WO2022267509A1 - Method for training smt printing parameter optimization model, device, and storage medium - Google Patents

Abstract

A method for training an SMT printing parameter optimization model, a device, and a storage medium. The method comprises: receiving initial production data; performing influencing factor reconstruction according to the initial production data to obtain an influencing factor data packet; loading an optimization model to be trained; and training said optimization model according to the influencing factor data packet to obtain a trained optimization model.

Description

Training method, equipment and storage medium for SMT printing parameter optimization model

Cross References to Related Applications

This application is based on the Chinese patent application with the application number "202110715871.1" and the filing date is June 24, 2021, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is hereby incorporated by reference. Application.

technical field

This application belongs to the field of intelligent manufacturing technology, and involves surface mount technology (Surface Mounted Technology, referred to as: SMT) production line printing and solder paste inspection (SolderPaste Inspection, referred to as: SPI) stage, especially relates to a SMT printing parameter optimization model training , equipment and storage media.

Background technique

In surface mount technology, most quality problems occur in the solder paste printing stage, so how to set more reasonable and correct printing process parameters is particularly important.

When setting process parameters, there are two main ways, one is to rely on human experience and use trial printing to set, and the other is to use data mining models to optimize the printing process parameters of SMT production lines. However, there are various problems and deficiencies in these methods, including: insufficient analysis of various process parameters involved in the production process, missing some important parameters that affect the quality of solder paste printing, and the analysis of the relationship between printing parameters and quality by a single model Insufficient and insufficient analysis of print quality indicators. As a result, the accuracy of relevant quality prediction models and process parameter optimization is still not high.

Therefore, there is an urgent need for a method that can better improve the accuracy of process parameter optimization and prediction.

Contents of the invention

An embodiment of the present application provides a training method for an SMT printing parameter optimization model, the method comprising the following steps: receiving initial production data; reconstructing influencing factors according to the initial production data to obtain influencing factor data packets; loading to be trained Optimizing the model: training the optimization model to be trained according to the influencing factor data package to obtain a trained optimization model.

The embodiment of the present application also provides a computer device, the device includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and when executing the computer program Steps for realizing the training method of the SMT printing parameter optimization model as described above.

The embodiment of the present application also provides a storage medium for computer-readable storage, the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to realize The steps of the training method of the aforementioned SMT printing parameter optimization model.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present application more clearly, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application. Ordinary technicians can also obtain other drawings based on these drawings on the premise of not paying creative work.

Fig. 1 is the schematic flow chart of the training method of a kind of SMT printing parameter optimization model that an embodiment of the present application provides;

Fig. 2 is a schematic flow chart of the steps of obtaining the influencing factor data package provided by an embodiment of the present application;

FIG. 3 is a schematic flowchart of the steps of obtaining the influencing factor data package provided by another embodiment of the present application;

Fig. 4 is a schematic flowchart of the steps of obtaining the influencing factor data package provided by another embodiment of the present application;

FIG. 5 is a schematic flowchart of the steps of training the optimization model to be trained provided by an embodiment of the present application;

FIG. 6 is a schematic flowchart of the steps of iterative training provided by an embodiment of the present application;

FIG. 7 is a schematic flow diagram of the steps of obtaining a trained optimization model provided by an embodiment of the present application;

Fig. 8 is a schematic flow chart of the process parameter prediction and recommended steps provided by an embodiment of the present application;

Fig. 9 is a schematic structural block diagram of a computer device provided by an embodiment of the present application.

detailed description

The main purpose of the embodiments of the present application is to propose a training, equipment and storage medium for an SMT printing parameter optimization model, aiming at improving the prediction result of printing quality in the SMT production line and improving the accuracy of process parameter recommendation.

The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of them. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the scope of protection of this application.

The flow charts shown in the drawings are just illustrations, and do not necessarily include all contents and operations/steps, nor must they be performed in the order described. For example, some operations/steps can be decomposed, combined or merged, so the actual order of execution may be changed according to the actual situation.

As shown in Figure 1, Figure 1 is a schematic flow diagram of a training method for an SMT printing parameter optimization model provided by an embodiment of the present application, the method comprising the following steps:

Step S101, receiving input initial production data.

SMT is Surface Mount Technology (Surface Mount Technology), which is called Surface Mount Technology in English. It is the most popular technology and process in the electronic assembly industry. SMC/SMD for short, chip components in Chinese) are installed on the surface of a printed circuit board (Printed Circuit Board, referred to as: PCB) or on the surface of other substrates, and are soldered and assembled by reflow soldering or dip soldering. technology.

When performing surface assembly, it is not done randomly. It is necessary to determine the actual assembly or placement method according to the actual situation, including the actual information of the product. At the same time, how to set the placement parameters during placement is also It is very necessary, so it is necessary to select the relevant process parameters reasonably and correctly to realize the surface mount of the device.

In one embodiment, in order to better realize the placement of components, more correct process parameters are required, and in order to have better process parameters during placement, relevant models will be pre-built and trained at this time, In order to recommend a more suitable combination of process parameters according to the actual situation. When training the SMT printing parameter optimization model, the input initial production data is firstly received, and then the parameter optimization model to be trained is trained according to the obtained initial production data.

Step S102 , reconstruct the influencing factors according to the initial production data, and obtain the reconstructed influencing factor data package.

After the input initial production data is received, the initial production data needs to be processed, specifically, the received initial production data is reconstructed to obtain a reconstructed influencing factor data package.

In fact, when training the SMT printing parameter optimization model (hereinafter referred to as the optimization model), firstly, the training samples for training are determined, and then the optimization model to be trained is trained according to the obtained training samples. Specifically, firstly, the input initial production data is received, and then the impact factors are reconstructed on the received initial production data, so as to obtain the impact factor data package after the reconstruction of the influence factors is completed.

Exemplarily, the initial production data input for model training includes a lot, specifically including manpower, machine, material, method, environment, measurement and collection of all collectable data in the printing stage of the SMT production line and SPI detection, and then the The received data is further processed.

For example, taking the solder paste printing process of the SMT production line as an example, when obtaining the original production data of the printing stage of the SMT production line, the obtained original production data includes the following six types of data, namely: time batch data, PCB board attribute data, Process parameter data, small elements, printing process data and SPI detection data. Among them, PCB board attribute data includes PCB board length, board width and board height; process parameter data includes scraper pressure, scraper speed, demoulding speed, demoulding distance, automatic cleaning count and cleaning speed; small elements include scraper separation distance, Squeegee separation speed, manual cleaning count, cleaning options, table printing height compensation, automatic cleaning and manual cleaning; printing process data include printing time, production count, scraper count, Mask count, Mark_1X axis coordinates, Mark_1Y axis coordinates, Mark_2X axis coordinates , Mark_2 Y-axis coordinates, work separation delay, start offset, end offset, down offset, cleaning supply time, wiping paper waiting distance, machine waiting time, solder paste usage count, average pressure, minimum pressure, maximum pressure and Post-squeegee counting; SPI detection data includes solder paste volume and area.

It should be noted that when the initial production data for training the optimization model is received, the obtained initial production data is not directly used for training, but the influencing factors are reconstructed for the initial production data. Specifically, the initial production data records many different factors that have an impact on the actual SMT printing. At the same time, there are many of these factors. If all the data are directly used for model training at this time, it will increase the cost of model training. At the same time, it will also make the efficiency of model training extremely low, and because all the data is used for model training, the model will be too single, and the model will not be accurate when recommending process parameters. Therefore, after receiving the initial production data, the influencing factors of the initial production data will be reconstructed to fully consider the impact of feature hidden information on the model, so that the trained model can have higher prediction accuracy and recommendation accuracy.

Referring to FIG. 2 , FIG. 2 is a schematic flowchart of the steps of obtaining the influencing factor data package provided by an embodiment of the present application.

In one embodiment, after receiving the inputted initial production data, the initial production data will be reconstructed for influencing factors, so as shown in FIG. 2 , the specific steps for reconstructing influencing factors include step S201 to step S202.

Step S201. Filter and screen the initial production data to obtain an original data set that does not contain fixed attributes.

After the initial production data is obtained, since all the data contained in the initial production data are not necessarily the data that have an impact on SMT printing, after receiving the initial production data, all the features contained in the initial production data will be /Data is filtered to obtain the original data set used for model training.

In one embodiment, when obtaining the initial production data, there will be some data that will not affect the actual process, such as a certain type of data in which the value of all data is 1 (a fixed value), Obviously, this type of data will not affect the overall SMT printing, so this type of data needs to be filtered out.

Exemplarily, the obtained initial production data includes manpower, machine, material, method, environment, measurement, collection of SMT production line printing stage and SPI detection all collectable data, and the original data obtained after data filtering The set includes PCB board attribute parameters, printing process parameters, printing process parameter data and SPI detection data, etc. For data or factors that are not these attributes, it may be data that needs to be filtered out. Taking the solder paste printing process of the SMT production line as an example, after obtaining the relevant initial production data, the corresponding data filtering and screening will be carried out to remove some characteristic factors that have no influence on the SMT process, and then obtain the The original data set, and the original data set obtained at this time can be shown in Table 1 below:

Table 1

It should be noted that when screening data, the data that needs to be filtered can be manually screened, or can be automatically screened. Screening can be achieved by setting corresponding thresholds, and then it is determined which factor or factors can be filtered and screened out through the set thresholds. Specifically, the filtering of initial production data includes:

Determine the characteristics contained in the initial production data, and determine the characteristic change value of the characteristic; obtain the preset change value, and filter the data contained in the initial production data according to the characteristic change value and the preset change value; wherein, the characteristic change value The first feature that is less than the preset change value is deleted; the second feature whose feature change value is greater than or equal to the preset change value is used as the original data set.

Therefore, when filtering the initial production data, determine the features contained in the initial production data, where the feature is a data category, and then summarize and analyze all the values corresponding to the feature to determine the feature change value corresponding to each feature. At the same time, the pre-set change value is obtained, and then the initial production data is filtered by comparing the characteristic change value with the change preset change value, and at the same time, it is determined which or which features/factors need to be filtered out, wherein, The first feature whose feature change value is greater than or equal to the preset change value is deleted, and the second feature whose feature change value is greater than or equal to the preset change value is used as the original data set, so as to use the obtained original data set to train the optimization model.

When determining the feature change value corresponding to a certain feature, it can be determined by calculating the variance corresponding to this feature, and then using a set threshold to determine whether it is the data that needs to be filtered and screened out. Under normal circumstances, in order to ensure the accuracy of model training, when filtering the initial production data, the same type of data is filtered out. Then when setting the preset change value, the preset change value to zero.

Step S202, using feature crossover and principal component analysis to reconstruct the influencing factors of the original data set, and obtain a reconstructed influencing factor data package.

Among them, the principal component analysis FI-PCA of feature intersection is used to reconstruct the features of the data, and at the same time, after the feature reconstruction is completed, the dimensionality reduction processing is performed on the relevant data of the reconstructed features.

After obtaining the original data set after filtering and screening, use feature crossover and principal component analysis to realize the reconstruction of the influencing factors of the original data set, specifically, first perform feature crossover, and then perform principal component analysis to obtain Influencing factor data package for model training.

In practical applications, the specific process of reconstruction of influencing factors includes: first, data specification is performed on the original data set, and the influence of data dimension on the reconstruction of influencing factors is eliminated; Crossover, realize the characteristic crossover of all influencing factors, complete the crossover between influencing factors, and then conduct principal component analysis on the obtained new influencing factors to determine the principal components, and then realize dimensionality reduction to obtain reconstructed influencing factors, and finally combine the obtained new influencing factors The obtained reconstructed influencing factors are combined with the process parameters to be optimized to obtain the data package of influencing factors in the printing stage of the SMT production line.

Referring to FIG. 3 , FIG. 3 is a schematic flowchart of the steps of obtaining the influencing factor data package provided by another embodiment of the present application.

Wherein, the process includes step S301 to step S304.

Step S301. Perform data reduction on the original data set to obtain a reduced original data set.

When the original data set is obtained, since the data recorded in the original data set is recorded according to the actual data, there will be a huge difference between the data values, which will not only affect the reconstruction of the influencing factors, but also It has an impact on subsequent model training, such as reducing the speed of model training. Therefore, reasonable processing of data can reduce the impact of data differences and improve the efficiency of subsequent model training.

When reducing the data recorded in the original data set, the data is actually normalized to reduce the data between (0, 1), so as to eliminate the influence of dimension on data mining speed and model convergence speed. influences.

The calculation formula for normalization reduction is as follows:

Among them, x ^* is the value obtained by normalization, x is the actual value corresponding to a feature before normalization, x _max is the maximum value of this type of feature, and x _min is the minimum value of this type of feature value.

Step S302. Determine the process parameters to be optimized corresponding to the optimization model to be trained, and remove the process parameters to be optimized contained in the reduced original data set to obtain a first data set.

The initial production data received during model training contains process parameters that need to be optimized, and it is necessary to determine the impact of other parameters on the optimized process parameters in addition to the optimized process parameters during model training, Therefore, after data reduction, for the original data set obtained after reduction, it is necessary to extract the process parameters to be optimized contained in it, and then process the characteristic data of other parts that do not contain the process parameters to be optimized.

Therefore, when the original data set after reduction is obtained, the process parameters to be optimized corresponding to the optimization model currently being trained are determined, and then the process parameters corresponding to the process parameters to be optimized contained in the original data set after data reduction are completed The data are eliminated to obtain the first data set.

Exemplary, for example, when the optimization model optimizes and recommends process parameters, the process parameters include: scraper pressure, scraper speed, demoulding speed, demoulding distance, automatic cleaning count and cleaning speed, then the original data after the protocol When the data set is eliminated, the data corresponding to the six features are eliminated, and the data set obtained after data elimination is the first data set, which is used for training and optimizing the optimization model to be trained.

At this time, after performing normalized data reduction processing and removing process parameters to be optimized based on the original data set obtained in Table 1, the original data set obtained at this time can be shown in Table 2 below:

Table 2

0

1

2

...

26

27

28

-0.200753	0.423355	0.583249	...	0.731365	0.211944	-0.267736
-0.047214	0.425975	0.585725	...	0.756390	0.273257	-0.267736
-0.091925	0.431216	0.590678	...	0.756829	0.297846	-0.267736
...	...	...	...	...	...	...
0.052419	0.231809	-1.241435	...	0.760561	0.224869	-0.267736

Step S303, performing feature cross multiplication on the influencing factors included in the first data set to obtain a second data set.

After completing the data elimination of the reduced original data set to obtain the first data set, the data contained in the obtained first data set will be processed, specifically, when the first data set is obtained, the first data set will be processed. The influencing factors contained in a data set are subjected to feature cross multiplication, and then the corresponding second data set is obtained after the processing is completed.

In one embodiment, after the first data set is obtained, the data contained in the first data set is processed by feature intersection, wherein the feature intersection is mined by performing functional operations on the original features of the sample data. The hidden information features between the original features, and for the printing stage of the SMT production line, in addition to the process parameters, there are small elements, printing process parameters, etc., the physical and logical relationship between these parameter features is unclear, and the number is huge. Hiding features allows the model to learn more complex features, thereby improving the accuracy of the algorithm.

Exemplarily, when performing feature crossing, a new feature is obtained by crossing features with features, where the actual calculation formula of feature crossing is as follows:

x _i,j = x _i × x _j

Among them, x _i represents the i-th data of the feature, x _j is the j-th feature of the feature, and x _{i, j} is the product interaction result of x _i and x _j .

After performing feature crossover on the original data set after reduction and data elimination in this way, the corresponding second data set is obtained, and the second data set obtained at this time can be shown in Table 3 below:

table 3

0	1	2	...	26	27	28
0.040302	-0.084990	-0.117089	...	0.044920	-0.056745	0.071683
0.002229	-0.020112	-0.027655	...	0.074670	-0.073161	0.071683
0.008450	-0.039640	-0.054298	...	0.088712	-0.079744	0.071683
...	...	...	...	...	...	...
0.008414	-0.040275	-0.054863	...	0.044920	-0.056745	0.071683

Step S304 , using principal component analysis to analyze each influencing factor included in the second data set to obtain an influencing factor data package.

Principal Component Analysis (PCA for short) is a linear dimensionality reduction method, which can effectively eliminate the correlation between features and improve the expressive ability of features. After the original features are crossed, although more hidden features are extracted, the number of crossed features generated is huge, and it is necessary to reduce the dimensionality and replace the original features with a small number of unrelated features to improve the speed of model training.

Therefore, after the feature intersection of the first data set is completed, principal component analysis is used to analyze the obtained second data set, so as to obtain the corresponding influencing factor data package after the analysis process is completed. Specifically, when analyzing each influencing factor contained in the second data set, determine the degree of influence of each influencing factor on the actual SMT process, so as to determine a number of influencing factors in the second data set as having a greater impact on the SMT process. The reconstructed influencing factors with great influence, and after obtaining the reconstructed influencing factors, combine the reconstructed influencing factors with the process parameters to be optimized and specific parameters to obtain the influencing factor data package.

Exemplarily, the specific parameters include solder paste volume and area, so the characteristics or parameters contained in the obtained influencing factor data package at this time include: reconstructing influencing factors, process parameters to be optimized (squeegee pressure, scraper speed, demoulding speed, ejection distance, automatic cleaning count, and cleaning speed), and solder paste volume and area.

Referring to FIG. 4 , FIG. 4 is a schematic flowchart of the steps of obtaining the influencing factor data package provided by another embodiment of the present application. Wherein, when the principal component analysis is used to analyze and process the second data set to obtain the influencing factor data package, it includes:

Step S401, using principal component analysis to calculate the eigenvalues corresponding to each influencing factor in the second data set, and sorting the eigenvalues;

Step S402, according to the selection method from large to small, select several eigenvalues from the sorted eigenvalues to add, and when the added value is greater than the set preset threshold for the first time, determine the number of eigenvalues to be added. The features corresponding to the eigenvalues are reconstruction influencing factors;

Step S403, combining the reconstructed influencing factors and the process parameters to be optimized to obtain an influencing factor data package.

When using principal component analysis to analyze and process the second data set, first calculate the eigenvalues corresponding to each influencing factor in the second data set according to principal component analysis, wherein each influencing factor in the second data set is not reconstructed The final influencing factors, and sort the obtained eigenvalues, such as sorting from large to small or small to large, and after the sorting is completed, select and add the eigenvalues according to the selection method from large to small, At the same time, compare the value obtained by the addition with the preset threshold to obtain the corresponding reconstruction influencing factors in the second data set. Finally, after obtaining the reconstruction influencing factors, compare the reconstruction influencing factors with the process parameters to be optimized. Merge the combination to get the impact factor data package.

In practical applications, different influencing factors have different influences on the actual SMT process, and feature crossover is to extract the influencing factors at a deeper level, and to improve model training by mining hidden features is more accurate. After the feature crossover is completed, due to the huge amount of data after the crossover, corresponding processing is required, such as dimensionality reduction processing, to extract more characteristic part of the data as input data for model training.

For each influencing factor in the obtained second data set, it has a certain influence on the model, but the degree of influence is different. Feature crossover is to improve the accuracy of model training, and to ensure model training The efficiency makes it impossible to use all the data obtained by feature crossover as the input of model training, so it is necessary to further judge all the influencing factors contained in the second data set to select more appropriate influencing factors as model training input of.

When calculating the eigenvalue corresponding to each influencing factor in the second data set, it is to determine the influence degree of each influencing factor on the actual process in the SMT process, and the larger the eigenvalue, the greater the influence, so the calculated For the eigenvalues corresponding to each influencing factor, the eigenvalues are sorted and then selected to obtain the reconstructed influencing factors. When selecting the influencing factors, select according to the size of the actual eigenvalues from large to small, and at the same time sum the obtained eigenvalues to obtain the summation result, and then select the corresponding influencing factors according to the obtained summation value as Refactoring factors.

Exemplarily, when using principal component analysis to process the second data, the specific processing process is as follows:

When the second data set Z={z ₁ ,z ₂ ,…,z _p } is obtained, the second data set is centralized to eliminate the impact of the dimension difference between features, and the centralization formula is as follows Shown:

Among them, X is the data set after the centralization of the second data set,

is the average of the data contained in the dataset.

After the centering of the second data set is completed, the covariance matrix of the data set X obtained by centering is calculated, and the calculation formula of the covariance matrix is as follows:

After the covariance matrix is obtained, the eigenvalue decomposition is performed on the covariance matrix, and all the eigenvalues are arranged in descending order from large to small, which is recorded as {λ ₁ ,λ ₂ ,…,λ _p }, and the corresponding eigenvector is {ω ₁ ,ω ₂ ,…,ω _p }, where λ is the eigenvalue, ω is the eigenvector, and then select the eigenvector corresponding to the first d largest eigenvalue for mapping, and map p-dimensional data to d-dimensional data, specifically The mapping method is as follows:

In the formula,

Represents the data after mapping.

And the parameter d is determined by the proportion of information, and the calculation method of the proportion of information is as follows:

Among them, η is the proportion of information.

Step S103, loading the optimization model to be trained.

In one embodiment, when constructing the optimized model to be trained, the constructed optimized model to be trained is a deep neural network model with multiple outputs, and when the optimized model to be trained is trained, it is the hidden layer of the deep neural network with multiple outputs The number of nodes and the activation function are optimized, and finally multiple groups of optimized multi-output deep neural network prediction models are obtained.

Among them, the deep neural network is a fully connected neural network, mainly composed of input layer, hidden layer and output layer, which has better fitting ability and prediction effect on massive data than traditional machine learning models. The hidden layer nodes and activation functions of deep neural networks are often set according to empirical formulas, and different combinations of hidden layer nodes and activation functions will affect the accuracy of deep neural networks. It should be noted that, for the loaded optimized model to be trained, in addition to the multi-output deep neural network model constructed based on the deep neural network, it may also be constructed based on network models of other structures.

In fact, the optimized model to be trained is a model used to predict and recommend relevant process parameters that need to be used in the SMT printing process, and can also judge the quality of the combination of different input process parameters. The optimized model to be trained is obtained based on the deep neural network. After the construction is completed, the model is trained and optimized, and several better nodes are selected in the multi-node optimized model to be trained for integration to obtain training. Good optimization model.

Step S104: Train the optimization model to be trained according to the influencing factor data package to obtain a trained optimization model.

After the optimized model to be trained is loaded, the optimized model to be trained is trained and optimized using the influencing factor data package obtained in advance, and finally the converged optimized model is recorded and stored when the optimized model to be trained is determined to converge , that is, the trained optimization model is obtained when the convergence is determined.

In one embodiment, when the model to be trained and optimized is trained, an elite-retaining genetic algorithm is used to realize the training and optimization of the model. Genetic algorithm is a random global search and optimization method developed by imitating the biological evolution mechanism in nature, drawing on Darwin's theory of evolution and Mendel's genetic theory. The elite-retaining genetic algorithm improves the traditional genetic algorithm. The best individual that has appeared so far in the evolution process of the group, that is, the elite individual, is directly copied to the next generation without pairing and crossover, and the fitness value of the new generation group is The smallest individual is eliminated, so as to speed up the global convergence to find the best solution.

Referring to FIG. 5, FIG. 5 is a schematic flowchart of the steps of training the optimization model to be trained according to an embodiment of the present application, wherein the steps include steps S501 to S502.

Step S501, initialize the optimization model to be trained, obtain corresponding initialization parameters, and encode the influencing factor data packet according to a preset encoding method to obtain encoded data;

Step S502: Perform several iterations of training on the initialized optimization model to be trained according to the encoded data, so as to adjust the initialization parameters, and obtain a trained optimization model when convergence is determined.

When training and optimizing the model to be trained and optimized, the model parameters or parameter information in the model to be trained and optimized are optimized by using the elite-reserving genetic algorithm, and then the training and optimization of the model are completed at the time of convergence to obtain the training optimization The final optimized model is then available for users to use.

When optimizing and training the model, first initialize the optimized model to be trained, that is, initialize the relevant parameters and parameter information of the model, and at the same time record the corresponding initialization parameters during initialization, where the initialization parameters are also related to the model. Parameter information, and when the model is initialized, the data contained in the obtained influencing factor data package will be encoded according to the preset encoding method to obtain the corresponding encoded data, and then the obtained encoded data will be used The optimization model to be trained is trained and optimized to adjust the initialization parameters, and finally the trained optimization model is obtained and stored and recorded when the convergence is determined.

In one embodiment, when the optimization model to be trained is trained, it is implemented by using the elite-retaining genetic algorithm, so several iterations of training are performed during the actual training, and each training will update the relevant parameters of the model Adjust and optimize.

Referring to FIG. 6 , FIG. 6 is a schematic flowchart of the steps of iterative training provided by an embodiment of the present application. Wherein this step includes step S601 to step S603.

Step S601, input the encoded data into the initialized optimization model to be trained, so as to perform the first training on the initialized optimization model to be trained, and obtain the first intermediate parameters after the first training;

Step S602, adjusting the optimization model to be trained based on the first intermediate parameters to obtain a first intermediate optimization model to be trained;

Step S603, input the encoded data into the first intermediate optimization model to be trained, so as to perform a second training on the optimization model to be trained, and obtain second intermediate parameters after the second training, so as to By analogy, the optimization model to be trained is trained several times.

When performing iterative training, the initial parameters of the model to be trained and optimized are adjusted and optimized according to the influencing factor data package. Specifically, after obtaining the encoded data corresponding to the influencing factor data package, the encoded data is input into the initialized model to be trained In the optimization model, the optimized model after initialization is trained for the first time, and the first intermediate parameter after the first training is obtained when the first training is completed, and the optimized model to be trained is adjusted according to the first intermediate parameter, Obtain the first intermediate optimized model to be trained, and then input the coded data into the first intermediate optimized model to be trained, so that the optimized model to be trained is trained for the second time, and so on to complete several trainings of the optimized model to be trained.

When training the optimized model to be trained, it is determined that the training is completed when it converges, and there are many conditions for judging the convergence, such as setting a number of training times, and for example, one or some parameters in the model meet the preset conditions.

When the training of the model is completed, the optimized model obtained when the training is completed will be recorded. Specifically, referring to FIG. 7, when the optimized model obtained and stored is obtained and stored, it includes:

Step S701, when determining the convergence, sort the objective function values obtained in each training, so as to select N objective function values from the sorted objective function values according to the selection method from large to small, wherein the encoded data input into the optimization model to be trained to obtain the objective function value, and output an objective function value for each training;

Step S702, determining N sets of parameter information corresponding to the N objective function values, wherein the parameter information includes the number of hidden layer network nodes and corresponding activation functions;

Step S703, input the N sets of parameter information into the N sub-models in the optimization model to be trained, and integrate the N sub-models to obtain a trained optimization model, so as to carry out the trained optimization model storage.

When the trained optimization model is obtained, the obtained optimization model is a neural network model with multiple outputs, so when the convergence is determined, the objective function values obtained during each training are sorted, and then according to The method selects N objective function values from the sorted objective function values, and then determines the parameter information corresponding to the selected N objective function values, where the parameter information includes the number of hidden layer network nodes and the activation function corresponding to the hidden layer , and finally determine N sub-models in the optimization model according to the obtained N sets of parameter information, and integrate the determined N sub-models to obtain a trained optimization model.

Exemplarily, when initializing, set the encoding method of the elite reserved genetic algorithm to RI (real number encoding), and also use binary encoding and gray encoding, which are determined according to actual usage requirements, and set the population size to 200 at the same time to optimize training The maximum number of iterations is 300, the mutation probability is 0.05, and the crossover probability is 0.95. The number of nodes in hidden layers 1 and 2 is optimized in the range [10,128], and the activation function optimization space is [relu, sigmoid, tanh].

In an embodiment, when acquiring the above-mentioned objective function, it is first necessary to determine the calculation method of the objective function, and then calculate the corresponding objective function value for each training according to relevant data information. The objective function can be as follows:

Among them, y _{area, i} is the actual value of the solder paste volume, y _{vol, i} is the actual value of the area of the solder paste,

is the sample predicted value of the volume,

is the sample prediction value of the area, and m is the sample size.

For example, when 300 iterative training is completed, 300 objective function values will be obtained, and each objective function value will correspond to a parameter information, and then the obtained 300 objective function values will be sorted according to the size, and then the sorting is completed Then select 5 objective function values from large to small, and determine the hidden layer network points corresponding to the 5 objective function values and the corresponding activation functions. The number of objective function values is determined according to actual needs. This sets N to be 5. When the sorting and selection are completed, the obtained deep neural network structure and objective function value can be shown in Table 4 below:

Table 4

Number of hidden layer 1 nodes	Number of hidden layer 2 nodes	Activation function 1	Activation function 2	forecast error
36	112	2	0	0.1308

117	68	2	0	0.1325
35	126	2	0	0.1402
65	110	2	0	0.1442
36	109	2	0	0.1465

Among them, the activation function represented by 0 in the activation function is relu, 1 represents sigmoid, 2 represents tanh, and the prediction error is the average relative percentage error. After obtaining the parameter information shown in the above table, the model will be adjusted according to the parameter information to obtain 5 deep neural network models, and then the obtained 5 deep neural network models will be integrated to obtain the final optimization Model. When performing model integration, stacking stack integration is performed on the obtained five neural network models, and then an optimized model of the integrated deep neural network is obtained.

In an embodiment, refer to FIG. 8 . FIG. 8 is a schematic flowchart of the steps of predicting and recommending process parameters provided by an embodiment of the present application. Wherein this step includes: step S801 to step S804.

Step S801, when receiving the parameter optimization instruction, load the trained optimization model, and receive the input process parameters to be analyzed;

Step S802, input the process parameters to be analyzed into the optimization model, and calculate the function value based on the process parameters to be analyzed;

Step S803, when the function value is greater than a preset threshold, output the predicted process parameters according to the optimization model.

After the model training and optimization are completed, the stored optimized model can be used for quality prediction of parameter quality and recommendation of process parameters.

When receiving a parameter optimization instruction, load the trained optimization model, and receive the input process parameters to be analyzed, and then input the received process parameters to be analyzed into the loaded optimization model to calculate the to-be-analyzed The function value corresponding to the process parameter, and then compare the obtained function value with the preset threshold value to determine whether the process parameter to be analyzed is qualified, wherein when it is determined that the obtained function value is greater than the set preset threshold value , indicating that the process parameters to be analyzed received at this time are not qualified process parameter combinations. At this time, the predicted process parameters will be output according to the loaded optimization model to obtain a more reasonable process parameter combination. After determining the obtained function When the value is less than or equal to the preset threshold, it is determined that the setting of the process parameter is qualified, and at this time, a qualified prompt message will be output.

In one embodiment, when the model is used after training, a new objective function is set in the model, where the new objective function is:

g(·)＝|vol-100|+|area-100|

Among them, vol means volume, area means area, 100 means perfect volume and area, and |·| means absolute value.

After receiving the input process parameters to be analyzed, the volume and area predicted by the model will be output, and then brought into the objective function to determine whether the currently received process parameters to be analyzed are qualified. When it is determined to be qualified, it will directly output qualified instruction information, and when it is determined to be unqualified, it will predict the process parameters. When predicting and recommending process parameters, set the optimization interval of the process parameters to be optimized, optimize the objective function through the elite-retaining genetic algorithm, obtain the optimal solution, and obtain the recommended process parameter combination for recommendation and feedback.

Exemplarily, when recommending a combination of process parameters, several sets of recommended process parameter combinations for recommendation are shown in Table 5 below:

table 5

Blade pressure	Squeegee speed	Release speed	Release distance	Automatic cleaning count	cleaning speed	objective function
5.34	63.21	2.40	1.13	12	43.04	0.193
11.85	50.70	2.93	0.13	30	36.68	2.594
8.48	47.64	2.95	0.13	27	33.55	3.002

It can be seen from the above table that according to the value of the objective function, it can be determined that when the process parameters scraper pressure is 5.34, scraper speed is 63.21, demolding speed is 2.40, automatic cleaning speed is 12 and cleaning speed is 43.04, it has a better effect and also That is, the process parameter corresponding to the combination is the optimal process parameter combination at this time.

In the training method, equipment and storage medium of the SMT printing parameter optimization model described above, when training, the training sample data for training is firstly reconstructed to the influencing factors, specifically, the relationship between the influencing factors is established through feature crossover. Then, principal component analysis is performed on the constructed features to complete the reconstruction of influencing factors, and the influencing factors are deeply excavated, and then different deep neural network models are integrated to obtain the final optimization model. It solves the problem of poor prediction effect of a single fixed model, greatly improves the accuracy of the model, and finally searches out the combination of process parameters for the best printing quality through the elite-retaining genetic algorithm, which has better applicability.

Referring to FIG. 9 , FIG. 9 is a schematic structural block diagram of a computer device provided by an embodiment of the present application.

Exemplarily, the device may be a tablet computer, a notebook, or a desktop computer.

The device also includes a processor, memory for storing a computer program.

The processor is configured to execute the computer program and implement any one of the SMT printing parameter optimization model training methods provided in the embodiments of the present application when executing the computer program.

It should be understood that the processor may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP for short), application-specific integrated circuits (Application Specific Integrated Circuit, referred to as: ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, referred to as: FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Wherein, the general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

Embodiments of the present application also provide a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor implements the The training method of the SMT printing parameter optimization model described in any one.

Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, the functional modules/units in the system, and the device can be implemented as software, firmware, hardware, and an appropriate combination thereof. In hardware embodiments, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be composed of several physical components. Components cooperate to execute. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit . Such software may be distributed on computer-readable storage media, which may include computer-readable storage media (or non-transitory media) and communication media (or transitory media).

As known to those of ordinary skill in the art, the term computer-readable storage medium includes both volatile and non-volatile media implemented in any method or technology for storage of information, such as computer-readable instructions, data structures, program modules, or other data. Volatile, removable and non-removable media. Computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cartridges, tape, magnetic disk storage or other magnetic storage devices, Or any other medium that can be used to store desired information and that can be accessed by a computer. In addition, as is well known to those of ordinary skill in the art, communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

Exemplarily, the computer-readable storage medium may be an internal storage unit of the electronic device described in the foregoing embodiments, such as a hard disk or a memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk equipped on the electronic device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD ) card, flash memory card (Flash Card), etc.

The electronic device and the computer-readable storage medium provided by the foregoing embodiments display at least two virtual keyboards in different display areas on the display screen when the user inputs information, so that information can be input through at least two virtual keyboards; The software infers the difficulty of input information by monitoring the state of the sensor, which enhances the security of information input.

The above is only a specific embodiment of the application, but the scope of protection of the application is not limited thereto. Any person familiar with the technical field can easily think of various equivalents within the scope of the technology disclosed in the application. Modifications or replacements, these modifications or replacements shall be covered within the scope of protection of this application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims (10)

1. A kind of training method of SMT printing parameter optimization model, described method comprises the following steps:

   receive initial production data;

   Reconstructing the influencing factors according to the initial production data to obtain the influencing factor data package;

   Load the optimized model to be trained;

   The optimization model to be trained is trained according to the influencing factor data package to obtain a trained optimization model.
2. The method according to claim 1, wherein said reconstruction of influencing factors according to said initial production data to obtain an influencing factor data package includes:

   Filtering the initial production data to obtain an original data set not included in the fixed attribute, wherein the data corresponding to the fixed attribute being an influencing factor is a fixed value;

   Reconstructing the influencing factors of the original data set by using feature crossover and principal component analysis to obtain a reconstructed influencing factor data package.
3. The method according to claim 2, wherein said utilizing feature crossover and principal component analysis to reconstruct the influencing factors of said original data set to obtain a reconstructed influencing factor data package, comprising:

   performing data reduction on the original data set to obtain a reduced original data set;

   Determining the process parameters to be optimized corresponding to the parameter to be trained optimization model, and removing the process parameters to be optimized contained in the original data set to obtain the first data set;

   performing feature cross multiplication on the influencing factors contained in the first data set to obtain a second data set;

   Each influencing factor included in the second data set is analyzed by principal component analysis to obtain an influencing factor data package.
4. The method according to claim 3, wherein said analysis of each influencing factor contained in said second data set by using principal component analysis to obtain an influencing factor data package includes:

   Using principal component analysis to calculate the eigenvalues corresponding to each influencing factor in the second data set, and sorting the eigenvalues;

   According to the selection method from large to small, select a number of eigenvalues from the sorted eigenvalues to add, and when the added value is greater than the set preset threshold for the first time, determine the eigenvalues to be added. The corresponding features are reconstruction influencing factors;

   Merge the reconstructed influencing factors with the process parameters to be optimized to obtain an influencing factor data package.
5. The method according to any one of claims 1 to 4, wherein the training of the optimization model to be trained according to the influencing factor data package to obtain a trained optimization model includes:

   Initializing the optimization model to be trained, obtaining corresponding initialization parameters, and encoding the influencing factor data packet according to a preset encoding method to obtain encoded data;

   Perform several iterations of training on the initialized optimization model to be trained according to the encoded data, so as to adjust the initialization parameters, and obtain a trained optimization model when convergence is determined.
6. The method according to claim 5, wherein said performing several iterations of training on said initialized optimization model to be trained according to said coded data comprises:

   Inputting the encoded data into the initialized optimization model to be trained, so as to perform the first training on the initialized optimization model to be trained, and obtain the first intermediate parameters after the first training;

   adjusting the optimization model to be trained based on the first intermediate parameters to obtain a first intermediate optimization model to be trained;

   inputting the encoded data into the first intermediate optimization model to be trained, so as to perform a second training on the first intermediate optimization model to be trained, and obtain second intermediate parameters after the second training, thereby By analogy, the optimization model to be trained is trained several times.
7. The method according to claim 6, wherein said obtaining a trained optimization model when determining convergence comprises:

   When the convergence is determined, the objective function values obtained by each training are sorted, so as to select N objective function values from the sorted objective function values according to the selection mode from large to small, wherein the encoded data is input to the The objective function value is obtained in the optimized model to be trained, and an objective function value is output for each training;

   Determining N sets of parameter information corresponding to the N objective function values, wherein the parameter information includes the number of hidden layer network nodes and the corresponding activation function;

   The N sets of parameter information are input into the N sub-models in the optimization model to be trained, and the N sub-models are integrated to obtain a trained optimization model, wherein the optimization model is based on a deep neural network. Constructed multi-output deep neural network model.
8. The method according to any one of claims 1 to 7, wherein, after said training the optimization model to be trained according to the influencing factor data package, and obtaining the trained optimization model, further comprising:

   When the parameter optimization instruction is received, the trained optimization model is loaded, and the input process parameters to be analyzed are received;

   inputting the process parameter to be analyzed into the optimization model, and calculating a function value based on the process parameter to be analyzed;

   When the function value is greater than a preset threshold, the predicted process parameters are output according to the optimization model.
9. A computer device comprising a memory and a processor;

   The memory is used to store computer programs;

   The processor is configured to execute the computer program and realize the steps of the method for training the SMT printing parameter optimization model according to any one of claims 1 to 8 when the computer program is executed.
10. A storage medium for computer-readable storage, the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement any one of claims 1 to 8 A step in the training method of the SMT printing parameter optimization model.

Images