A reflow solder joint morphology prediction method

Abstract

The application discloses a reflow soldering solder joint morphology prediction method and belongs to the technical field of reflow soldering solder joint morphology detection. The method is characterized in that a reflow soldering process is modeled by a machine learning method, the final result is quickly calculated by mathematical modeling, the experimental time of experimenters is saved, and engineering and technical personnel are helped to understand and master the connection and rules between different material data; the prediction accuracy is high, the correlation between sample input and output is found by data mining on a large number of samples, the generalization ability of the prediction model is enhanced by a fusion algorithm, and the robustness of the prediction model is ensured; the method can be used for solder joint morphology prediction of other welding methods, reflow soldering is taken as a training object, and in fact, other welding methods can also be used as training objects to model a welding process, and the final solder joint morphology parameters are quickly obtained.

Description

Technical Field

[0001] This invention relates to the field of reflow solder joint morphology detection technology, and specifically to a method for predicting the morphology of reflow solder joints. Background Technology

[0002] With the widespread application of Surface Mount Technology (SMT), reflow soldering technology, an important component of SMT, is also constantly evolving. Reflow soldering involves heating gas to a certain temperature to melt solder paste, which is then used to solder components onto a PCB (Printed Circuit Board). For reflow soldering technology, solder joint morphology serves as an evaluation criterion and is crucial for its inspection.

[0003] Currently, traditional methods are mostly used for inspecting the solder joint morphology of reflow soldering. While these methods offer high accuracy, they require experimental testing, resulting in long testing cycles and low material utilization. The existing technologies can be summarized as follows: First, the experimental cycle is long. Whether through computer simulation or actual experiments, process engineers need to spend a significant amount of time testing suitable parameters to obtain the ideal solder joint morphology. Second, material utilization is low, leading to high economic costs. Each experiment wastes a large amount of experimental materials, resulting in high labor and material costs.

[0004] To address the issue of low efficiency in reflow solder joint morphology testing in existing technologies, and the lack of an effective solution, a reflow solder joint morphology prediction method is proposed. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to solve the problem of low efficiency in reflow solder joint morphology testing in the prior art, and to provide a method for predicting the morphology of reflow solder joints.

[0006] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0007] S1: Data Preparation

[0008] Data acquisition is performed on the morphology of reflow solder joints, collecting relevant information, inputting production input parameters, and outputting solder joint morphology information, using feature vectors (x... i ,y i The position of the i-th sample in feature space O is represented by the form ), where x i Let y be the input space for the i-th group of reflow solder joint morphology samples. i For the output space of the i-th group of reflow solder joint morphology samples;

[0009] S2: Data Cleaning

[0010] The data collected in step S1 is cleaned and processed.

[0011] S3: Normalization

[0012] The samples processed in step S2 are normalized using the Min-Max standardization method;

[0013] S4: Data Dimensionality Reduction

[0014] The samples processed in step S3 are input into the prediction model for training. The EFB algorithm is used to bundle the high-dimensional data, changing the data dimension from Data*Feature to Data*Bundle.

[0015] S5: Decision Tree Growth

[0016] The decision tree growth strategy is adopted to complete the decision tree growth;

[0017] S6: Model Training

[0018] The GOSS algorithm is used to train the prediction model, and the prediction accuracy of the prediction model is improved by calculating the information gain.

[0019] S7: Gradient Calculation

[0020] The gradients in the data space are calculated and sorted according to their absolute values.

[0021] S8: Model Optimization

[0022] The model is trained on the training and validation sets to obtain the corresponding model parameters. The output is a regression prediction model trained using the existing training and validation sets, implemented using R. 2 The standard value is used as the evaluation standard for the prediction effect of the regression prediction model. The regression prediction model is optimized based on the standard value of the prediction effect. That is, the optimization direction is based on the standard value of the prediction effect, the termination condition is set, and the regression prediction model is iteratively optimized.

[0023] S9: Calculation of Impact

[0024] After optimization in step S8, a reflow solder joint morphology prediction model is obtained. Through statistical calculation, the degree of influence of production input parameters on the final output parameters is obtained, which serves as the basis for adjusting the input.

[0025] S10: Simulation Experiment

[0026] The prepared production input parameters are input into the reflow solder joint morphology prediction model. The final predicted value is obtained through regression. The predicted value is compared with the ideal solder joint morphology. The production input parameter values ​​are adjusted by adjusting the degree of influence of each production input parameter on the output parameter in step S9 until the ideal solder joint morphology value is obtained. Actual experiments are then conducted.

[0027] Furthermore, in step S2, the data cleaning method is the 3σ criterion. Assuming the data volume is n, the standard deviation σ is calculated using the Bessel formula:

[0028]

[0029] Among them, V i Let x be the data value of the i-th sample. i With the sample mean The deviation, when |V i If |>3σ, the data is considered abnormal and is deleted.

[0030] Furthermore, in step S3, during the normalization process, let the sample value be X, and the maximum and minimum values ​​be Xmax and Xmin, respectively. max and X min Then the processed sample X * for:

[0031]

[0032] Furthermore, in step S4, the prediction model is implemented based on the LightGBM algorithm.

[0033] Furthermore, in step S5, by judging all the leaves, the leaf with the greatest information gain after splitting is found, and then the node of that leaf is split. Through such cyclical splitting, the decision tree growth is finally completed.

[0034] Furthermore, in step S6, let x be the portion of the input space on the dataset. i The feature dimensions of the data space are s and g. i To determine the gradient descent direction of the objective function during each iteration, the GOSS algorithm, after traversing all split nodes, finds the split point k and the splitting feature variable J in the data space. The resulting two datasets are R1 and R2, respectively. The information gain V generated at split point k is then... J|O (d) is:

[0035]

[0036] Furthermore, in step S7, the data points with large gradient magnitudes ('a') are selected from largest to smallest to form dataset A. Then, from the remaining datasets with smaller gradient magnitudes, 'b' data points are randomly selected to form dataset B. When calculating the information gain, the portion with smaller gradient magnitudes is multiplied by a constant. Among them, a larger gradient magnitude means that the gradient magnitude is greater than or equal to the first set threshold, and a smaller gradient magnitude means that the gradient magnitude is less than or equal to the second set threshold, wherein the second set threshold is less than the first set threshold.

[0037] Furthermore, in step S8, R 2 The formula for calculating the value is as follows:

[0038]

[0039] in, For the i-th sample, For the true value of the i-th sample, This represents the average value of the sample.

[0040] Furthermore, in step S10, the regression methods include linear regression, quadratic support vector machine regression, and rational quadratic Gaussian process regression.

[0041] The present invention has the following advantages over the prior art:

[0042] 1. Compared with the traditional reflow solder joint morphology detection method, the reflow solder joint morphology prediction method of the present invention uses machine learning to model the reflow soldering process and quickly calculates the final result through mathematical modeling, saving experimental time for experimental personnel and helping engineering technicians to understand and master the relationship and rules between different material data.

[0043] 2. This invention has high prediction accuracy. By performing data mining on a large number of samples, it finds the correlation between sample input and output, and enhances the generalization ability of the prediction model through fusion algorithms, thus ensuring the robustness of the prediction model.

[0044] 3. This invention can be applied to predict weld joint morphology using other welding methods. While this invention uses reflow soldering as the training object, other welding methods can also be used to model the welding process and quickly obtain the final weld joint morphology parameters. Attached Figure Description

[0045] Figure 1 This is a flowchart of the reflow solder joint morphology prediction method in an embodiment of the present invention;

[0046] Figure 2This is a schematic diagram of the method for collecting reflow solder joint morphology data in an embodiment of the present invention; (wherein, 1 represents the observation lens, 2 represents the solder, 3 represents the tantalum capacitor, H represents the climb height, L1 represents the external length of the connecting plate, and L2 represents the side connection length)

[0047] Figure 3 This is a flowchart of the reflow solder joint morphology prediction model in an embodiment of the present invention;

[0048] Figure 4 This is a flowchart of constructing a histogram in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the Leaf-Wise decision tree growth strategy in an embodiment of the present invention;

[0050] Figure 6a This is a predicted diagram of the solder joint thickness in an embodiment of the present invention;

[0051] Figure 6b This is an absolute error diagram of the solder joint thickness in an embodiment of the present invention;

[0052] Figure 6c This is a predicted diagram of the solder joint width in an embodiment of the present invention;

[0053] Figure 6d This is an absolute error diagram of the solder joint width in an embodiment of the present invention;

[0054] Figure 6e This is a prediction diagram of solder climb-up in an embodiment of the present invention;

[0055] Figure 6f This is an absolute error diagram of solder climb-up in an embodiment of the present invention;

[0056] Figure 7a This is a diagram illustrating the influence of solder joint thickness in an embodiment of the present invention;

[0057] Figure 7b This is a diagram illustrating the influence of solder joint width in an embodiment of the present invention;

[0058] Figure 7c This is a diagram illustrating the influence of solder creep in an embodiment of the present invention. Detailed Implementation

[0059] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0060] Example 1

[0061] like Figure 1As shown, this embodiment provides a technical solution: a method for predicting the morphology of reflow solder joints, comprising the following steps:

[0062] Step 1: Collect data on the morphology of reflow solder joints, including relevant information, input production parameters, and output solder joint morphology information. The feature vector (x...) is used to... i ,y i The position of the i-th sample in feature space O is represented by the form ), where x i Let y be the input space for the i-th group of reflow solder joint morphology samples. i For the output space of the i-th group of reflow solder joint morphology samples;

[0063] Step 2: Perform data cleaning on the data collected in Step 1. The data cleaning method used is the 3σ criterion. Let the data volume be n, and calculate the standard deviation σ using the Bessel formula:

[0064]

[0065] Among them, V i Let x be the data value of the i-th sample. i With the sample mean The deviation, when |V i If |>3σ, it is considered abnormal data and the data is deleted;

[0066] Step 3: Normalize the cleaned samples from Step 2 using Min-Max standardization. Let the sample value be X, and the maximum and minimum values ​​be Xmax and Xmin, respectively. max and X min Then the processed sample X * for:

[0067]

[0068] Step 4: Input the samples into the prediction model for training. Use the EFB (Exclusive Feature Bundle) algorithm to bundle the high-dimensional data. By constructing a histogram, the data dimension is changed from Data*Feature to Data*Bundle. Since the amount of data in the bundle after bundling is much smaller than the amount of data in the feature, the training efficiency is improved. Here, Data represents a single data point, Feature represents a feature, and Bundle represents the bundled feature.

[0069] Step 5: Adopt a leaf-wise decision tree growth strategy. By judging all leaves, find the leaf with the largest information gain after splitting, and then split the node of that leaf. Through this cyclical splitting, the decision tree growth is finally completed. In this way, splitting is done on a single node instead of splitting all nodes at a time greatly reduces the amount of computation and improves computational efficiency.

[0070] Step 6: Train the prediction model using the GOSS (Gradient-based One-Side Sampling) algorithm. Improve the prediction accuracy of the prediction model by calculating information gain. Assume that the input space portion of the dataset is x. i The feature dimensions of the data space are s and g. i To determine the gradient descent direction of the objective function during each iteration, the GOSS algorithm, after traversing all split nodes, finds the split point k and the splitting feature variable J in the data space. The resulting two datasets are R1 and R2, respectively. The information gain V generated at split point k is then... J|O (d) is:

[0071]

[0072] Step 7: Based on Step 6, calculate the gradient in the data space and sort them according to their absolute values. Select the data points with the largest gradient magnitudes ('a') from largest to smallest to form dataset A. Then, randomly select 'b' data points from the remaining datasets with smaller gradient magnitudes to form dataset B. To prevent the original distribution of the data space from being disrupted, multiply the smaller gradient magnitudes by a constant when calculating the information gain. 'a' and 'b' are predefined numbers of data points. The selection method involves taking 'a' data points from the dataset in descending order to form the larger dataset, and then randomly selecting 'b' data points from the remaining dataset. Once the optimal split point 'k' is selected... * Its corresponding characteristic variable J * Calculate the maximum information gain at that point. for:

[0073]

[0074] Step 8: The prediction model uses the LightGBM algorithm as its foundation, trains on the training and validation sets, obtains the corresponding model parameters, and outputs a regression prediction model trained on the existing training and validation sets. This model is then processed using R-Square (i.e., R...). 2The standard value is used as a criterion for evaluating the predictive performance of a regression prediction model. The regression prediction model is optimized based on the standard value of the predictive performance. That is, the optimization direction is set with the standard value of the predictive performance, and the termination condition is set to iteratively optimize the regression prediction model. This evaluation index assesses the proportion of the variance of the prediction model for the variable. Its value ranges between [0,1]. The closer it is to 1, the more accurate the prediction model is. The closer it is to 0, the less valuable the prediction model is.

[0075] R 2 The formula for calculating the value is as follows:

[0076]

[0077] in, For the i-th sample, For the true value of the i-th sample, This represents the average value of the sample.

[0078] Step 9: After optimization in Step 8, the reflow solder joint morphology prediction model (i.e., the optimal regression prediction model) is obtained. Through statistical calculation, the degree of influence of production input parameters on the final output parameters is obtained, which serves as the basis for adjusting the input.

[0079] Step 10: The experimenter inputs the prepared production input parameters into the reflow solder joint morphology prediction model, obtains the final predicted value through regression, compares the predicted value with the ideal solder joint morphology, and adjusts the production input parameter values ​​by adjusting the degree of influence of each production input parameter on the output parameter in step 9 until the ideal solder joint morphology value is obtained, and then conducts the actual experiment.

[0080] Example 2

[0081] This embodiment collects reflow soldering process data and final solder joint morphology parameter data of tantalum capacitors on an SMT production line to verify the method in Embodiment 1, including the following steps:

[0082] 1. Data Preparation. This embodiment collects 2000 data points, including tantalum capacitor reflow soldering process data and final solder joint morphology parameters from an SMT production line. These include input process parameters (production input parameters) such as peak temperature, cooling rate, and solder paste thickness, and output solder joint morphology parameters (output solder joint morphology information) such as solder joint thickness, solder joint width, and solder climb-up. Each data point is organized into a spatial direction (X, Y), and the entire dataset is a vector space S.

[0083] 2. Data Cleaning. Data cleaning is performed on 2000 data points. Assuming the data volume n = 2000, the standard deviation σ is calculated using the Bessel formula:

[0084]

[0085] Among them, V i Let x be the data value of the i-th sample. i With the sample mean The deviation, when |V i When |>3σ, it is judged as abnormal data and deleted; after cleaning the data, 1700 data points are retained as the dataset for training the prediction model.

[0086] 3. Normalization. The Min-Max standard normalization method is used. Let the sample value be X, and the maximum and minimum values ​​be Xmax and Xmin, respectively. max and X min Then the processed sample X * for:

[0087]

[0088] 4. Dimensionality Reduction. The samples are input into the prediction model for training. The EFB algorithm is used to bundle high-dimensional data, and a histogram is constructed to reduce the data dimension from Data*Feature to Data*Bundle.

[0089] 5. Decision tree growth. A leaf-wise decision tree growth strategy is employed, such as... Figure 5 As shown, by judging all the leaves, the leaf with the largest information gain after splitting is found, and then the node of that leaf is split. Through such a cyclical splitting, the decision tree growth is finally completed. In this way, splitting is done on a single node instead of splitting all nodes at a time greatly reduces the amount of computation and improves the computational efficiency.

[0090] 6. Model Training. The GOSS algorithm is used to train the prediction model. Information gain is calculated to improve the prediction accuracy. Assume the input space portion of the dataset is x. i The feature dimensions of the data space are s and g. i To determine the gradient descent direction of the objective function during each iteration, the GOSS algorithm, after traversing all split nodes, finds the split point k and the splitting feature variable J in the data space. The resulting two datasets are R1 and R2, respectively. The information gain V generated at split point k is then... J|O (d) is:

[0091]

[0092] 7. Gradient Calculation. Calculate the gradient across the data space and sort the data by absolute value. Select the 'a' data points with the largest gradient magnitudes from largest to smallest to form dataset A. Then, randomly select 'b' data points from the remaining datasets with smaller gradient magnitudes to form dataset B. To prevent the original distribution of the data space from being disrupted, multiply the smaller gradient magnitudes by a constant when calculating the information gain. When the optimal split point k is selected * Its corresponding characteristic variable J * Calculate the maximum information gain at that point. for:

[0093]

[0094] 8. Model Performance Evaluation. The model uses the LightGBM algorithm as its foundation, trains on the training and validation sets, obtains the corresponding model parameters, and outputs a regression prediction model trained on the existing training and validation sets. R-Square (i.e., R...) is used for evaluation. 2 The ) value is used as a standard for evaluating the prediction effect of the prediction model. This evaluation index assesses the proportion of the variance of the prediction model for the variable. Its value ranges between [0,1]. The closer it is to 1, the more accurate the prediction model is. The closer it is to 0, the less valuable the prediction model is. The final results of the prediction effects of solder joint thickness, solder joint width and solder climb are 0.87802, 0.88462 and 0.86216, respectively.

[0095]

[0096] in, For the i-th sample, For the true value of the i-th sample, This represents the average value of the sample.

[0097] 9. Calculate the degree of influence. Based on the above steps, output a reflow solder joint morphology prediction model. Through statistical calculations, obtain the degree of influence of production input parameters on the final output parameters, such as... Figure 7a , 7b As shown in 7c, this serves as the basis for adjusting the input.

[0098] 10. Simulation Experiment. The experimenters input the prepared production input parameters into the reflow solder joint morphology prediction model, obtain the final predicted value through regression, compare the predicted value with the ideal solder joint morphology, and adjust the production input parameter values ​​according to the degree of influence of the production input parameters on the output parameters until the ideal solder joint morphology value is obtained. The actual experiment is then conducted, reducing the originally planned number of experiments to 1 / 10, and suggestions for parameter modification are provided.

[0099] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the morphology of reflow solder joints, characterized in that, Includes the following steps: S1: Data Preparation Data acquisition is performed on the morphology of reflow solder joints, collecting relevant information, inputting production input parameters, and outputting solder joint morphology information, using feature vectors. The form of the first Group of samples in feature space The location, among which For the first Input space for reflow solder joint morphology samples. For the first Output space of reflow solder joint morphology samples; S2: Data Cleaning The data collected in step S1 is cleaned and processed. S3: Normalization The samples processed in step S2 are normalized using the Min-Max standardization method; S4: Data Dimensionality Reduction The samples processed in step S3 are input into the prediction model for training. The EFB algorithm is used to bundle high-dimensional data, reducing the data dimensionality from... Become ,in, Represents a single data point. Indicates features, Indicates the characteristics after binding; S5: Decision Tree Growth The decision tree growth strategy is adopted to complete the decision tree growth; S6: Model Training The GOSS algorithm is used to train the prediction model, and the prediction accuracy of the prediction model is improved by calculating the information gain. S7: Gradient Calculation Calculate the gradient in the data space and sort them according to their absolute values; S8: Model Optimization The model is trained on the training and validation sets to obtain the corresponding model parameters, and outputs a regression prediction model trained based on the existing training and validation sets. The standard value is used as the evaluation standard for the prediction effect of the regression prediction model. The regression prediction model is optimized based on the standard value of the prediction effect. That is, the optimization direction is based on the standard value of the prediction effect, the termination condition is set, and the regression prediction model is iteratively optimized. S9: Calculation of Impact After optimization in step S8, a reflow solder joint morphology prediction model is obtained. Through statistical calculation, the degree of influence of production input parameters on the final output parameters is obtained, which serves as the basis for adjusting the input. S10: Simulation Experiment The prepared production input parameters are input into the reflow solder joint morphology prediction model. The final predicted value is obtained through regression. The predicted value is compared with the ideal solder joint morphology. The production input parameter values ​​are adjusted by adjusting the degree of influence of each production input parameter on the output parameter in step S9 until the ideal solder joint morphology value is obtained. Actual experiments are then conducted. In step S4, the prediction model is implemented based on the LightGBM algorithm; In step S5, by judging all the leaves, the leaf with the largest information gain after splitting is found, and then the node of that leaf is split. Through such cyclic splitting, the decision tree growth is finally completed. In step S6, let the input space portion of the dataset be... The feature dimension of the data space is , To determine the gradient descent direction of the objective function during each iteration, the GOSS algorithm searches for the split point in the data space after traversing all split nodes. and segmentation of feature variables At that time, the two datasets that were split were respectively Then at the dividing point Information gain generated for: ； In step S7, the gradient magnitude is larger. The dataset is composed of data points selected from largest to smallest. Then, select from the remaining datasets with smaller gradient magnitudes using a random selection method. The dataset consists of several data points. When calculating information gain, the portion of the gradient with a smaller magnitude is multiplied by a constant. In this context, a larger gradient magnitude means that the gradient magnitude is greater than or equal to a first set threshold, while a smaller gradient magnitude means that the gradient magnitude is less than or equal to a second set threshold, where the second set threshold is less than the first set threshold.

2. The method for predicting the morphology of reflow solder joints according to claim 1, characterized in that: In step S2, the data cleaning method is as follows: Criteria, assuming the data volume is... The standard deviation is calculated using the Bessel formula. : ； in, For the first Data values ​​of each sample With the sample mean The deviation, when If the data is deemed abnormal, it will be deleted.

3. The method for predicting the morphology of reflow solder joints according to claim 1, characterized in that: In step S3, during the normalization process, the sample value is set to... The maximum and minimum values ​​are respectively and The processed sample for: 。 4. The method for predicting the morphology of reflow solder joints according to claim 1, characterized in that: In step S8, The formula for calculating the value is as follows: ； in, For the first Predicted values ​​for each sample For the first The true value of each sample This represents the average value of the sample.

5. The method for predicting the morphology of reflow solder joints according to claim 1, characterized in that: In step S10, the regression methods include linear regression, quadratic support vector machine regression, and rational quadratic Gaussian process regression.

Images