A method for predicting multi-tin anomalies of PCB under strong noise

Abstract

The application relates to the field of artificial intelligence, in particular to a method for predicting multiple tin abnormalities of a PCB under strong noise. In the model training stage, training data is sent into an expanded causal convolution subnetwork and a soft threshold subnetwork to extract features and remove noise signals, so that a soft thresholded solder paste printing feature vector is obtained; a mixed attention subnetwork is used to perform attention weighting on the features in the channel and space to obtain a solder paste printing weighted spatial feature vector. The two kinds of features are fused and input into a time series modeling subnetwork to evaluate the influence of historical state information on the current state and model time series dependency information. The processed feature information is input into a full connection layer to obtain a multiple tin abnormality prediction result of the PCB solder paste printing. The application effectively processes noise signals in an actual scene and models the space-time dependency information of sensor data, thereby improving the accuracy of multiple tin abnormality prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to a method for predicting PCB board multi-tin abnormalities under strong noise. BACKGROUND

[0002] Surface Mount Technology (SMT) is a method of installing electronic components on a Printed Circuit Board (PCB). This technology mainly uses soldering technology to directly mount electronic components on the surface of the PCB board. The SMT production line generally includes tin paste printing, patching, reflow soldering and other main processes, and is the main production line in the electronic manufacturing industry. PCB board tin paste printing is an important step in the SMT production line, which prints tin paste on the electronic component welding position of the PCB board to provide soldering material for the subsequent component mounting. In this process, the printing quality of the tin paste directly affects the quality of component welding, so it is a key link to determine the performance of the production line.

[0003] PCB board multi-tin abnormality refers to the printing amount of tin paste exceeding the specified standard during the tin paste printing process, resulting in excessive tin paste on the PCB board. This situation may cause short circuit, false welding or cold welding problems in the subsequent welding process, and further cause the performance of electronic components to decline or even fail.

[0004] Solving the problem of PCB board multi-tin abnormality is of great significance to ensure the normal production of the SMT production line. Since multi-tin abnormality is one of the key factors leading to the failure of the SMT production line, it accounts for 60-70% of the failure rate of the entire production line, therefore, effectively detecting and preventing multi-tin abnormality can significantly improve the stability and yield of the production line, reduce production costs, and improve product quality, thereby having an important impact on the production efficiency and product quality of the entire electronic manufacturing industry. Currently, the main problems facing PCB board multi-tin abnormality prediction include large industrial field noise and difficulty in effectively modeling the spatiotemporal dependence information of sensor data. In the past, more tin paste printing detection equipment (AOI) was used to make post-judgment on the quality of PCB board tin paste printing. However, this method cannot predict the time of multi-tin abnormality occurrence in advance, which may cause unplanned downtime accidents and seriously affect normal production.

[0005] With the rapid development of deep learning technology in the field of predictive maintenance, many deep learning methods have been proposed to solve the problem of abnormality prediction. However, in the tin paste printing process of the SMT production line, the noise in the actual scene is large, and the spatiotemporal dependence information of the sensor data is difficult to effectively model, so the existing deep learning methods are prone to failure. SUMMARY

[0006] To solve the above problems, the application provides a method for predicting multiple tin abnormalities of a PCB under strong noise.

[0007] The method comprises:

[0008] Step one, collect the data of the tin paste printing quality of the PCB, and after pretreatment, use it as a training set;

[0009] Step two, perform feature processing on the training set, increase the receptive field and modeling timing information, remove related noise, and learn context information to obtain a tin paste printing fusion feature vector;

[0010] Step three, input the tin paste printing fusion feature vector as historical state information into a timing attention module to obtain a tin paste printing state containing context information at time step t

[0011] Step four, input the tin paste printing state containing context information at time step t into a reset gate r t and an update gate z t for processing to obtain a tin paste printing candidate hidden state c' at time step t t ;

[0012] Step five, calculate the tin paste printing hidden state h t at time step t based on the tin paste printing candidate hidden state c' at time step t t ;

[0013] Step six, input the tin paste printing hidden state h t at time step t into a fully connected layer to predict the maximum tin paste volume in all pads of the PCB at the next time step;

[0014] Step seven, compare the ratio of the predicted maximum tin paste volume in all pads of the PCB to the standard volume with a standard threshold value to obtain a multiple tin abnormality prediction result of the tin paste printing of the PCB.

[0015] Further, step two specifically comprises:

[0016] perform feature processing on the training set through a convolution layer to obtain a tin paste printing feature vector;

[0017] input the tin paste printing feature vector into a dilated causal convolution subnetwork to obtain a tin paste printing extended timing feature;

[0018] input the tin paste printing extended timing feature into a soft threshold subnetwork to obtain a soft thresholded tin paste printing feature vector;

[0019] input the tin paste printing feature vector into a hybrid attention subnetwork to obtain a tin paste printing weighted spatial feature vector;

[0020] The soft-thresholded solder paste printing feature vector is fused with the solder paste printing weighted space feature vector to obtain a solder paste printing fusion feature vector.

[0021] Further, step three specifically includes:

[0022] The solder paste printing fusion feature vector is input into a time sequence attention module as historical state information, and the historical state information h′ of the PCB board solder paste printing data is obtained. t , t∈[1, n], n represents the time step of the solder paste printing feature vector, the influence size α′ of each historical state on the current state is obtained through the time sequence attention module, an intermediate vector i′ after weighting is calculated, and the current time feature vector after processing by the time sequence attention module is defined as the solder paste printing state containing context information at time step t. t t

[0023] Further, the reset gate r in step four specifically includes: t

[0024] The reset gate r is: t

[0025]

[0026] Wherein, W r , U r represent weight matrices, b r represents a bias matrix, sigmoid represents an activation function, and h t-1 is the solder paste hidden state at time step t-1.

[0027] Further, the update gate z in step four specifically includes: t

[0028] The update gate z is: t

[0029]

[0030] Wherein, W z、 Uz represents a weight matrix, b z represents a bias matrix, and h t-1 is the solder paste hidden state at time step t-1.

[0031] Further, step four specifically includes:

[0032] The solder paste printing candidate hidden state c′ at time step t is: t

[0033] ​​​​​​​

[0034] wherein tanh represents an activation function, W c and U c both represent weight matrices, • represents a matrix multiplication operation, b c represents a bias matrix, h t-1 is the solder paste printing hidden state of time step t-1.

[0035] Further, step five specifically comprises:

[0036] the solder paste printing hidden state h t of time step t is:

[0037] h t = z t · h t-1 + (1-z t )· c t ';

[0038] wherein h t-1 is the solder paste printing hidden state of time step t-1.

[0039] Further, the standard threshold value in step seven is 290%.

[0040] Further, step seven specifically refers to:

[0041] determining whether the ratio of the maximum solder paste volume in all pads of the predicted PCB to the standard volume exceeds the standard threshold value, if yes, the PCB has a solder paste abnormality, if not, the PCB does not have a solder paste abnormality.

[0042] The system in the application corresponds to the method, and the specific preferred scheme of the method is also applicable to the system.

[0043] The one or more technical solutions provided in the embodiments of the application have at least the following technical effects or advantages:

[0044] The application effectively processes the noise signal in the actual scene through the expanded causal convolution subnetwork and the soft threshold subnetwork, and effectively models the space-time dependent information of the sensor data through the hybrid attention subnetwork and the time series modeling subnetwork, thereby improving the accuracy of the solder paste abnormality prediction. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 a system framework diagram of the solder paste abnormality prediction network provided by the embodiments of the application;

[0046] Figure 2 an ASATCN-ABGRU model and an ASATCN module of the solder paste abnormality prediction network coding module provided by the embodiments of the application;

[0047] Figure 3 An internal structure diagram of a channel attention module provided for an embodiment of the present application is shown in the figure;

[0048] Figure 4 An internal structure diagram of a spatial attention module provided for an embodiment of the present application is shown in the figure;

[0049] Figure 5 An internal structure diagram of a time sequence attention module provided for an embodiment of the present application is shown in the figure;

[0050] Figure 6 An internal structure diagram of a gated recurrent unit module provided for an embodiment of the present application is shown in the figure;

[0051] Figure 7 A prediction result of the percentage of the predicted maximum volume of the PCB board solder paste to the standard value provided for an embodiment of the present application. DETAILED DESCRIPTION

[0052] The present application is described in detail below in combination with the drawings and specific embodiments. Before the technical solutions of each embodiment of the present application are described in detail, the terms and names involved are explained. In this specification, components with the same name or the same reference number represent similar or identical structures, and are limited to the purpose of illustration only.

[0053] The present application obtains a multi-solder anomaly prediction result of PCB board solder paste printing by constructing a multi-solder anomaly prediction network. The multi-solder anomaly prediction network is a deep learning model containing a dilated causal convolution subnetwork, a soft threshold subnetwork, a hybrid attention subnetwork, a time sequence modeling subnetwork, and a fully connected layer, as shown in Figure 1 .

[0054] In the model training phase, the training data is input into the dilated causal convolution subnetwork and the soft threshold subnetwork to extract features and remove noise signals, obtaining a soft thresholded solder paste printing feature vector; at the same time, the hybrid attention subnetwork is used to perform attention weighting on the features in the channel and space to obtain a solder paste printing weighted spatial feature vector. Next, the soft thresholded solder paste printing feature vector and the solder paste printing weighted spatial feature vector are fused to obtain a solder paste printing fusion feature vector.

[0055] The solder paste printing fusion feature vector is input into the time sequence modeling subnetwork to evaluate the influence of historical state information on the current state and model the time sequence dependency information. Then, the processed feature information is input into the fully connected layer to obtain a multi-solder anomaly prediction result of PCB board solder paste printing.

[0056] The mean square error (MSE) between the prediction result and the true result is calculated as a loss function, and then the model parameters are optimized by back propagation and gradient descent method to improve the prediction accuracy.

[0057] The present application combines various deep learning techniques, including convolutional neural networks, attention mechanisms, and time series modeling, to achieve accurate prediction of multiple tin abnormalities in PCB solder paste printing.

[0058] The present application will be explained below in conjunction with a specific embodiment, and the specific steps are as follows:

[0059] 1. Data preparation

[0060] Data of PCB solder paste printing quality is collected from an industrial site. After preprocessing including normalization, the collected sensor data set is divided into a training set and a test set according to the 4:1 principle.

[0061] This embodiment takes five days of data collection as an example, and the PCB solder paste printing data of one day is taken as the test set, while the data of the other four days is taken as the training set to train the parameters of the prediction model.

[0062] 2. Data encoding

[0063] The training set is input into the ASATCN-ABGRU model proposed for PCB multiple tin abnormality prediction, as shown in the left half of Figure 2 First, the feature vector of the solder paste printing is obtained by feature processing through the convolutional layer. In the convolutional layer, the training set first extracts multi-scale information of the original data through one-dimensional convolution (Conv 1D), then reduces the model complexity through one-dimensional maximum pooling (MaxPooling 1D), and reduces the risk of model overfitting through the dropout block (Dropout 1).

[0064] As shown in the right half of Figure 2 The solder paste printing feature vector is input into the ASATCN module to obtain the solder paste printing extended time series feature by expanding the receptive field and modeling longer sequence time series information through the dilated causal convolution subnetwork. Specifically, the dilated causal convolution subnetwork here includes batch normalization (BN), activation function (LeakyReLU / ReLU), dropout block (Dropout2), and dilated causal convolution block (Dilated Causal Conv1D). Among them, batch normalization (BN) is used to speed up the convergence speed during model training, and also plays a certain degree of regularization role; the role of the activation function is to provide the network with nonlinear modeling capability; the dropout block is used to reduce the risk of model overfitting; the role of the dilated causal convolution block is to increase the receptive field of the model to model longer time series dependencies.

[0065] The solder paste printing extended time series feature is input into the soft threshold subnetwork, as shown in Figure 2The tin paste printing extension timing feature is averaged, passed through a channel attention module (NAM module), and the importance information of each channel is obtained. Specifically, as shown in Figure 3 In the channel attention module (NAM module), first, the channel scaling factors γ0, γ1,..., γ C are passed through batch normalization processing, and C represents the number of channels. Then, the channel weight information is obtained by multiplying the proportion of each channel scaling factor ω0, ω1,..., ω C , and then the channel attention size is obtained through the nonlinear mapping of the sigmoid activation function. Then, the weighted channel feature vector is obtained by multiplying the tin paste printing extension timing feature. The soft threshold feature vector is obtained by twice subtraction and once maximum value operation. Finally, the soft threshold tin paste printing feature vector is obtained by multiplication operation, so as to obtain the useful feature information of the data in the training set and remove the related noise signal.

[0066] The tin paste printing feature vector is input into the hybrid attention subnetwork, which mainly includes a channel attention module (NAM module) as shown in Figure 3 and a spatial attention module as shown in Figure 4 , which are respectively used to obtain channel features and spatial features within the channel. Specifically, in the spatial attention SA module, first, the input features are processed through the maximum pooling (MaxPool) and average pooling (AvgPool), then the feature vectors obtained by processing are fused, and the feature is extracted through the one-dimensional convolution (Conv1D). Then, the spatial attention size is obtained by using the activation function (sigmoid) for nonlinear mapping, and then the tin paste printing weighted spatial feature vector is obtained by multiplying the input feature.

[0067] The soft threshold tin paste printing feature vector and the tin paste printing weighted spatial feature vector are fused to obtain the tin paste printing fusion feature vector.

[0068] 3. Data decoding

[0069] The tin paste printing fusion feature vector is input into the decoding layer. First, it is input into the timing attention module. The tin paste printing fusion feature vector represents the feature information at different times after encoding processing. These feature information after encoding processing is taken as the historical state information. According to the influence size of the historical information on the current state information in the timing attention module, the current state information at different times after timing weight processing is calculated in turn. As shown in Figure 5 For the historical state information h′ t, t e [1, n], n represents the time step of the solder paste printing feature vector, and the influence of each historical state on the current state is obtained through the time sequence attention module t Then the weighted intermediate vector i' is calculated t The current time feature vector processed by the time sequence attention module can be expressed as the context information containing solder paste printing state at time step t

[0070] The context information containing solder paste printing state is input into the stacked bidirectional gated recurrent unit network (BiGRU) to perform time sequence modeling on the PCB solder paste printing quality data. Specifically, the bidirectional gated recurrent unit network (BiGRU) is obtained by connecting two gated recurrent unit modules (GRU) in front and back, as shown in Figure 6 In the gated recurrent unit module, it mainly includes reset gate r t and update gate z t . The reset gate r t is composed of the context information containing solder paste printing state at time step t and the solder paste printing hidden state h t-1 at time step t-1 after linear transformation and addition, and then the sigmoid activation function, that is, the reset gate r t is:

[0071]

[0072] where W r , U r represent weight matrices, and b r represents a bias matrix. Due to the existence of the sigmoid activation function, the output value is between 0 and 1, which is used to select how much information to leave.

[0073] The update gate z t and the reset gate r t are similar, that is,

[0074]

[0075] where W z , U z represent weight matrices, and b z represents a bias matrix.

[0076] c' t is the candidate hidden state of the solder paste printing at time step t, which is determined by the reset gate r t , the solder paste printing hidden state h t-1 at time step t-1, and the context information containing solder paste printing state at time step t , that is,

[0077]

[0078] wherein tanh denotes an activation function, W c and U c denote weight matrices, · denotes a matrix multiplication operation, and b c denotes a bias matrix.

[0079] The hidden state of the solder paste printing at time step t is h t , which is determined by the candidate hidden state of the solder paste printing at time step t c t ', the update gate z t , the hidden state of the solder paste printing at time step t-1 h t-1 together determine:

[0080] h t = z t · h t-1 + (1-z t )· c t '.

[0081] 4. Multi-solder anomaly prediction

[0082] The hidden state of the solder paste printing at time step t is sent to a dense layer for predicting the maximum solder paste volume in all pads of the PCB at the next time step, and the multi-solder anomaly prediction result of the PCB solder paste printing is obtained by comparing the ratio of the maximum solder paste volume in all pads of the PCB to the standard volume with a standard threshold.

[0083] Figure 7 The prediction result of the percentage of the maximum solder paste volume of the PCB to the standard value obtained based on the present application is shown, and the green line in the figure is the standard threshold of the ratio of the solder paste volume to the standard volume, and the exceeding of this threshold is the multi-solder anomaly of the PCB.

[0084] The standard threshold of the ratio of the solder paste volume to the standard volume in the present embodiment is 290%.

[0085] The loss function between the predicted PCB multi-solder anomaly result and the real PCB multi-solder anomaly result is calculated by the MSE index, and the update of the model parameters is realized by back propagation and gradient descent.

[0086] The above-described embodiments only describe the preferred embodiments of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for predicting abnormal tin content in PCB boards under strong noise, characterized in that: The following steps are involved: Step 1: Collect data on PCB solder paste printing quality and use it as a training set after preprocessing; Step 2: Perform feature processing on the training set to increase the receptive field and model timing information, remove related noise, and learn context information to obtain the solder paste printing fusion feature vector; Step 3: Input the solder paste printing fusion feature vector as the historical state information into the temporal attention module to obtain the time step Solder paste printing status including context information ; Step 4: Set the time step Solder paste printing status including context information Input reset gate and update gate Processing to get the time step Solder paste printing candidate hidden state ; Step 5, based on time step Solder paste printing candidate hidden state Calculate time steps The hidden state of solder paste printing is ; Step 6: Set the time step Solder paste printing hidden state Send it to the fully connected layer to predict the maximum solder paste volume of all pads on the PCB board in the next time step; Step 7: Compare the ratio of the predicted maximum solder paste volume to the standard volume of all pads on the PCB board with the standard threshold to obtain the prediction result of excessive solder paste printing on the PCB board; Step 2 specifically includes: The training set is processed by convolutional layer to obtain solder paste printing feature vector; The solder paste printing feature vector is passed into the dilated causal convolutional sub-network to obtain the solder paste printing extended time series feature; Input the solder paste printing extended time series features into the soft threshold sub-network to obtain the soft thresholded solder paste printing feature vector; Input the solder paste printing feature vector into the hybrid attention sub-network to obtain the solder paste printing weighted spatial feature vector; The soft-thresholded solder paste printing feature vector is fused with the solder paste printing weighted spatial feature vector to obtain the solder paste printing fused feature vector.

2. A method for predicting abnormal tin content in PCB boards under strong noise according to claim 1, characterized in that: Step three specifically includes: The solder paste printing fusion feature vector is input into the temporal attention module as the historical state information. , Represents the time step of the solder paste printing feature vector, and uses the temporal attention module to obtain the impact of each historical state on the current state , calculate the weighted intermediate vector , define the current moment feature vector after processing by the temporal attention module as the time step Solder paste printing status including context information .

3. The method for predicting abnormal tin content in PCBs under strong noise according to claim 1, wherein: Reset the gate as described in step 4 , specifically including: Reset Gate for: ； in, 、 represents the weight matrix, represents the bias matrix, sigmoid represents the activation function, is the time step The solder paste printing is hidden.

4. The method for predicting abnormal tin content in PCBs under strong noise according to claim 1, wherein: Update the gate as described in step 4 , specifically including: Update Gate for: ； in, 、 represents the weight matrix, represents the bias matrix, is the time step The solder paste printing is hidden.

5. The method for predicting abnormal tin content in PCBs under strong noise according to claim 1, wherein: Step 4 specifically includes: Time step Solder paste printing candidate hidden state for: ； in, represents the activation function, and Both represent weight matrices, represents the matrix product operation, represents the bias matrix, is the time step The solder paste printing is hidden.

6. The method for predicting abnormal tin content in PCBs under strong noise according to claim 1, wherein: Step 5 specifically includes: Time step Solder paste printing hidden state for: ； in, is the time step The solder paste printing is hidden.

7. The method for predicting abnormal tin content in a PCB under strong noise according to claim 1, wherein: The standard threshold described in step seven is 290%.

8. The method for predicting abnormal tin content in PCBs under strong noise according to claim 1, wherein: Step 7 specifically refers to: It is determined whether the ratio of the maximum solder paste volume to the standard volume in all pads of the predicted PCB board exceeds the standard threshold. If so, the PCB board has excessive solder paste abnormality; otherwise, the PCB board has no excessive solder paste abnormality.

Images