A PCB solder paste printing quality prediction method based on temporal reconstruction graph convolution

Abstract

The present invention discloses a PCB solder paste printing quality prediction method based on time reconstruction graph convolution, comprising obtaining historical data of solder paste volume detection sequences and corresponding detection times of a PCB board and performing preprocessing; establishing a time reconstruction module to reconstruct the historical data according to the time intervals of the corresponding detection times to obtain reconstructed data; establishing a spatiotemporal graph convolution module to extract the correlation between the solder paste volumes of the obtained reconstructed data and the variation pattern of the solder paste volumes themselves; constructing a linear regressor based on the correlation between the solder paste volumes and the variation pattern of the solder paste volumes themselves to predict the variation pattern of the solder paste volumes. The present invention constructs a sparse graph based on the temporal correlation of the solder paste volumes on the pad to improve the coordination of similar evolution patterns while avoiding interference from different evolution patterns, and then constructs a temporal attention based on the intervals of production moments to reconstruct the data and predict the PCB solder paste printing quality.

Description

Technical Field

[0001] The present invention relates to the technical field of PCB solder paste printing quality prediction, and in particular to a PCB solder paste printing quality prediction method based on time reconstruction graph convolution. Background Art

[0002] Due to their wide range of uses, PCBs are processed in huge quantities each year. The vast majority of defective products can be traced back to the solder paste printing process. Therefore, it is necessary to predict PCB solder paste printing quality so that relevant parameters of the solder paste printer can be adjusted in advance to improve yield.

[0003] The difficulties in improving solder paste printing quality with existing technologies are: first, PCB motherboards are usually produced at unequal time intervals, and timing prediction is closely related to time intervals. The solder paste volume evolution law learned based on the equal interval method may be distorted; second, the solder paste volume evolution laws of different pads are different. The solder paste volume evolution laws with smaller differences have similarities and can be predicted collaboratively, but the solder paste volume evolution laws with larger differences may interfere with each other, making it difficult to use a unified model for prediction. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology. To achieve the above purpose, a PCB solder paste printing quality prediction method based on time reconstruction graph convolution is adopted to solve the problems raised in the above background technology.

[0005] A PCB solder paste printing quality prediction method based on time reconstruction graph convolution, the specific steps include:

[0006] Step S1: Obtain historical data of solder paste volume detection sequences and corresponding detection times of PCB boards, and perform preprocessing;

[0007] Step S2: Establish a time reconstruction module to reconstruct historical data according to the time interval of the corresponding detection time to obtain reconstructed data;

[0008] Step S3: Establish a spatiotemporal graph convolution module to extract the correlation between the solder paste volumes of the reconstructed data and the change pattern of the solder paste volume itself;

[0009] Step S4: construct a linear regressor based on the correlation between the solder paste volumes and the change pattern of the solder paste volumes themselves, and predict the change pattern of the solder paste volumes.

[0010] As a further technical solution of the present invention: the specific steps in step S1 include:

[0011] Step S11: Obtain the historical solder paste volume detection sequence of the PCB board and the corresponding detection time, and record the solder paste volume of the PCB board detected at time t as N is the number of solder paste volume tests, and the i-th solder paste volume test result is recorded as i=1,2,…,N, and get the historical solder paste volume detection sequence n is the number of PCB boards;

[0012] Step S12: Use the sliding window method to divide the historical solder paste volume detection sequence S, where the historical window size is K, the translation step size is s, the prediction window size is T, and the blocking threshold is α; the historical window solder paste volume detection sequence is obtained as The prediction window solder paste volume detection sequence is

[0013] Step S13: Construct weight label matrix Shifting the sequence by the prediction window Mark the defective solder paste printing position at any time and any position, where the corresponding defective position of the weight label matrix L is marked as 1, otherwise it is 0;

[0014] Step S14: Constructing an adjacency matrix The correlation of solder paste volume is calculated based on the Pearson correlation coefficient, which is formulated as follows:

[0015]

[0016] in, From time t1 to time t n The set of solder paste volumes at time i, yes The mean of yes The variance of i,j is the Pearson correlation coefficient between the i-th solder paste volume and the j-th solder paste volume;

[0017] Based on the threshold function T λ (c) Filter the correlation. When the correlation c i,j When it is less than λ, the two solder pastes are correlated, and negatively correlated or uncorrelated interference is eliminated. The threshold function T λ (c) is formulated as follows:

[0018]

[0019] And thus construct the adjacency matrix:

[0020]

[0021] As a further technical solution of the present invention: the specific steps in step S2 include:

[0022] Step S21: normalize the obtained production moment of the detection time:

[0023] When the PCB board is produced, the monitoring sample is obtained, the production time is obtained, and the production time vector is set And normalize the vector t, the formula is as follows:

[0024]

[0025] Get the normalized production time vector

[0026] S22: construct temporal attention according to the obtained production moment, and perform weighted reconstruction on the historical data based on the constructed temporal attention;

[0027] The normalized real production time vector coding:

[0028]

[0029]

[0030] Then calculate the self-attention matrix of production time, which is formulated as follows:

[0031]

[0032] Then use the softmax constraint on Att to convert the attention scores between solder pastes into a probability distribution between [0, 1]:

[0033]

[0034] Finally, the obtained self-attention matrix is ​​used to analyze the i-th solder paste from time t1 to time t N Feature set at a moment To refactor:

[0035]

[0036] in, Represents the reconstructed sequential solder paste printing features, and the reconstructed features are assembled into graph G′ t =(R′ t ,A), As input to the spatiotemporal graph convolution.

[0037] As a further technical solution of the present invention: the specific steps in step S3 include:

[0038] Step S31: Construct a spatiotemporal graph convolution module to capture the correlation between the evolution laws of solder pastes and the volume change law of a single solder paste, where each spatiotemporal graph convolution module includes two Gated CNNs, one GCN, and one BatchNormalization layer;

[0039] Step S32: Using the Gated CNN, learn the volume change pattern of a single solder paste, which is formulated as follows:

[0040]

[0041] Among them, W c1 and W c2 is the convolution kernel, b c1 and b c2 is the bias, they are all trainable parameters, σ is the nonlinear activation function sigmoid, is the element-wise product of matrices, represents the input of the k-th layer of Gated CNN, represents the output of the k-th layer of Gated CNN,

[0042] Step S33: Using the GCN, the correlation of volume change patterns between solder pastes is mined, and the formula is as follows:

[0043]

[0044] Among them, g θ It's the core. is the graph convolution operation, θ0 is a trainable parameter, represents the k-th layer GCN input, is the k-th layer GCN output, is the identity matrix, D ii =∑ j A ij is the degree matrix.

[0045] As a further technical solution of the present invention: the specific steps in step S4 include:

[0046] Step S41: Based on the output of the last layer of Gated CNN After the fully connected layer, the final output is:

[0047]

[0048] Step S42: weighted mean absolute error loss based on design Reducing falsely low error indicators:

[0049]

[0050] in, is the i-th solder paste quality category label at time t, the abnormal label is 1, and the normal label is 0. is the volume of the ith solder paste at time t, is the corresponding predicted value, ω is the category weight of the abnormal sample;

[0051] The overall loss function is obtained according to the weighted mean absolute error loss as follows:

[0052]

[0053] Among them, θ is the model parameter and μ is the hyperparameter.

[0054] Compared with the prior art, the present invention has the following technical effects:

[0055] Using this technical solution, a spatiotemporal graph neural network based on time reconstruction is first used to construct a sparse graph based on the temporal correlation of solder paste volume on pads, enhancing the coordination of similar evolution patterns while avoiding interference from different evolution patterns. Temporal attention is then used to reconstruct the data based on production intervals. This overcomes the shortcomings of existing methods by constructing a sparse graph based on temporal correlations and then reconstructing the raw data using temporal attention designed based on production intervals, thereby predicting PCB solder paste printing quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings:

[0057] Figure 1 A schematic diagram of the method steps of the embodiment disclosed in this application;

[0058] Figure 2 This is a flowchart of the method disclosed in the embodiment of this application. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] Please refer to Figure 1 and Figure 2 In an embodiment of the present invention, a method for predicting PCB solder paste printing quality based on time reconstruction graph convolution includes the following specific steps:

[0061] Step S1: Obtain the historical data of the solder paste volume detection sequence and the corresponding detection time of the PCB board and perform preprocessing. The specific steps include:

[0062] Step S11: Collect the historical PCB board solder paste volume detection sequence and the corresponding detection time, and record the PCB board solder paste volume detected at time t as N is the number of solder paste volume tests, and the i-th solder paste volume test result is recorded as i=1,2,…,n. Finally, the historical solder paste volume detection sequence is obtained n is the number of PCB boards.

[0063] Step S12: Use the sliding window method to divide the historical solder paste volume detection sequence S, with the historical window size being K, the translation step size being s, the prediction window size being T, and the blocking threshold being α. Thus, the historical window solder paste volume detection sequence is obtained. Prediction window solder paste volume detection sequence

[0064] Step S13: Construct weight label matrix Since defective solder paste printing is a very important issue with a low probability, it is important to focus on printing defects. If there is a defective tin or a defective tin at any time and any position, the corresponding position in the weight label matrix L is marked as 1, otherwise it is 0.

[0065] Step S14: Constructing an adjacency matrix To calculate the correlation of solder paste volume, we use the Pearson correlation coefficient. The formula is as follows:

[0066]

[0067] From time t1 to time t n The set of solder paste volumes at time i. yes The mean of yes The variance of c i,j is the Pearson correlation coefficient between the i-th solder paste volume and the j-th solder paste volume.

[0068] The threshold function T is introduced λ (c) Filter the correlation. When the correlation c i,j When it is less than λ, the two solder pastes are not unrelated, thus avoiding negative correlation or no correlation interference. λ (c) is formalized as follows:

[0069]

[0070] The adjacency matrix is ​​constructed from this

[0071] Step S2: Establish a time reconstruction module to reconstruct historical data according to the corresponding detection time interval to obtain reconstructed data. The specific steps include:

[0072] Step S21: normalize the obtained production moment of the detection time:

[0073] Unlike regular sampling of maintenance targets by sensors, monitoring samples can only be obtained when PCBs are produced, so the sampling intervals are irregular. Let the real production time vector First, the vector t is normalized and the formula is as follows:

[0074]

[0075] Get the normalized production time vector

[0076] Step S22: constructing time attention according to the obtained production moment, and performing weighted reconstruction on the historical data based on the constructed time attention;

[0077] In order to overcome the distortion of solder paste volume timing prediction caused by irregular production time intervals, the production moment is used to construct temporal attention, and the original data is weightedly reconstructed according to the temporal attention.

[0078] The normalized real production time vector coding:

[0079]

[0080]

[0081] Then calculate the self-attention matrix of production time, which is formalized as follows:

[0082]

[0083] Then, we use softmax to constrain Att and convert the attention scores between solder pastes into a probability distribution between [0, 1], while highlighting the relationship between solder pastes.

[0084]

[0085] Finally, the obtained attention matrix is ​​used to analyze the i-th solder paste from time t1 to time t N Feature set at a moment Perform reconstruction.

[0086]

[0087] in, Represents the reconstructed sequential solder paste printing features, and the reconstructed features are assembled into graph G′ t =(R′ t ,A), As input to the spatiotemporal graph convolution.

[0088] Step S3: Establish a spatiotemporal graph convolution module to extract the correlation between the solder paste volumes of the reconstructed data and the change pattern of the solder paste volume itself. The specific steps include:

[0089] Step S31: For solder paste volume prediction, the PCBs at each production moment are represented as a graph, with continuously produced PCBs forming a spatiotemporal graph. Collaborative prediction of solder pastes with similar volume variation patterns can mitigate the adverse effects of random fluctuations in the volume of individual solder pastes. Pads with significantly different solder paste volume variation patterns are not correlated, preventing mutual interference. Therefore, a spatiotemporal graph convolution module is constructed to capture the correlation between the evolution patterns of solder pastes and the volume variation patterns of individual solder pastes.

[0090] Each spatiotemporal graph convolution module consists of two Gated CNNs, a GCN, and a Batch Normalization layer.

[0091] Step S32: Gated CNN learns the volume change pattern of a single solder paste, which is formulated as follows:

[0092]

[0093] Among them, w c1 and W c2 is the convolution kernel, b c1 and b c2 is the bias, they are all trainable parameters, σ is the nonlinear activation function sigmoid, is the element-wise product of matrices, represents the input of the k-th layer of Gated CNN, represents the output of the k-th layer of Gated CNN,

[0094] Step S33: GCN is used to explore the correlation of volume change patterns between solder pastes, which is formulated as follows:

[0095]

[0096] Among them, g θ It's the core. is the graph convolution operation, θ0 is a trainable parameter, represents the k-th layer GCN input, is the k-th layer GCN output, is the identity matrix, D ii =∑ j A ij is the degree matrix.

[0097] Step S4: construct a linear regressor based on the correlation between solder paste volumes and the variation pattern of solder paste volumes themselves to predict the variation pattern of solder paste volumes. The specific steps include:

[0098] Step S41: Based on the output of the last layer of Gated CNN After the fully connected layer, the final output is:

[0099]

[0100] Step S42: The goal of our task is to accurately predict the volume of solder paste to detect products with abnormal printing quality in advance. Considering that the number of samples with abnormal solder paste volume is far less than that of normal samples, the weight of abnormal samples must be increased to prevent false low error indicators. Here, a weighted mean absolute error loss is designed.

[0101]

[0102] is the i-th solder paste quality category label at time t, the abnormal label is 1, and the normal label is 0. is the volume of the ith solder paste at time t, is the corresponding predicted value, and ω is the category weight of the abnormal sample.

[0103] The overall loss function is as follows:

[0104]

[0105] θ is a model parameter and μ is a hyperparameter.

[0106] Step S43: predict the change pattern of solder paste volume to obtain the prediction result of PCB solder paste printing quality.

[0107] Although the embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents, and all should be included within the scope of protection of the present invention.

Claims

1. A PCB solder paste printing quality prediction method based on time reconstruction graph convolution, characterized in that: The specific steps include: Step S1: Obtain historical data of solder paste volume detection sequences and corresponding detection times of PCB boards, and perform preprocessing; Step S2: Establish a time reconstruction module to reconstruct historical data according to the corresponding detection time interval to obtain reconstructed data. The specific steps include: Step S21: normalize the obtained production moment of the detection time: When the PCB board is produced, the monitoring sample is obtained, the production time is obtained, and the production time vector is set , and for the vector Normalize and formulate as follows: ； Get the normalized production time vector ; S22: construct temporal attention according to the obtained production moment, and perform weighted reconstruction on the historical data based on the constructed temporal attention; The normalized real production time vector coding: ； ； Then calculate the self-attention matrix of production time, which is formulated as follows: ； Then use the softmax pair Constraint, convert the attention score between solder pastes into a probability distribution between [0, 1]: ； Finally, the obtained self-attention matrix is ​​used to analyze the i-th solder paste from Time has come To refactor: ； in, , Represents the reconstructed sequential solder paste printing features, and the reconstructed features are assembled into a graph , As input to spatiotemporal graph convolution; Step S3: Establish a spatiotemporal graph convolution module to extract the correlation between the solder paste volumes of the reconstructed data and the change pattern of the solder paste volume itself; Step S4: construct a linear regressor based on the correlation between the solder paste volumes and the change pattern of the solder paste volumes themselves, and predict the change pattern of the solder paste volumes.

2. The PCB solder paste printing quality prediction method based on time reconstruction graph convolution according to claim 1 is characterized in that: The specific steps in step S1 include: Step S11: Obtain the historical solder paste volume detection sequence and corresponding detection time of the PCB board, and The volume of solder paste on the PCB board detected at any time is recorded as , N is the number of solder paste volume tests, and the i-th solder paste volume test result is recorded as , , and obtain the historical solder paste volume detection sequence , is the number of PCB boards; Step S12: Use the sliding window method to detect the volume of the historical solder paste Divide, where the history window size is K, each translation step is s, the prediction window size is T, and the congestion threshold is ; The historical window solder paste volume detection sequence is obtained as , the prediction window solder paste volume detection sequence is ; Step S13: Construct weight label matrix Shifting the sequence by the prediction window Mark the bad solder paste printing position at any time and any position, where the weight label matrix The corresponding bad position is marked as 1, otherwise it is 0; Step S14: Constructing an adjacency matrix , the correlation of solder paste volume is calculated based on the Pearson correlation coefficient, which is formulated as follows: ； in, Indicates from Time has come The set of solder paste volumes at time i, yes The mean of yes The variance of is the Pearson correlation coefficient between the i-th solder paste volume and the j-th solder paste volume; Based on threshold function Filter the correlation. Less than When , the two solder pastes are correlated, and the negatively correlated or uncorrelated interference is eliminated. The threshold function The formula is as follows: ； And thus construct the adjacency matrix: 。 3. The PCB solder paste printing quality prediction method based on time reconstruction graph convolution according to claim 1 is characterized in that: The specific steps in step S3 include: Step S31: Construct a spatiotemporal graph convolution module to capture the correlation between the evolution laws of solder pastes and the volume change law of a single solder paste, where each spatiotemporal graph convolution module includes two Gated CNNs, one GCN, and one BatchNormalization layer; Step S32: Using the Gated CNN, learn the volume change pattern of a single solder paste, which is formulated as follows: ； in, and is the convolution kernel, and is the bias, they are all trainable parameters, is the nonlinear activation function sigmoid, is the element-wise product of matrices, represents the input of the k-th layer of Gated CNN, represents the output of the k-th layer of GatedCNN, ; Step S33: Using the GCN, the correlation of volume change patterns between solder pastes is mined, and the formula is as follows: ； in, It's the core. is the graph convolution operation, is a trainable parameter, represents the k-th layer GCN input, is the k-th layer GCN output, is the identity matrix, is the degree matrix.

4. The PCB solder paste printing quality prediction method based on time reconstruction graph convolution according to claim 1 is characterized in that: The specific steps in step S4 include: Step S41: Based on the output of the last layer of Gated CNN , after passing through the fully connected layer, the final output is: ； Step S42: weighted mean absolute error loss based on design , reducing false low error metrics: ； in, is the i-th solder paste quality category label at time t, the abnormal label is 1, and the normal label is 0. is the volume of the ith solder paste at time t, is the corresponding predicted value, is the category weight of abnormal samples; The overall loss function is obtained according to the weighted mean absolute error loss as follows: ； in, are model parameters, is a hyperparameter.

Images