fitness is, and whether the optimization criterion is met is judged, if the optimization criterion is met, the threshold value corresponding to the maximum fitness among the 100 fitness values is obtained, the threshold value is the optimal solution of the objective function, the algorithm is stopped, and (3b9) is executed, and the SPI detection optimal threshold value is output; if the 100 fitness degrees are not met, the next step is carried out, and the optimization process of the genetic algorithm is continued.
(3b5) Selecting an individual: and screening to generate individuals again according to the fitness value, wherein the probability of selecting the individuals with high fitness value is high, and conversely, the probability of selecting the individuals with low fitness value is low or even eliminated. That is, the threshold value with a high fitness value is selected with a high probability, and the threshold value with a low fitness value is selected with a low probability.
(3b6) And (3) crossing: and generating the child individuals by using a certain crossing mode according to the crossing probability. For example, the crossover probability is defined as 0.1, a number is randomly generated, and if the number is less than 0.1, crossover operation is performed between thresholds; if the number is greater than 0.1, no crossover operations are performed between the thresholds. The crossing mode comprises one-point crossing, two-point crossing, multi-point crossing and uniform crossing.
(3b7) Mutation: generating offspring individuals by utilizing a certain mutation mode according to the mutation probability. For example, the variation probability is 0.1, a number is randomly generated, if the number is less than 0.1, the threshold value is varied, and the threshold value becomes a new threshold value; if the number is greater than 0.1, the threshold value is not changed, and a new generation of individuals is generated. The variation mode includes basic bit variation, uniform variation, boundary variation, non-uniform variation and Gaussian approximate variation.
(3b8) And (3) circularly calculating the fitness: a new generation of population is generated from steps (3b6) and (3b7) and then returns to (3b4) for the next round of fitness calculation until the best individual and its characterized optimal solution are obtained. Namely, the fitness of the next generation individual is calculated.
(3b9) And stopping the algorithm, and outputting the SPI detection optimal threshold value for the SPI optimal threshold value in the solder paste printing production process.
In the optimization of the target function, the genetic algorithm is selected to optimize the target function, and compared with other algorithms, the method has better global optimization and parallelism.
A detailed specific example is given below to further illustrate the present invention.
Example 4
The SPI threshold optimization method based on SMT big data is the same as the SPI threshold optimization method in the embodiment 1-3, and referring to FIG. 1, the implementation steps of the invention are as follows:
(1a) and (6) collecting data.
And collecting relevant data of three main stations of the SMT production line, namely SPI detection data, AOI detection data and Repair data, aiming at the welding pads of the same packaging type.
Referring to table 1, required field data of three stations of the SMT production line are collected.
Table 1 fields required for constructing a data packet
| 3D SPI station | AOI station | Repair station |
| PCB bar code | PCB bar code | PCB bar code |
| Pad bit number | Pad bit number | Pad bit number |
| Volume of solder paste | Manual detection | Type of defect |
| Area of solder paste | Machine detection | |
| Height of solder paste |
(1b) And (6) associating the data.
Referring to table 2, the collected data is correlated by fields (PCB bar code, pad number) common to SPI inspection data, AOI inspection data, and Repair data.
TABLE 2 data association Table
| Serial number | PCB bar code | Pad bit number | Volume (%) | Area (%) | Height (%) | Type of defect |
| …… | …… | …… | …… | …… | …… | …… |
(1c) And (6) classifying the data.
Referring to table 3, table 3 shows the standard for dividing the pad data of the same package type, and for the pad data of the same package, the data needs to be divided into normal data and abnormal data according to whether the pad is defective and the type of defect. Classifying the data of the pad subjected to the cold joint as abnormal data I; classifying the data with the defect type of continuous tin on the bonding pad as abnormal data II; and taking the data of the pads qualified by the SPI detection and the AOI detection as normal data.
TABLE 3 pad data partition criteria for the same package type
| Serial number | Data ofType (B) | |
| 1 | Normal data | No defect occurs, and the bonding pad is qualified through SPI detection and |
| 2 | Abnormal data I | Data with defect type of cold joint |
| 3 | Abnormal data II | Data indicating the type of occurrence of defects as tin-bonding |
(1d) Data preprocessing:
and (4) performing outlier elimination and data deduplication on the normal data and the abnormal data to improve the accuracy of kernel density estimation. There are two main processes for outlier detection: 1) eliminating the values of the volume, the area and the height of the solder paste which are zero; 2) the continuous tin and the insufficient soldering data caused by the non-tin paste amount need to be eliminated.
(2a) data sampling: the optimal sample size required by the nuclear density estimation is about 2000, so that the data type imbalance problem is solved by selecting a proper sampling method according to the specific conditions of normal data and abnormal data, and the accuracy of the nuclear density estimation is improved.
(2b) Data normalization: normalized formula is
Wherein x is the true value (volume, area, thickness, etc.) of the solder paste parameters,Height), xmax、xminThe maximum value and the minimum value in the real values of the solder paste parameters are respectively, y is a value after the solder paste is normalized, the convergence rate of a threshold estimation program can be improved after normalization, and the precision of the model is improved.
(2c) Determining the optimal window width: calculating the optimal window width h of the normal data and the abnormal data according to the optimal window width formulaopt。
Wherein, the sigma is an intermediate variable,
n is the sample size, xiThe observed values are the solder paste sample values,
mean values of solder paste samples.
(2d) And estimating the probability density of the SPI detection parameter.
And selecting a Gaussian kernel function, and performing probability density estimation on the normal data and the abnormal data by using a kernel density estimation method to obtain probability density expressions of the normal data and the abnormal data.
The kernel density estimation expression is:
wherein, f (x) is the probability density value to be estimated. n is the number of samples. h is the window width. K (x) is a kernel function, which generally satisfies symmetry and ═ k (x) dx ═ 1.
Referring to FIG. 2, FIG. 2 is a graph of probability density of SPI parameter in the present invention, wherein a curve L1 is a probability density curve of normal data, and a curve L1 is expressed as f (x | w)1) (ii) a L2 is the probability density curve of abnormal data, and the expression of L2 is f (x | w)2). If the dotted line m is set as a solder paste SPI detection parameter threshold (also called an alarm threshold), part of normal data obviously exceeds the threshold, so that misjudgment is generated, and the misjudgment probability is the small area enclosed by the dotted line m and a normal curve L1; in addition, a part of the abnormal data falls within the detection threshold, and a false-positive is generated, and the probability of false-positive is equal to the area of the region enclosed by the dotted line m and the curve L2. The sum of the misjudgment probability and the missed judgment probability is the target function.
(3) SPI detection threshold estimation:
(3a) establishing an objective function: and obtaining the SPI false rate and the SPI missing rate according to the Bayesian decision minimum error rate criterion.
The SPI false positive rate is:
the SPI miss rate is:
wherein f (x | w)1) Is the probability density function expression of normal data. f (x | w)2) Is a probability density function expression of abnormal data. And t is a detection value of the SPI of the solder paste.
And taking the SPI detection value as an operation variable, and taking a value corresponding to the condition that the sum of the misjudgment rate and the missed judgment rate is the lowest as an objective function MinF (x).
In the formula, P1Is the SPI false positive rate. P2Is the SPI miss rate. MinF (x) is the minimum value of the sum of the false rate and the missed rate.
(3b) Solving an optimal threshold value: and optimizing the target function by using a genetic algorithm to obtain an optimal threshold value by taking the minimum probability sum of misjudgment and missed judgment as a target.
Referring to fig. 3, the optimization process involved in the genetic algorithm flow is described as follows:
the first step is as follows: and (4) initial population. An initial population is randomly generated according to the population size, and each individual represents the genotype of a chromosome.
The second step is that: and calculating the fitness. Calculating the fitness value of each individual, judging whether the fitness value meets the following optimization criterion, if so, obtaining the best individual and the optimal solution represented by the best individual, and stopping the algorithm; if not, the process proceeds to the next step.
The third step: an individual is selected. Screening the regenerated individuals according to the fitness value, wherein the probability of selecting the individuals with high fitness value is high, and conversely, the probability of selecting the individuals with low fitness value is low, or even eliminated.
The fourth step: and (4) crossing. Generating the filial individuals by using a certain crossing mode according to a certain probability.
The fifth step: and (5) carrying out mutation. Generating the offspring individuals by using a certain variation mode according to a certain probability.
And a sixth step: and (5) circularly calculating the fitness. And generating a new generation of population from the fourth step and the fifth step, and then returning to the second step.
The optimization criteria in genetic algorithms generally have different determination modes according to different problems. One of the following is generally adopted as a judgment condition: in the invention, when one of the four options of the maximum fitness, the average fitness, the evolution algebra and the ratio of the maximum fitness to the average fitness exceeds a preset value, the condition is used for stopping optimization judgment, namely an optimization criterion.
The technical effects of the present invention will be described below by experiments and data thereof.
Example 5
SPI threshold value optimization method based on SMT big data as same as embodiments 1-4
(1a) Data collection and data association.
And selecting pad solder paste data with the packaging type of 0.5QFN of a batch of PCBs, wherein the pad solder paste data comprises SPI detection data, AOI detection data and replay data, and qualified data and defective data exist.
Referring to table 4, table 4 shows data associated with pad solder paste of package type 0.5 QFN.
TABLE 4 solder pad paste data for package type 0.5QFN
| Serial number | PCB bar code | Pad bit number | Volume (%) | Area (%) | Height (%) | Type of |
| 1 | 71071700015 | D3A2 | 60.98 | 83.53 | 137.26 | Insufficient solder joint |
| 2 | 7107170006 | C1A2 | 50.21 | 62.83 | 140.58 | Insufficient solder joint |
| 3 | 71071600003 | D1A3 | 112.15 | 97.69 | 137.77 | Qualified |
| …… | …… | …… | …… | …… | …… | …… |
| 98540 | 7107160004 | C1A3 | 200.89 | 106.55 | 158.67 | Tin connection |
| 98541 | 705987300265 | A1B2 | 65.24 | 78.21 | 139.85 | Insufficient solder joint |
| 98542 | 70598730012 | C1D1 | 109.84 | 103.3 | 127.61 | Qualified |
(1b) Data classification and data preprocessing.
Referring to table 3, the data of table 4 are divided into normal data, cold solder data, and continuous solder data. And (3) performing outlier rejection on the insufficient solder data and the continuous solder data: 1) eliminating the volume value of the cold solder data and the continuous tin data as 0; 2) eliminating the value higher than 100% in the cold joint data; values below 100% in the wicking data are excluded. And then, the normal data, the cold solder data and the continuous tin data are subjected to de-duplication. After all pretreatment operations were completed, some data are shown in the table below.
Referring to table 5, normal data for the PCB package is 0.5QFN pad.
TABLE 5 Normal data
Referring to table 6, table 6 is the cold joint data for the PCB package with 0.5QFN pads.
TABLE 6 rosin joint data
| Serial number | Volume (%) | Type of |
| 1 | 76.68 | Insufficient solder joint |
| 2 | 40.53 | Insufficient solder joint |
| 3 | 67.24 | Insufficient solder joint |
| 4 | 63.04 | Insufficient solder joint |
| 5 | 36.68 | Insufficient solder joint |
| …… | …… | …… |
| 2028 | 86.13 | Insufficient solder joint |
| 2029 | 95.09 | Insufficient solder joint |
| 2030 | 64.42 | Insufficient solder joint |
| 2031 | 55.54 | Insufficient solder joint |
| 2032 | 45.46 | Insufficient solder joint |
Referring to table 7, table 7 shows the solder joint data for the 0.5QFN pads for the PCB package.
TABLE 7 tin linkage data
| Serial number | Volume (%) | Type of |
| 1 | 122.15 | |
| 2 | 160.53 | Tin connection |
| 3 | 140.81 | |
| 4 | 230.55 | Tin connection |
| 5 | 166.98 | Tin connection |
| …… | …… | …… |
| 680 | 151.25 | Tin connection |
| 681 | 112.29 | Tin connection |
| 682 | 220.89 | Tin connection |
| 683 | 160.21 | Tin connection |
| 684 | 200.26 | Tin connection |
And 2, estimating the state of the SPI detection parameter.
(2a) And (6) sampling data.
Referring to tables 5-7, there are 84129 pieces of normal data, 2032 pieces of cold solder data and 684 pieces of continuous solder data. The number of normal data is far larger than the sum of the quantity of the cold solder data and the continuous tin data, and the data are unbalanced. The sample data size required by the nuclear density estimation is about 2000. And undersampling the normal data, and oversampling the continuous tin data to balance the various types of data. The volume of the cold joint data sample is proper, and sampling is not needed.
Referring to table 8, the normal data after undersampling is shown.
TABLE 8 undersampled Normal data
| Serial number | Volume (%) |
| 1 | 109.41 |
| 2 | 99.39 |
| 3 | 104.93 |
| 4 | 110.20 |
| 5 | 118.52 |
| …… | …… |
| 2125 | 106.73 |
| 2126 | 114.80 |
| 2127 | 134.58 |
| 2128 | 117.41 |
| 2129 | 105.34 |
Referring to Table 9, the tin linkage data after oversampling is shown.
TABLE 9 oversampled wicking data
| Serial number | Volume (%) | Type of |
| 1 | 122.15 | |
| 2 | 140.53 | Tin connection |
| 3 | 170.81 | |
| 4 | 130.55 | Tin connection |
| 5 | 166.22 | Tin connection |
| …… | …… | …… |
| 1980 | 151.25 | Tin connection |
| 1981 | 162.29 | Tin connection |
| 1982 | 140.89 | Tin connection |
| 1983 | 160.22 | Tin connection |
| 1984 | 179.26 | Tin connection |
(2b) And (6) normalizing the data.
In order to conveniently process data subsequently, convergence is ensured when an algorithm program runs, so that the model precision is improved, and normal data are normalized to a [0,1] interval.
Referring to table 10, the results of normalization of the normal data are shown.
TABLE 10 normalization of Normal data results
| Serial number | Volume of |
| 1 | 0.081264 |
| 2 | 0.087189 |
| 3 | 0.096994 |
| 4 | 0.101439 |
| 5 | 0.101650 |
| …… | …… |
| 2125 | 0.865970 |
| 2126 | 0.868086 |
| 2127 | 0.870132 |
| 2128 | 0.892000 |
| 2129 | 0.898349 |
Referring to table 11, the cold joint data is normalized.
TABLE 11 normalization of cold joint data
| Serial number | Volume of |
| 1 | 0.114301 |
| 2 | 0.122221 |
| 3 | 0.125135 |
| 4 | 0.141771 |
| 5 | 0.142565 |
| …… | …… |
| 2028 | 0.593203 |
| 2029 | 0.598130 |
| 2030 | 0.669890 |
| 2031 | 0.698019 |
| 2032 | 0.746945 |
Referring to Table 12, the results of normalization of the wicking data are shown.
TABLE 12 normalization of tin linkage data
| Serial number | Volume of |
| 1 | 0.012527 |
| 2 | 0.017487 |
| 3 | 0.038086 |
| 4 | 0.042206 |
| 5 | 0.043845 |
| …… | …… |
| 1980 | 0.808390 |
| 1981 | 0.820077 |
| 1982 | 0.845510 |
| 1983 | 0.944173 |
| 1984 | 0.965570 |
(2c) And determining the optimal window width. And calculating the optimal window width of the normal data and the abnormal data according to the optimal window width formula. Referring to table 13, the window widths are optimized for the abnormal data i, the abnormal data ii, and the normal data.
TABLE 13 optimal window widths for abnormal data I, abnormal data II, and normal data
| Serial number | Sample data type | |
| 1 | Abnormal data I (rosin joint data) | 0.0161756 |
| 2 | Abnormal data II (continuous tin data) | 0.0384185 |
| 3 | Normal data | 0.01944043 |
(2d) And estimating the probability density of the SPI detection parameter.
The kernel density estimation method introduced by the invention belongs to non-parameter estimation, the non-parameter estimation has a wide application range compared with parameter estimation, can process large sample data, has better robustness, does not need to assume the distribution type of the sample in advance, and does not depend on a specific mathematical model.
Firstly, a Gaussian kernel function is selected, and kernel density estimation is carried out on abnormal data I (namely, insufficient solder data) and normal data. Referring to fig. 4, a blue solid line is a normal data probability density curve, and a red dotted line is a probability density curve of abnormal data i.
The normal data probability density function expression is:
the probability density function expression of the abnormal data I is as follows:
referring to fig. 5, kernel density estimation is performed on abnormal data ii (i.e., the wicking data) and normal data, where a blue solid line is a probability density curve of the normal data, and a red dotted line is a probability density curve of the abnormal data ii.
The expression of the probability density function of the normal data is f (x | w)1)。
The probability density function expression of the abnormal data II is as follows:
step 3, estimating an SPI detection threshold value:
(3a) and establishing an objective function by taking the lower threshold as an operation variable.
When the lower limit of the SPI detection threshold is the LCL, the objective function is the sum of the small area surrounded by the blue solid line and the LCL and the small area surrounded by the red dotted line and the LCL in fig. 4.
The SPI false positive rate is:
the SPI miss rate is:
establishing an objective function according to the sum of the misjudgment rate and the missed judgment rate, wherein the expression is as follows:
(3b) and (3) establishing an objective function by taking the upper threshold as an operation variable:
when the upper limit of the SPI detection threshold is UCL, the objective function is the sum of the small region area surrounded by the blue solid line and UCL and the small region area surrounded by the red dotted line and UCL in fig. 5.
The SPI error determination expression is:
the SPI miss rate expression is:
establishing an objective function according to the sum of the misjudgment rate and the missed judgment rate, wherein the expression is as follows:
(3c) and solving an optimal threshold value.
With MinF1(x) For the objective function, in the interval [0,90 ]]Internally, optimizing by using genetic algorithm, and searching minimum MinF of objective function1(x) 0.0206, the optimum lower threshold LCL*＝69.218。
With MinF2(x) As an objective function, in the interval [100,170 ]]Internally, optimizing by using genetic algorithm, and searching minimum MinF of objective function2(x) 0.0156, when the optimal threshold value UCL is reached*＝145.4902。
In the optimization of the target function, the genetic algorithm is selected to optimize the target function, and compared with other algorithms, the method has better global optimization and parallelism.
While the foregoing is directed to embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.
In short, the solder paste detection threshold optimization method based on SMT big data mainly solves the problems that a traditional SPI threshold setting method based on manual experience lacks theoretical guidance, has certain blindness, cannot effectively find potential defects in the PCB printing process and reduce insufficient soldering and continuous soldering defects caused by solder paste amount and the like. The implementation scheme is as follows: constructing a threshold estimation data packet, and processing by collecting SPI detection data, AOI detection data and Repair data of an SMT production line and data; estimating the state of the SPI parameter, and introducing a kernel density estimation method KDE of non-parameter estimation into state estimation of solder paste detection parameters by using a data-driven process modeling technology; and optimizing the SPI threshold, establishing a target function of the SPI detection threshold by taking a misjudgment rate and a missed judgment rate as optimization targets and combining a minimum error rate Bayesian decision criterion, optimizing the target function by selecting a genetic algorithm, solving the optimal upper limit and the optimal lower limit of the detection threshold, and setting the optimal threshold as the SPI threshold of the surface mounting technology. The whole scheme of the invention has rigorous and complete design, the threshold setting method has theoretical and feasible properties, can effectively control the defects of insufficient solder, continuous solder and the like caused by the solder paste amount to flow into the SMT subsequent process in the surface mounting process, and improves the overall yield of the PCB.
Claims (3)
1. A solder paste detection threshold value optimization method based on SMT big data is characterized by comprising the following steps:
(1) constructing a threshold estimation data packet:
(1a) data collection: collecting relevant data of three main stations of an SMT production line, namely SPI detection data, AOI automatic optical detection data and Repair data aiming at a bonding pad printed by the surface mounting technology of the same packaging type;
(1b) data association: the data are related into pad data marked by PCB serial numbers for PCB serial numbers common to SPI detection data, AOI detection data and Repair data;
(1c) data classification: dividing the associated pad SPI detection value into normal data and abnormal data according to whether a defect occurs or not;
(1d) data preprocessing: performing outlier elimination and data deduplication on the normal data and the abnormal data to form threshold value estimation data packets of the normal data and the abnormal data;
(2) and (3) estimating the state of the SPI parameter:
(2a) data sampling: selecting a proper sampling method to solve the problem of data type imbalance according to the specific conditions of normal data and abnormal data;
(2b) data normalization: normalizing the data of the normal data and abnormal data threshold estimation data packets;
(2c) determining the optimal window width: respectively calculating the optimal window widths of the normal data and the abnormal data according to a window width optimal formula;
(2d) estimation of SPI parameter probability density: selecting a Gaussian kernel function, and performing probability density estimation on normal data and abnormal data by using a kernel density estimation method to obtain probability density expressions of the normal data and the abnormal data;
(3) estimation of SPI threshold:
(3a) establishing an objective function: according to a Bayesian decision minimum error rate criterion, calculating a probability density expression of normal data to obtain an SPI misjudgment rate, and calculating a probability density expression of abnormal data to obtain an SPI misjudgment rate; taking an SPI detection value as an operation variable, and taking a value corresponding to the condition that the sum of the misjudgment rate and the missed judgment rate is the lowest as a target function;
(3b) solving an optimal threshold value: and optimizing the objective function by using a genetic algorithm, solving an optimal threshold value, and setting the optimal threshold value as an SPI threshold value of the surface mounting technology.
2. The method of claim 1, wherein the establishing an objective function in step (3a) comprises the steps of:
(3a1) establishing an objective function: obtaining the SPI false rate and the SPI missing rate according to the Bayesian decision minimum error rate criterion,
the SPI false positive rate is:
P1＝∫t +∞f(x|w1)dx
the SPI miss rate is:
in the formula, w1Represents normal data; w is a2Representing anomalous data; f (x | w)1) Probability density function expression of normal data; f (x | w)2) A probability density function expression of abnormal data; t is a solder paste SPI detection value;
(3a2) taking the SPI detection value as an operation variable, and taking a value corresponding to the condition that the sum of the misjudgment rate and the missed judgment rate is the lowest as an objective function MinF (x);
in the formula, P1The SPI misjudgment rate is obtained; p2The SPI miss rate is obtained; MinF (x) is the minimum value of the sum of the false rate and the missed rate.
3. The SMT big data-based solder paste detection threshold optimization method according to claim 1, wherein in the step (3b), a genetic algorithm is used to optimize an objective function to obtain an SPI optimal threshold, with a goal that a sum of probabilities of erroneous judgment and missed judgment is minimized;
the optimization process involved in the genetic algorithm flow is described as follows:
(3b1) self-defining the size, the cross probability and the variation probability of a population;
(3b2) initial population: randomly generating an initial population according to the size of the user-defined population, wherein each individual represents the genotype of a chromosome;
(3b3) and (3) optimization criterion: when one of the four options of the maximum fitness, the average fitness, the evolution algebra and the ratio of the maximum fitness to the average fitness exceeds a preset value;
(3b4) calculating the fitness: calculating the fitness value of each individual, judging whether the fitness value meets the optimization criterion, if so, obtaining the best individual and the optimal solution of the characterization of the best individual, stopping the algorithm, and executing (3b 9); if not, the next step is carried out;
(3b5) selecting an individual: screening the regenerated individuals according to the fitness value, wherein the probability of selecting the individuals with high fitness value is high, and on the contrary, the probability of selecting the individuals with low fitness value is low or even eliminated;
(3b6) and (3) crossing: generating offspring individuals by utilizing a certain crossing mode according to the crossing probability;
(3b7) mutation: generating offspring individuals by utilizing a certain mutation mode according to the mutation probability;
(3b8) and (3) circularly calculating the fitness: generating a new generation of population from (3b6) and (3b7), and then returning to (3b4) to perform the next round of fitness calculation until the best individual and the optimal solution characterized by the best individual are obtained;
(3b9) and stopping the algorithm, and outputting the SPI detection optimal threshold value for the SPI optimal threshold value in the solder paste printing production process.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810650120.4A CN108960306B (en) | 2018-06-22 | 2018-06-22 | Solder paste detection threshold value optimization method based on SMT big data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810650120.4A CN108960306B (en) | 2018-06-22 | 2018-06-22 | Solder paste detection threshold value optimization method based on SMT big data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108960306A CN108960306A (en) | 2018-12-07 |
| CN108960306B true CN108960306B (en) | 2022-03-04 |
Family
ID=64491535
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810650120.4A Active CN108960306B (en) | 2018-06-22 | 2018-06-22 | Solder paste detection threshold value optimization method based on SMT big data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108960306B (en) |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110427593B (en) * | 2018-12-19 | 2022-12-02 | 西安电子科技大学 | SMT printing parameter optimization method based on industrial big data |
| CN109657718B (en) * | 2018-12-19 | 2023-02-07 | 广东省智能机器人研究院 | Data-driven SPI defect type intelligent identification method on SMT production line |
| CN109597968B (en) * | 2018-12-29 | 2021-06-08 | 西安电子科技大学 | SMT big data-based solder paste printing performance influence factor analysis method |
| CN111421954B (en) * | 2019-01-10 | 2022-02-18 | 鸿富锦精密电子(天津)有限公司 | Intelligent judgment feedback method and device |
| CN110533278B (en) * | 2019-07-15 | 2022-03-22 | 西安电子科技大学 | SMT production line detection threshold setting method based on particle swarm optimization algorithm |
| CN112801328B (en) * | 2019-11-14 | 2023-10-31 | 富联精密电子(天津)有限公司 | Product printing parameter setting device, method and computer readable storage medium |
| CN111010818B (en) * | 2019-11-15 | 2020-10-16 | 河北科技大学 | Smearing path planning method and device applied to tin paste smearing equipment |
| CN110996547B (en) * | 2019-11-15 | 2020-10-16 | 河北科技大学 | Driving parameter calculation method and device applied to tin paste smearing equipment |
| CN111899248B (en) * | 2020-08-05 | 2024-02-02 | 佛山市南海区广工大数控装备协同创新研究院 | Automatic PCB solder paste defect detection method based on machine learning |
| CN114375107A (en) * | 2020-10-15 | 2022-04-19 | 中兴通讯股份有限公司 | Method, device and equipment for reconstructing unstructured influence factors of solder paste printing of SMT (surface mount technology) production line |
| CN112632856B (en) * | 2020-12-21 | 2023-09-19 | 江苏警官学院 | Method for controlling speed and temperature of conveyor belt of reflow oven |
| CN113094980B (en) * | 2021-03-26 | 2023-05-02 | 西安交通大学 | Solder paste printing quality prediction method and system based on IGA-DNN |
| CN114004193B (en) * | 2021-10-14 | 2024-01-19 | 苏州浪潮智能科技有限公司 | Identification method and related device for tin-plating risk of PCB |
| CN115187587B (en) * | 2022-09-07 | 2022-11-29 | 常州铭赛机器人科技股份有限公司 | Detection method and imaging system for continuous tin detection |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6738505B1 (en) * | 1999-05-04 | 2004-05-18 | Speedline Technologies, Inc. | Method and apparatus for detecting solder paste deposits on substrates |
| CN103559399A (en) * | 2013-11-04 | 2014-02-05 | 广东粤海控股有限公司 | Water transportation pipe network leakage positioning method based on Bayesian decision theory and genetic algorithm |
| CN106033643A (en) * | 2015-03-10 | 2016-10-19 | 中国移动通信集团辽宁有限公司 | Data processing method and data processing device |
| CN107092654A (en) * | 2017-03-20 | 2017-08-25 | 山东科技大学 | Based on Change in Mean detect alarm normally with abnormal deviation data examination method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7103219B2 (en) * | 2001-04-12 | 2006-09-05 | Eastman Kodak Company | Population mixture modeling with an indeterminate number of sub-populations |
-
2018
- 2018-06-22 CN CN201810650120.4A patent/CN108960306B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6738505B1 (en) * | 1999-05-04 | 2004-05-18 | Speedline Technologies, Inc. | Method and apparatus for detecting solder paste deposits on substrates |
| CN103559399A (en) * | 2013-11-04 | 2014-02-05 | 广东粤海控股有限公司 | Water transportation pipe network leakage positioning method based on Bayesian decision theory and genetic algorithm |
| CN106033643A (en) * | 2015-03-10 | 2016-10-19 | 中国移动通信集团辽宁有限公司 | Data processing method and data processing device |
| CN107092654A (en) * | 2017-03-20 | 2017-08-25 | 山东科技大学 | Based on Change in Mean detect alarm normally with abnormal deviation data examination method and device |
Non-Patent Citations (5)
| Title |
|---|
| A novel approach for product makespan prediction in production life cycle;Jiantao Chang 等;《Int J Adv Manuf Technol(2015)》;20150419;1433-1448 * |
| QUANTIFYING THE IMPROVEMENTS IN THE SOLDER PASTE PRINTING PROCESS FROM STENCIL NANOCOATINGS AND ENGINEERED UNDER WIPE SOLVENTS;Chrys Shea 等;《International Conference on Soldering and Reliability》;20140513;1-50 * |
| 最大化约束密度单类分类器;赵加敏 等;《计算机科学》;20140228;第41卷(第2期);59-63 * |
| 河流水质参数估计的蒙特卡罗方法;严齐斌;《水利水电技术》;20061020;第37卷(第10期);14-16 * |
| 面向复杂工业大数据的实时特征提取方法;孔宪光 等;《西安电子科技大学学报（自然科学版）》;20161031;第43卷(第5期);70-74 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108960306A (en) | 2018-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108960306B (en) | Solder paste detection threshold value optimization method based on SMT big data | |
| TWI625615B (en) | Prediction model building method and associated predicting method and computer software product | |
| US7260444B2 (en) | Real-time management systems and methods for manufacturing management and yield rate analysis integration | |
| CN105302123B (en) | The monitoring method of on-line measurement data | |
| CN113469241A (en) | Product quality control method based on process network model and machine learning algorithm | |
| Tsai | Development of a soldering quality classifier system using a hybrid data mining approach | |
| CN102412168B (en) | Wafer defect defection method and system | |
| CN103676868A (en) | Automatic monitoring and intelligent analyzing system used in FPC manufacturing critical process | |
| CN110995153B (en) | Abnormal data detection method and device for photovoltaic power station and electronic equipment | |
| CN116307289B (en) | Textile processing procedure parameter detection and prediction method, system and storage medium | |
| CN113361958A (en) | Defect early warning method and system | |
| CN110856437B (en) | SMT production process control chart pattern recognition method | |
| TW201923688A (en) | Quality prediction method for multi-workstation system and system thereof | |
| CN104199417A (en) | Semiconductor coating technology statistical process control monitoring method | |
| CN108242411B (en) | Method and system for managing defects on a line | |
| Eger et al. | Knowledge capturing platform in multi-stage production systems for zero-defect manufacturing | |
| CN110910021A (en) | Method for monitoring online defects based on support vector machine | |
| CN105573269B (en) | The parameter monitoring system and method for semiconductor manufacturing board | |
| CN108304910B (en) | Air conditioner compressor slip sheet counting method based on machine vision | |
| Fallahnezhad et al. | Economic design of cumulative count of conforming control charts based on average number of inspected items | |
| CN110533278B (en) | SMT production line detection threshold setting method based on particle swarm optimization algorithm | |
| CN110307795B (en) | Automatic detection method for opening degree of crane hook | |
| CN117312802B (en) | SMT production line key process parameter optimization method based on man-machine fusion and storage medium | |
| JP5253745B2 (en) | Product manufacturing method and process management program | |
| Yang et al. | Modeling and implementation of a neurofuzzy system for surface mount assembly defect prediction and control |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |