Surface mount process parameter feedforward adaptive regulation method based on tin paste printing detection

Abstract

The surface mount process parameter feedforward adaptive regulation method based on tin paste printing detection relates to a process parameter adaptive regulation method and belongs to the technical field of electronic assembly process optimization. The method aims to solve the problem that a chip mounter is difficult to adaptively adjust parameters according to real-time working conditions, resulting in poor mounting consistency. The method first constructs a welding offset prediction network model for predicting the position deviation of an element after welding according to tin paste printing offset and process parameters; and then constructs a process parameter compensation model based on deep reinforcement learning to dynamically generate optimal compensation parameters based on the prediction results. In actual production, real-time offset data is obtained through a tin paste detection device to drive the above-mentioned model to adjust key process parameters such as mounting coordinates and height of the chip mounter online, thereby realizing feedforward adaptive compensation. The method can effectively improve the element mounting accuracy and process process capability, and provides an efficient cross-process fusion optimization scheme for precise electronic assembly.

Description

Technical Field

[0001] This invention relates to a method for adaptive control of process parameters, belonging to the field of electronic assembly process optimization technology. Background Technology

[0002] With the increasing miniaturization of electronic components, precision electronic surface mount technology (SMT) faces higher accuracy requirements. To systematically improve the process quality of production lines, modern SMT production lines deeply integrate next-generation information technologies such as cyber-physical systems (CPS) and the Industrial Internet of Things (IIoT) to build multi-source sensing, precise modeling, and dynamic optimization and control capabilities covering the entire process, driving the evolution of production lines towards intelligence and autonomy.

[0003] The intelligent surface mount technology (SMT) production line relies on an industrial Internet of Things (IIoT) architecture, integrating manufacturing units such as solder paste printers, pick-and-place machines, and reflow ovens, as well as online quality inspection systems for solder paste inspection, pre-reflow optical inspection, and post-reflow inspection, forming a real-time closed-loop data flow. For example... Figure 3 As shown, this system framework integrates and coordinates multi-dimensional data streams, including circuit board design data, real-time operating status, process parameter settings, and multi-dimensional detection indicators, to build an online compensation capability for process parameters.

[0004] Figure 4 The arch-type pick-and-place machine shown has become the mainstream solution in the surface mount technology (SMT) industry due to its high versatility and excellent production efficiency. The core of this equipment is an XYZ three-axis precision motion platform. The placement fixture is driven by an X-axis motor to move along the crossbeam guide rail, and through the coordinated control of the Y-axis motor, it achieves precise positioning on the worktable. The placement fixture is equipped with multiple equidistantly arranged suction rod assemblies. Each suction rod can be independently controlled by a Z-axis motor to complete the vertical stroke of the pick-and-place action, and is driven by an R-axis motor to rotate around the axis to precisely adjust the component placement angle. The suction rods have vacuum and air blowing functions, used for picking up and placing components respectively (hereinafter referred to as "pick-and-place"). During operation, the suction rods on the integrated placement fixture, equipped with nozzles, cyclically perform pick-and-place operations according to a preset program: in the pick-up stage, the nozzles pick up components from the feeder under vacuum; in the placement stage, the components are positioned with micron-level precision on the corresponding pads on the circuit board. This mechatronics system performs the aforementioned pick-and-place actions at high speed, alternately, and with precision, achieving high-efficiency production while ensuring placement accuracy, thus meeting the stringent quality and efficiency requirements of modern precision electronic assembly.

[0005] The mounting accuracy of micro-components is a key factor determining the soldering quality. Typical soldering defects caused by mounting deviations include: (1) Misalignment defects: When the overlap area between the component terminal and the target pad is insufficient, the solder cannot completely connect to the pad, resulting in a decrease in the mechanical strength of the solder, which in turn leads to poor soldering or electrical failure; (2) Cold soldering defects: Mounting position deviations weaken the thermal coupling effect between the terminal and the pad, hindering the full formation of intermetallic compounds, resulting in discontinuous microstructure or insufficient bonding strength of the solder joint; (3) Tombstone defects: such as Figure 5 As shown, when the component mounting position shifts to a certain pad, the uneven surface tension of the pads at both ends during reflow soldering can easily cause the component to stand up at one end, forming a "tombstone" phenomenon.

[0006] Currently, mainstream pick-and-place machines typically provide ±40°C. The mounting accuracy is sufficient to meet the assembly requirements of components with package sizes larger than 0.402; while high-end models can achieve ±25. Even higher positioning accuracy is crucial for assembling micro-components such as 0201 and 01005, but their purchase and maintenance costs are often several times higher than ordinary models. Therefore, mainstream models with better cost performance are often chosen for applications requiring standard precision, while high-reliability electronic products require high-end equipment. Currently, mainstream pick-and-place machines typically lack the ability to finely adjust parameters based on the specific state of each board, and their placement effect largely depends on macroscopic statistical laws and manual experience adjustments. This approach cannot fully adapt to the individual differences of each circuit board, resulting in limited overall placement accuracy and affecting the final product assembly quality. By sensing the actual production status of each placement point on each circuit board in real time during the production process and adaptively adjusting the placement parameters online accordingly, the inherent precision limitations of the equipment can be effectively overcome without modifying the hardware, enabling reliable placement of higher-precision components.

[0007] The main shortcomings of current research are:

[0008] In existing surface mount technology (SMT) systems, mounting equipment struggles to dynamically adjust process parameters in real time based on multi-source sensing information in response to fluctuations in environmental conditions and changes in mounting requirements. This limitation significantly restricts the adaptive control capability of the mounting process, especially when mounting high-precision components, easily leading to decreased consistency and making it difficult to consistently guarantee soldering quality and overall yield. Summary of the Invention

[0009] The purpose of this invention is to address the problem in existing surface mount technology (SMT) systems where mounting equipment struggles to dynamically adjust process parameters in real time based on multi-source sensing information in response to environmental fluctuations and changes in mounting conditions. This leads to decreased consistency and makes it difficult to consistently guarantee welding quality and overall yield. Therefore, this invention proposes a feedforward adaptive control method for SMT process parameters based on solder paste printing detection.

[0010] The specific process of the feedforward adaptive control method for surface mount process parameters based on solder paste printing inspection is as follows:

[0011] Step 1: Obtain historical chip mounter production process parameter data, historical component data, and corresponding historical quality inspection result data;

[0012] Step 2: Construct a soldering offset prediction network model. Based on the historical chip mounter production process parameter data, historical component data, and corresponding historical quality inspection result data obtained in Step 1, a trained soldering offset prediction network model is obtained.

[0013] Step 3: Construct a production process parameter compensation model. Based on the historical pick-and-place machine production process parameter data, historical component data, corresponding historical quality inspection result data, and the trained soldering offset prediction network model obtained in Step 1, a trained production process parameter compensation model is obtained.

[0014] Step 4: After the circuit board is printed with solder paste using a solder paste printer, the solder paste printing coordinate offset of the pads is detected by solder paste inspection equipment. The system combines the initialized pick-and-place machine production process parameters with the trained soldering offset prediction network model and the trained production process parameter compensation model to obtain the adjusted production process parameters. The circuit board is then transferred to the pick-and-place machine to complete the surface soldering production of all mounting points on the current circuit board.

[0015] The beneficial effects of this invention are as follows:

[0016] This invention discloses a feedforward adaptive control method for surface mount process parameters based on solder paste printing inspection, aiming to solve the problem that pick-and-place machines lack adaptive capabilities and struggle to dynamically respond to soldering accuracy requirements and environmental changes. This method constructs a prediction and process parameter feedforward adaptive compensation framework for component soldering quality, including: a high-precision soldering offset prediction model based on pick-and-place machine production process parameters and online quality inspection results, used to predict the positional offset of soldered components on the circuit board under the current production process parameters; and a production process parameter compensation model obtained by executing a dual-delay depth deterministic strategy gradient algorithm based on the soldering offset prediction model, used to dynamically adjust component placement process parameters by combining real-time inspection data of the current mounted circuit board. Experiments demonstrate that this method can improve soldering accuracy in mass production of circuit boards and optimize the process capability index, providing a cross-equipment process fusion optimization scheme for precision electrical balance component manufacturing.

[0017] Specifically, this invention has the following three innovative features:

[0018] An end-to-end causal prediction and compensation framework is constructed, which includes modules for pre-prediction of production quality and adaptive compensation of production process parameters, significantly improving the response speed and quality stability of process optimization.

[0019] A soldering offset prediction model integrating causal inference and deep learning is proposed to enable the prediction of component position offset after reflow soldering based on production process parameters and solder paste inspection results.

[0020] A multi-objective optimization model for production process parameters that balances deviation accuracy and production efficiency is proposed to achieve dynamic adjustment of production process parameters for different circuit boards, resulting in an average reduction of 25% in component soldering position deviation and a maximum improvement of 39.0% in process capability index. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 This is a schematic diagram of the integrated framework of the present invention;

[0023] Figure 3 A schematic diagram of a collaborative optimization and closed-loop control system for the entire surface welding process;

[0024] Figure 4 A schematic diagram of an arch-type chip mounter equipped with a chip mount carrier;

[0025] Figure 5 A diagram illustrating tombstone defects caused by excessive mounting offset;

[0026] Figure 6 This is a comparison chart of the process parameter compensation results. Detailed Implementation

[0027] Specific implementation method one: Combining Figure 1 , Figure 2 This embodiment describes the specific process of the feedforward adaptive control method for surface mount process parameters based on solder paste printing inspection:

[0028] Step 1: Obtain historical chip mounter production process parameter data, historical component data, and corresponding historical quality inspection result data;

[0029] Step 2: Execute the dual machine learning causal forest algorithm to analyze the causal relationship between the pick-and-place machine production process parameters and the online quality inspection results. Integrate the extended long short-term memory network and the convolutional neural network to build a high-precision soldering offset prediction network model. Based on the historical pick-and-place machine production process parameter data, historical component data and corresponding historical quality inspection result data obtained in Step 1, the trained high-precision soldering offset prediction network model is obtained.

[0030] Step 3: Construct a production process parameter compensation model. Based on the historical chip mounter production process parameter data, historical component data, corresponding historical quality inspection result data, and the trained high-precision soldering offset prediction network model obtained in Step 1, a trained production process parameter compensation model is obtained.

[0031] Step 4: After the circuit board is printed with solder paste using a solder paste printer, the solder paste printing coordinate offset of the pads is detected by solder paste inspection equipment. The system combines the initialized pick-and-place machine production process parameters with the trained high-precision soldering offset prediction network model and the trained production process parameter compensation model to obtain the adjusted production process parameters. The circuit board is then transferred to the pick-and-place machine to complete the surface soldering production of all mounting points on the current circuit board.

[0032] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step one, historical pick-and-place machine production process parameter data, historical component data, and corresponding historical quality inspection result data are obtained; the specific process is as follows:

[0033] Step 11: Obtain historical pick-and-place machine production process parameters; the specific process is as follows:

[0034] Historical pick-and-place machine production process parameters include: total number of circuit board mounting points. Serial number of each sticker installation point Component types at each mounting point , Length direction coordinates of mounting position Mounting position width direction coordinates Mounting height Placement of mounting components and air blowing delay And the speed of the suction rod rise ;

[0035] Steps 1 and 2: Obtain historical component data; the specific process is as follows:

[0036] Historical component data includes: total number of component types Index numbers of various components Length of various components ,width ,thickness and component type Required placement accuracy threshold ;

[0037] Step 13: Obtain historical online quality inspection results data; the specific process is as follows:

[0038] Historical online quality inspection results include: coordinate offset of solder paste printing on pads corresponding to placement points. Offset of mounting point coordinates Coordinate offset of the mounting points after reflow soldering .

[0039] The other steps and parameters are the same as in Specific Implementation Method 1.

[0040] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step two, a dual machine learning causal forest algorithm is executed to analyze the causal relationship between the pick-and-place machine production process parameters and the online quality inspection results. An extended long short-term memory network and a convolutional neural network are then integrated to construct a high-precision soldering offset prediction network model. Based on the historical pick-and-place machine production process parameter data, historical component data, and corresponding historical quality inspection result data obtained in step one, a trained high-precision soldering offset prediction network model is obtained. The specific process is as follows:

[0041] Step 2: Construct a dataset based on historical pick-and-place machine production process parameter data, historical component data, and corresponding historical quality inspection results data. ;

[0042] The input parameters for the high-precision solder paste printing coordinate offset prediction network model are set as follows: and process parameters and mounting coordinates Mounting height Placement of mounting elements and air blowing delay The upward speed of the suction rod at the mounting point The length of various components ,width ,thickness ;

[0043] Set the output parameters of the high-precision solder offset prediction network model to the coordinate offset of the corresponding placement point after reflow soldering. ;

[0044] Step 22: Transfer historical datasets Randomly and uniformly divide the dataset and dataset (dataset) and dataset (The amount of data is equal);

[0045] Steps 2 and 3: Generate dataset Treatment residual and the result residual ;

[0046] Step 24: Construct a causal forest and estimate the heterogeneous treatment effect;

[0047] Step 25: Construct a spatiotemporal feature extraction network; the specific process is as follows:

[0048] Step 251: Construct a Convolutional Neural Network (CNN) module. The CNN module consists of an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer.

[0049] Step 252: Construct the Extended Long Short-Term Memory (xLSTM) module;

[0050] Steps 2-5-3: The spatiotemporal feature extraction network includes a convolutional neural network module and an extended long short-term memory (xLSTM) module.

[0051] The working process of the spatiotemporal feature extraction network is as follows:

[0052] Data is input into the convolutional neural network module, and the convolutional neural network module outputs features. Data is input into the extended long short-term memory (xLSTM) network module, and the extended long short-term memory (xLSTM) network module outputs features.

[0053] Step 26: Constructing a causal attention mechanism; the specific process is as follows:

[0054] Step 261: Initialize the attention weight matrix The dimension is the same as the dimension of the xLSTM output feature;

[0055] Step 262: Set Attention Weights Attenuation strategy, initial =0.8, which decreases linearly with training cycles, every 10 training cycles, Decrease by 0.1 until... It reached the minimum value of 0.2;

[0056] Step 263: Design a dual-gated attention mechanism: data-driven gate and causal a priori gate ;

[0057] Attention weight ;

[0058] in, For Hadamard products, It is the sigmoid activation function;

[0059] Step 264: Weight attention The final result is obtained by performing a Hadamard product with the output features of the spatiotemporal feature extraction network;

[0060] Step 27: Constructing interventions to enhance input; the specific process is as follows:

[0061] Step 271: Define the input of the spatiotemporal feature extraction network The intervention transformation matrix is ;

[0062] Step 272: Design the intervention intensity decay function: ;in For the current training cycle;

[0063] Step 273: Constructing interventions to enhance input: ;

[0064] Step 28: Design the composite loss function ;

[0065] Step 29: Train the high-precision welding offset prediction network model to obtain the trained high-precision welding offset prediction network model.

[0066] Other steps and parameters are the same as in specific implementation method one or two.

[0067] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the dataset is generated in steps Two and Three. Treatment residual and the result residual The specific process is as follows:

[0068] Step 231: Transfer the dataset Randomly divided into A subset of equal size;

[0069] Step 232: Select a subset Used for evaluation Use the remaining The original training set is composed of several subsets;

[0070] Step 233: Select any one of the input and output parameters of the high-precision welding offset prediction network model as a variable. ;

[0071] Steps 2, 3, and 4: Randomly select 70% of the samples with replacement from the original training set in Step 2, 3, and 2 to form the Bootstrap training set.

[0072] Each sample is a sequence, and each sequence is... ;

[0073] Steps two, three, and five: Generate a decision tree using the Bootstrap training set;

[0074] The specific process is as follows:

[0075] Step 2351: Traverse all possible split points of the root node and select the one that results in a weighted sum of the mean square errors of the two child nodes after the split. Find the smallest split point, split the root node's sample into two child nodes (left and right). and ;

[0076] The root node is the Bootstrap training set, and all possible split points of the root node are all samples in the Bootstrap training set.

[0077] in ;

[0078] in,

[0079] This represents the number of samples in the current node. and These are the variables in the left and right child node samples, respectively. , They are respectively and The average value;

[0080] Steps two, three, five, two,

[0081] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0082] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0083] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0084] Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ;

[0085] Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ;

[0086] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0087] Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ;

[0088] Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ;

[0089] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0090] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0091] Determine the right child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0092] Determine the right child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0093] Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0094] Determine the right child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ;

[0095] Executing steps two, three, and five once yields one decision tree;

[0096] The maximum depth of a decision tree is 4.

[0097] The nodes that cannot be split further are called leaf nodes, and the predicted values ​​for a leaf node are the variables of all training samples within that leaf node. The average value;

[0098] Steps 2, 3, and 6: Determine the number of decision trees generated. Does it meet the requirements? If yes, proceed to steps two, three, and seven; if no, proceed to steps two, three, and four.

[0099] Steps two, three, and seven: The decision trees are ensembled into a random forest model, and the ensemble random forest model is used to evaluate the validation subset. Make predictions and obtain the predicted values ​​of variables from the validation subset samples. ;

[0100] subset for: The set that constitutes;

[0101] Predicted value for:

[0102] Any one of the parameters in;

[0103] Compute the validation set subset residuals: ;

[0104] If the variable is a process parameter, then To treat residual If the variable belongs to the coordinate offset after reflow soldering, then For the result residual ;

[0105] The specific process is as follows:

[0106] Step 2371: Let the random forest model obtained in step 236 contain... There are six decision trees, denoted as the first decision tree, the second decision tree, ..., the third decision tree. Decision trees; the corresponding input parameter vector is ;

[0107] Step 2372, for the subset of the verification set any sample in The corresponding samples are input into each decision tree in the random forest model, and each decision tree outputs the results for the variables. The predicted values ​​(5 trees, 5 values) are as follows:

[0108] Step 23721, Select the first Decision trees, which will store samples As input, from the first The root node of a decision tree (including samples) The traversal begins with the MSE corresponding to the root node.

[0109] Step 23722: At the current node, read the variables corresponding to the current node and the MSE determined in Step 235, and compare the samples. The relationship between the value of the corresponding variable and the size of the current node's MSE:

[0110] If the sample If the value of the corresponding variable is less than the MSE of the current node, then the sample is moved to the next node. Passed to the left child node of the current node;

[0111] If the sample If the value of the corresponding variable is greater than or equal to the current node MS, then the sample... Pass it to the right child node of the current node;

[0112] Step 23723, repeat step 23722 until the sample is complete. Falling into a leaf node (a node that cannot be split further is a leaf node);

[0113] Steps 2, 3, 7, 2, 4: Read the variables from the corresponding leaf nodes. The predicted value of the leaf node is denoted as the first. Decision trees for samples Predicted value The predicted value of a leaf node is the variable of all training samples within that leaf node. The average value;

[0114] Steps 2, 3, 7, and 3: Perform a validation set subset analysis. The pair of variables output by each sample in each decision tree The predicted values ​​are integrated to obtain the final predicted values ​​of the random forest model for the samples. Specifically:

[0115]

[0116] Steps two, three, seven, and four: Validating the subset of the validation set. All samples are processed sequentially through steps 2372 to 2373 to obtain a subset of the validation set. Predicted value (e.g., validation set subset) All 10 samples are included, and each of the 10 samples corresponds to one of the 10 final predicted values. 10 final predicted values To form a 10-row matrix );

[0117] Step 2375: Calculate the verification set subset residuals: (e.g., validation set subset) All 10 samples, each corresponding to one of the 10 variables, and each of the 10 variables... To form a 10-row matrix A 10-row matrix Subtract a 10-row matrix This yields a 10-row matrix residual. );

[0118] If the variable is a process parameter, then To treat residual If the variable belongs to the coordinate offset after reflow soldering, then For the result residual ;

[0119] Step 238: Determine whether the residual calculation for all variables has been completed. If not, return to step 233; if yes, proceed to step 239.

[0120] Steps 2, 3, and 9: Determine if all If all subsets have completed residual calculation, then return to step 232; otherwise, proceed to step 24.

[0121] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0122] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that step two and four involves constructing a causal forest and estimating the heterogeneous treatment effect; the specific process is as follows:

[0123] Step 241: From the dataset The Bootstrap sample set is formed by randomly selecting samples with replacement. ;

[0124] Step 242: Record the current node depth (Initially 0), number of samples in the current node ;

[0125] The current node is the Bootstrap sample set. Number of samples in the current node Bootstrap sample set All samples;

[0126] Steps two through three: splitting operation; the specific process is as follows:

[0127] Traverse all possible split points of the current node (all possible split points of the root node are the Bootstrap sample set). (all samples), the current node Split into two child nodes at the split point. and Select the split point that maximizes the heterogeneity criterion value for splitting;

[0128] The formula for calculating the heterogeneity criterion is as follows:

[0129]

[0130] in, , For the number of variables, This represents the estimated treatment effect within the child node. The split point;

[0131] child node medium sample Input the random forest model, select process parameters as variables, and calculate the residual values ​​after the random forest model outputs the predicted values. ; This represents the total number of samples at the current node.

[0132] child nodes medium sample Input the random forest model, select the coordinate offset after reflow soldering as the variable, and calculate the residual value after the random forest model outputs the predicted value;

[0133] child node Inner Sample The average treatment effect estimate of each variable;

[0134] child node Inner Sample The average treatment effect estimate of each variable;

[0135] This is the value of the heterogeneity criterion;

[0136] Each sample is a sequence, and each sequence is... ;

[0137] Step 244: Determine if the conditions for continued division are met: and If yes, proceed to steps two, four, and three; otherwise, proceed to steps two, four, and five. This represents the maximum depth of the current node. This represents the total number of samples at the current node.

[0138] Steps 2, 4, and 5: Determine if the required number of generated causal trees is met. If yes, proceed to step 241; otherwise, proceed to step 246. The current number of causal trees, This represents the maximum number of causal trees.

[0139] Steps 2, 4, and 6: For the database Each data point Calculate each data point Similarity weights in each causal tree ;

[0140] Where the training samples With data points In the causal tree If they fall into the same leaf node, ,otherwise ;

[0141] Data points For a sequence ;

[0142] Training samples For dataset medium sample A sequence

[0143] in Representative sample A corresponding mounting point;

[0144] Step 247: Estimating the heterogeneous treatment effect based on similarity weights: .

[0145] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0146] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that step two-eight involves designing a composite loss function. The specific process is as follows:

[0147] Step 281: Define the prediction loss term: ;

[0148] in, This is the predicted displacement value. This represents the true displacement value. This is the function for calculating the mean squared error;

[0149] Step 282: Define the causal consistency loss term: ;

[0150] in, Predicted displacement values ​​after intervention Compared with the displacement prediction value without intervention difference, This is a heterogeneous treatment effect;

[0151] Step 283: Designing Causal Weights Attenuation strategy: Cosine annealing strategy is adopted. ;

[0152] in, For the current training cycle, This represents the total number of training cycles.

[0153] Step 284: Define the composite loss function : .

[0154] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0155] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that step 29 involves training the high-precision welding offset prediction network model to obtain a trained high-precision welding offset prediction network model; the specific process is as follows: Step 291: Initialize the parameters of the high-precision welding offset prediction network model and set the total number of training cycles. Initialize the current training cycle ;

[0156] Step 292: Input Data Feature extraction is performed using a spatiotemporal feature extraction network, and a causal attention mechanism is applied (attention weights are obtained using the causal attention mechanism in step two-six). Pay attention weight The displacement prediction is generated by performing a Hadamard product with the output features of the spatiotemporal feature extraction network. The output layer generates displacement prediction values. ;

[0157] Step 293: Process the input data Intervention operations were performed to obtain Then, features are extracted through a spatiotemporal feature extraction network, and a causal attention mechanism is applied (attention weights are obtained by applying the causal attention mechanism in step two). Pay attention weight The Hadamard product is performed with the output features of the spatiotemporal feature extraction network to generate the predicted displacement value after intervention. The output layer generates the predicted displacement values ​​after intervention. ;

[0158] Step 294, Predicted Value Predicted displacement values ​​after intervention Compared with the true value Through the composite loss function Calculate the loss;

[0159] Step 295: Calculate the gradient using the backpropagation algorithm and update the parameters of the high-precision welding offset prediction network model;

[0160] Step 296, Judgment Has the total number of training cycles been reached? If yes, proceed to step twenty; otherwise, reduce the training period. Perform step 292;

[0161] Step 20: Save the network weight values ​​to obtain the trained welding offset prediction network model.

[0162] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0163] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that, in step three, a production process parameter compensation model is constructed. Based on the historical pick-and-place machine production process parameter data, historical component data, corresponding historical quality inspection result data, and the trained high-precision soldering offset prediction network model obtained in step one, a trained production process parameter compensation model is obtained. The specific process is as follows:

[0164] Step 3.1 Define the maximum number of training cycles. Maximum number of time steps per cycle Discount factor Policy network update delay Soft update rate Define the action space boundary Policy Network Policy network parameters Target Policy Network Target policy network parameters Two Critic networks , Parameters of two Critic networks , Two target Critic networks , Parameters of the two target Critic networks , Experience replay buffer The trained high-precision welding offset prediction network model Environmental status parameters Explore noise (mean is 0, variance is 0) Gaussian function), offset tolerance Batch size ;

[0165] Step 3.2: Construct a production process parameter compensation model; the specific process is as follows:

[0166] Step 321: Define the policy network Input: Historical environmental status parameters ;

[0167] Target Policy Network Input and Policy Networks The input is the same;

[0168] Step 3.2.2 Define the policy network Output: Historical Actions ;

[0169] Target Policy Network Output and Policy Network The outputs are the same;

[0170] Step 3.2.2. Define the input for each Critic network: historical solder paste printing coordinate offset. Historically adjusted mounting coordinates Mounting height Placement of mounting components and air blowing delay Mounting speed of suction rod Length of various components ,width and thickness ;

[0171] The target Critic network input is the same as the Critic network input;

[0172] Steps 3-4: Define the output of each Critic network: History Value estimation, ;

[0173] The target Critic network output is the same as the Critic network output;

[0174] Step 33: Action Selection and Exploration Strategy; the specific process is as follows:

[0175] Step 331, regarding the time step Policy Network Accept the current environmental state ;

[0176] Step 332: Generate deterministic actions Add exploration noise Clip to the boundary of motion space ;

[0177] in, For policy networks;

[0178] Indicates the action Crop to inside, if ,but ;like ,but ;like ,but No change;

[0179] Steps three and four: Calculate the reward function; the specific process is as follows:

[0180] Step 3-41: Use the trained high-precision welding offset prediction network model Predicted coordinate offset values ​​after reflow soldering : ;

[0181] in , The predicted coordinate offset of the placement point after reflow soldering;

[0182] Step 342: Calculate the offset Euclidean distance : ;

[0183] Step 3. Calculate the component size normalization factor. : ;

[0184] Steps 3 and 4: Calculation of Efficiency Bonus Items : ;

[0185] in, and These are the minimum and maximum values ​​of the suction rod's upward speed at the mounting point, respectively. and These are the minimum and maximum values ​​for the air blowing delay when placing the component at the mounting point; and These are the weights for speed and latency, respectively.

[0186] Steps 3, 4, and 5: Calculate the dynamic displacement weights. : ;

[0187] in, For components The offset tolerance, if If so, it is considered that the offset exceeds the offset tolerance.

[0188] Steps 3, 4, and 6: Calculate the total reward : ;

[0189] Step 35: Training and updating the production process parameter compensation model to obtain a trained production process parameter compensation model;

[0190] Step 36: Save the target network parameters This yields a well-trained production process parameter compensation model.

[0191] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0192] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that, in step three and five, the production process parameter compensation model is trained and parameters are updated to obtain a trained production process parameter compensation model; the specific process is as follows:

[0193] Step 3.51, Time Step Calculate the target action: for each state Input into the policy network Obtain the target action Execute the target action Get the state ;

[0194] Step 352: Determine the current environment status Input a pre-trained high-precision welding offset prediction network model Obtain the offset Euclidean distance ;

[0195] Calculate total reward Calculation based on two Critic networks value, ;

[0196] Based on total reward and calculate ,when or time step hour, ,otherwise Update the termination flag. ;

[0197] in, For the goal value;

[0198] Step 353, Storage Experience:

[0199] Experience {the environment state before interaction} ,action ,award The new state after the action is performed Termination mark Stored in the experience replay buffer In the middle, if If yes, proceed to step 351; otherwise, proceed to step 354.

[0200] Steps 3, 5, and 4: Small-batch data sampling: from the experience replay buffer Random sampling Number of samples;

[0201] Step 3.55: Calculate the two Critic loss functions. and Perform gradient descent to update the Critic network parameters. and Update the two Critic networks; the specific process is as follows:

[0202]

[0203]

[0204]

[0205]

[0206] in, Represents the loss function Critic network parameters gradient, Represents the loss function Critic network parameters The gradient;

[0207] Indicates sample goal value, ;

[0208] Indicates sample Input the first Critic network calculation value;

[0209] Indicates sample Input the second Critic network calculation value;

[0210] Steps 3, 5, and 6: Policy Network Delay Update: If the current time step... Calculate the policy gradient Execute gradient ascent to update network parameters ;

[0211] Policy gradient Represented as:

[0212]

[0213] in, Indicates the first Sample, Indicates the number of samples;

[0214] This indicates calculating the gradient of action A; mod represents the modulo operation; for example, to illustrate the result of mod, if we calculate 7... At time 2, since the quotient of 7 divided by 2 is 3 and the remainder is 1, therefore 7 2 = 1;

[0215] Step 357, Target Network Soft Update:

[0216] Update target Critic network parameters , ;

[0217] Update target policy network parameters ;

[0218] in, Indicates the soft update rate. Indicates the policy network parameters;

[0219] Step 358, if time step If so, proceed to step 36; otherwise, proceed to step 351.

[0220] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0221] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that, in step four, after the circuit board undergoes solder paste printing using a solder paste printer, the solder paste printing coordinate offset of the pads is detected by a solder paste inspection device. The initial pick-and-place machine production process parameters are then fed into a trained high-precision soldering offset prediction network model and a trained production process parameter compensation model to obtain adjusted production process parameters. The circuit board is then transferred to the pick-and-place machine to complete the surface soldering production of all mounting points on the current circuit board. The specific process is as follows:

[0222] Step 4: After the solder paste is printed on the circuit board by the solder paste printer, it is sent to the solder paste inspection equipment (SPI) for inspection. The SPI equipment checks whether the solder paste at each mounting point on the current circuit board is misaligned. If the solder paste at each mounting point is not misaligned, a new circuit board is inspected; if the solder paste at each mounting point is misaligned, the SPI equipment obtains the result of the misalignment. The circuit board is then transferred to a pick-and-place machine for production.

[0223] Step 42: Adjust the solder paste offset of each mounting point on the current circuit board. Process parameters and mounting coordinates Mounting height Placement of mounting components and air blowing delay Mounting speed of suction rod Length of various components ,width ,thickness Input a trained high-precision solder joint offset prediction network model, and output the coordinate offset of the corresponding mounting point after reflow soldering. ;

[0224] based on Calculate the offset Euclidean distance : ;

[0225] when Instead of performing step four and three, select the current production process parameters for surface welding production;

[0226] when Proceed to step four three;

[0227] Step 43: Place the current circuit board The placement position length direction coordinates in the initial production process parameters of the pick-and-place machine and width direction coordinates Mounting height Placement of mounting components and air blowing delay Mounting speed of suction rod and the length of various components ,width ,thickness The target policy network, which is input into the production process parameter compensation model, outputs the adjusted process parameter mounting coordinates after the model has been trained. Mounting height Placement of mounting elements and air blowing delay The upward speed of the suction rod at the mounting point ;

[0228] Step 4: Place the adjusted process parameters onto the coordinate system. Mounting height Placement of mounting elements and air blowing delay The upward speed of the suction rod at the mounting point The circuit boards currently being produced by the pick-and-place machine are now being surface-mounted.

[0229] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0230] The beneficial effects of the present invention are verified using the following embodiments:

[0231] Example 1:

[0232] The surface mount process parameter feedforward adaptive control method based on solder paste printing detection in this embodiment is prepared according to the following steps:

[0233] The example considers the adaptive compensation process of process parameters for a six-suction beam-type chip mounter.

[0234] Without using any process parameter compensation scheme, the pick-and-place machine was set according to normal production process parameters, using the traditional "place on pad" method. The average deviation in the soldering length direction of the corresponding circuit board was... Standard deviation in length direction Average deviation in width direction Standard deviation in width direction Process capability index . This indicates that the corresponding mounting result does not meet the process accuracy requirements.

[0235] The process parameters were set using the "pick-and-place process control based on the bootstrapping method for quality enhancement in surface mount technology" from the paper "A pick-and-place process control based on the bootstrapping method for quality enhancement in surface mount technology". The average deviation in the soldering length direction of the corresponding circuit board was determined. Standard deviation in length direction Average deviation in width direction Standard deviation in width direction Process capability index .

[0236] Using the adaptive compensation method for process parameters of this invention, if a soldering offset prediction model is not used, but instead an offset prediction model trained with the xLSTM algorithm from "xlstm: Extended long short-term memory", the average deviation in the soldering length direction of the circuit board corresponding to the component placement process parameter compensation result is... Standard deviation in length direction Average deviation in width direction Standard deviation in width direction Process capability index Compared to the traditional "mount-to-pad" method, the process parameter feedforward adaptive compensation method using the xLSTM algorithm improves the process capability index by a significant margin. .

[0237] Using the process parameter feedforward adaptive compensation method of this invention, if an offset prediction model is adopted, the average deviation in the circuit board soldering length direction corresponding to the component mounting process parameter compensation result is... Standard deviation in length direction Average deviation in width direction Standard deviation in width direction Process capability index Compared to the traditional "mount-to-pad" method, the process capability index... Improved Compared to the process parameter adaptive method using the xLSTM algorithm, the process capability index... Improved .

[0238] Figure 6The results of adaptive compensation of process parameters using the above algorithm were further compared. The horizontal axis represents the deviation of the component welding coordinates in the length direction, and the vertical axis represents the deviation of the component welding coordinates in the width direction. From Figure 6 As can be seen, by using the adaptive compensation of process parameters of the present invention, the coordinate deviation of the welded component obtained is significantly smaller than that obtained by other production methods.

[0239] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A feedforward adaptive control method for surface mount process parameters based on solder paste printing inspection, characterized in that: The specific process of the method is as follows: Step 1: Obtain historical chip mounter production process parameter data, historical component data, and corresponding historical quality inspection result data; Step 2: Construct a soldering offset prediction network model. Based on the historical chip mounter production process parameter data, historical component data, and corresponding historical quality inspection result data obtained in Step 1, a trained soldering offset prediction network model is obtained. Step 3: Construct a production process parameter compensation model. Based on the historical pick-and-place machine production process parameter data, historical component data, corresponding historical quality inspection result data, and the trained soldering offset prediction network model obtained in Step 1, a trained production process parameter compensation model is obtained. Step 4: After the circuit board is printed with solder paste using a solder paste printer, the solder paste printing coordinate offset of the pads is detected by solder paste inspection equipment. The system combines the initialized pick-and-place machine production process parameters with the trained soldering offset prediction network model and the trained production process parameter compensation model to obtain the adjusted production process parameters. The circuit board is then transferred to the pick-and-place machine to complete the surface soldering production of all mounting points on the current circuit board.

2. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 1, characterized in that: Step one involves acquiring historical pick-and-place machine production process parameter data, historical component data, and corresponding historical quality inspection result data; the specific process is as follows: Step 11: Obtain historical pick-and-place machine production process parameters; the specific process is as follows: Historical pick-and-place machine production process parameters include: total number of circuit board mounting points. Serial number of each sticker installation point Component types at each mounting point , Length direction coordinates of mounting position Mounting position width direction coordinates Mounting height Placement of mounting components and air blowing delay And the speed of the suction rod rise ; Steps 1 and 2: Obtain historical component data; the specific process is as follows: Historical component data includes: total number of component types Index numbers of various components Length of various components ,width ,thickness and component type Required placement accuracy threshold ; Step 13: Obtain historical online quality inspection results data; the specific process is as follows: Historical online quality inspection results include: coordinate offset of solder paste printing on pads corresponding to placement points. Offset of mounting point coordinates Coordinate offset of the mounting points after reflow soldering .

3. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 2, characterized in that: In step two, a soldering offset prediction network model is constructed. Based on the historical chip mounter production process parameter data, historical component data, and corresponding historical quality inspection result data obtained in step one, a trained soldering offset prediction network model is obtained. The specific process is as follows: Step 2: Construct a dataset based on historical pick-and-place machine production process parameter data, historical component data, and corresponding historical quality inspection results data. ; Set the input parameters of the solder paste printing coordinate offset of the solder pad to predict the solder offset network model. and process parameters and mounting coordinates Mounting height Placement of mounting elements and air blowing delay The upward speed of the suction rod at the mounting point The length of various components ,width ,thickness ; Set the output parameters of the solder offset prediction network model to the coordinate offset of the corresponding placement point after reflow soldering. ; Step 22: Transfer historical datasets Randomly and uniformly divide the dataset and dataset ; Steps 2 and 3: Generate dataset Treatment residual and the result residual ; Step 24: Construct a causal forest and estimate the heterogeneous treatment effect; Step 25: Construct a spatiotemporal feature extraction network; the specific process is as follows: Step 251: Construct a convolutional neural network module, which consists of an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer. Step 252: Construct the Extended Long Short-Term Memory (xLSTM) module; Steps 2-5-3: The spatiotemporal feature extraction network includes a convolutional neural network module and an extended long short-term memory (xLSTM) module. The working process of the spatiotemporal feature extraction network is as follows: Data is input into the convolutional neural network module, and the convolutional neural network module outputs features. Data is input into the extended long short-term memory (xLSTM) network module, and the extended long short-term memory (xLSTM) network module outputs features. Step 26: Constructing a causal attention mechanism; the specific process is as follows: Step 261: Initialize the attention weight matrix The dimension is the same as the dimension of the xLSTM output feature; Step 262: Set Attention Weights Attenuation strategy, initial =0.8, which decreases linearly with training cycles, every 10 training cycles, Decrease by 0.1 until... It reached the minimum value of 0.2; Step 263: Design a dual-gated attention mechanism: data-driven gate and causal a priori gate ; Attention weight ; in, For Hadamard products, It is the sigmoid activation function; Step 264: Weight attention The final result is obtained by performing a Hadamard product with the output features of the spatiotemporal feature extraction network; Step 27: Constructing interventions to enhance input; the specific process is as follows: Step 271: Define the input of the spatiotemporal feature extraction network The intervention transformation matrix is ; Step 272: Design the intervention intensity decay function: ;in For the current training cycle; Step 273: Constructing interventions to enhance input: ; Step 28: Design the composite loss function ; Step 29: Train the welding offset prediction network model to obtain the trained welding offset prediction network model.

4. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 3, characterized in that: In steps two and three, a dataset is generated. Treatment residual and the result residual ; The specific process is as follows: Step 231: Transfer the dataset Randomly divided into A subset of equal size; Step 232: Select a subset Used for evaluation Use the remaining The original training set is composed of several subsets; Step 233: Select any one of the input and output parameters of the welding offset prediction network model as a variable. ; Steps 2, 3, and 4: Randomly select 70% of the samples with replacement from the original training set in Step 2, 3, and 2 to form the Bootstrap training set. Each sample is a sequence, and each sequence is... ; Steps two, three, and five: Generate a decision tree using the Bootstrap training set; The specific process is as follows: Step 2351: Traverse all possible split points of the root node and select the one that results in a weighted sum of the mean square errors of the two child nodes after the split. Find the smallest split point, split the root node's sample into two child nodes (left and right). and ; in ; in, This represents the number of samples in the current node. and These are the variables in the left and right child node samples, respectively. , They are respectively and The average value; Steps two, three, five, two, Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ; Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the right child node? No further splitting; if not, traverse the right child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the right child node is selected. The middle sample is divided into left and right child nodes. and ; Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the left child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Determine the right child node Is the number of samples less than 2? If so, is it the left child node? No further splitting; if not, traverse the left child node. Choose from all possible split points that result in a weighted sum of the mean squared errors of the two child nodes after the split. The smallest split point is used to split the node, and the left child node is selected. The samples are divided into left and right child nodes. and ; Steps 2, 3, and 6: Determine the number of decision trees generated. Does it meet the requirements? If yes, proceed to steps two, three, and seven; otherwise, proceed to steps two, three, and four. Steps two, three, and seven: The decision trees are ensembled into a random forest model, and the ensemble random forest model is used to evaluate the validation subset. Make predictions and obtain the predicted values ​​of variables from the validation subset samples. ; subset for: The set that constitutes; Predicted value for: Any one of the parameters in; Compute the validation set subset residuals: ; If the variable is a process parameter, then To treat residual If the variable belongs to the coordinate offset after reflow soldering, then For the result residual ; The specific process is as follows: Step 2371: Let the random forest model obtained in step 236 contain... There are six decision trees, denoted as the first decision tree, the second decision tree, ..., the third decision tree. Decision trees; the corresponding input parameter vector is ; Step 2372, for the subset of the verification set any sample in The corresponding samples are input into each decision tree in the random forest model, and each decision tree outputs the results for the variables. The predicted values ​​are as follows: Step 23721, Select the first Decision trees, which will store samples As input, from the first The traversal begins from the root node of the decision tree; Step 23722: At the current node, read the variables corresponding to the current node and the MSE determined in Step 235, and compare the samples. The relationship between the value of the corresponding variable and the size of the current node's MSE: If the sample If the value of the corresponding variable is less than the MSE of the current node, then the sample is moved to the next node. Passed to the left child node of the current node; If the sample If the value of the corresponding variable is greater than or equal to the current node MS, then the sample... Pass it to the right child node of the current node; Step 23723, repeat step 23722 until the sample is complete. It falls into a leaf node; Steps 2, 3, 7, 2, and 4: Read the variables from the corresponding leaf nodes. The predicted value of the leaf node is denoted as the first. Decision trees for samples Predicted value ; Steps 2, 3, 7, and 3: Perform a validation set subset analysis. The pair of variables output by each sample in each decision tree The predicted values ​​are integrated to obtain the final predicted values ​​of the random forest model for the samples. Specifically: Steps two, three, seven, and four: Validating the subset of the validation set. All samples are processed sequentially through steps 2372 to 2373 to obtain a subset of the validation set. Predicted value ; Step 2375: Calculate the verification set subset residuals: ; If the variable is a process parameter, then To treat residual If the variable belongs to the coordinate offset after reflow soldering, then For the result residual ; Step 238: Determine whether the residual calculation for all variables has been completed. If not, return to step 233; if yes, proceed to step 239. Steps 2, 3, and 9: Determine if all If all subsets have completed residual calculation, then return to step 232; otherwise, proceed to step 24.

5. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 4, characterized in that: In step two and four, a causal forest is constructed and the heterogeneous treatment effect is estimated; The specific process is as follows: Step 241: From the dataset The Bootstrap sample set is formed by randomly selecting samples with replacement. ; Step 242: Record the current node depth Number of samples in the current node ; Steps two through three: splitting operation; the specific process is as follows: Iterate through all possible split points of the current node and split the current node. Split into two child nodes at the split point. and Select the split point that maximizes the heterogeneity criterion value for splitting; in The formula for calculating the heterogeneity criterion is: in, , For the number of variables, This represents the estimated treatment effect within the child node. The split point; child node medium sample Input the random forest model, select process parameters as variables, and calculate the residual values ​​after the random forest model outputs the predicted values. ; This represents the total number of samples at the current node. child nodes medium sample Input the random forest model, select the coordinate offset after reflow soldering as the variable, and calculate the residual value after the random forest model outputs the predicted value; child node Inner Sample The average treatment effect estimate of each variable; child node Inner Sample The average treatment effect estimate of each variable; This is the value of the heterogeneity criterion; Each sample is a sequence, and each sequence is... ; Step 244: Determine if the conditions for continued division are met: and If yes, proceed to steps two, four, and three; otherwise, proceed to steps two, four, and five. This represents the maximum depth of the current node. This represents the total number of samples at the current node. Steps 2, 4, and 5: Determine if the required number of generated causal trees is met. If yes, proceed to step 241; otherwise, proceed to step 246. The current number of causal trees, This represents the maximum number of causal trees. Steps 2, 4, and 6: For the database Each data point Calculate each data point Similarity weights in each causal tree ; Where the training samples With data points In the causal tree If they fall into the same leaf node, ,otherwise ; Data points For a sequence ; Training samples For dataset medium sample A sequence ; in Representative sample A corresponding mounting point; Step 247: Estimating the heterogeneous treatment effect based on similarity weights: .

6. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 5, characterized in that: In step two-eight, a composite loss function is designed. The specific process is as follows: Step 281: Define the prediction loss term: ; in, This is the predicted displacement value. This represents the true displacement value. This is the function for calculating the mean squared error; Step 282: Define the causal consistency loss term: ; in, Predicted displacement values ​​after intervention Compared with the displacement prediction value without intervention difference, This is a heterogeneous treatment effect; Step 283: Designing Causal Weights Attenuation strategy: Cosine annealing strategy is adopted. ; in, For the current training cycle, This represents the total number of training cycles. Step 284: Define the composite loss function : .

7. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 6, characterized in that: In step 29, the welding offset prediction network model is trained to obtain a trained welding offset prediction network model; the specific process is as follows: Step 291: Initialize the parameters of the welding offset prediction network model and set the total number of training cycles. Initialize the current training cycle ; Step 292: Input Data Feature extraction is performed using a spatiotemporal feature extraction network, and a causal attention mechanism is applied to generate displacement prediction values ​​in the output layer. ; Step 293: Process the input data Intervention operations were performed to obtain The feature extraction is then performed by a spatiotemporal feature extraction network, and a causal attention mechanism is applied to generate the predicted displacement value after intervention in the output layer. ; Step 294, Predicted Value Predicted displacement values ​​after intervention Compared with the true value Through the composite loss function Calculate the loss; Step 295: Calculate the gradient using the backpropagation algorithm and update the parameters of the welding offset prediction network model; Step 296, Judgment Has the total number of training cycles been reached? ; If yes, proceed to step twenty; otherwise, set the training cycle to zero. Perform step 292; Step 20: Save the network weight values ​​to obtain the trained welding offset prediction network model.

8. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 7, characterized in that: In step three, a production process parameter compensation model is constructed. Based on the historical pick-and-place machine production process parameter data, historical component data, corresponding historical quality inspection result data, and the trained soldering offset prediction network model obtained in step one, a trained production process parameter compensation model is obtained. The specific process is as follows: Step 3.1 Define the maximum number of training cycles. Maximum number of time steps per cycle Discount factor Policy network update delay Soft update rate Define the action space boundary Policy Network Policy network parameters Target Policy Network Target policy network parameters Two Critic networks , Parameters of two Critic networks , Two target Critic networks , Parameters of the two target Critic networks , Experience replay buffer The trained welding offset prediction network model Environmental status parameters Explore noise Offset tolerance Batch size ; Step 3.2: Construct a production process parameter compensation model; the specific process is as follows: Step 321: Define the policy network Input: Historical environmental status parameters ; Target Policy Network Input and Policy Networks The input is the same; Step 3.2.2 Define the policy network Output: Historical Actions ; Target Policy Network Output and Policy Network The outputs are the same; Step 3.2.

2. Define the input for each Critic network: historical solder paste printing coordinate offset. Historically adjusted mounting coordinates Mounting height Placement of mounting components and air blowing delay Mounting speed of suction rod Length of various components ,width and thickness ; The target Critic network input is the same as the Critic network input; Steps 3-4: Define the output of each Critic network: History Value estimation, ; The target Critic network output is the same as the Critic network output; Step 33: Action Selection and Exploration Strategy; the specific process is as follows: Step 331, regarding the time step Policy Network Accept the current environmental state ; Step 332: Generate deterministic actions Add exploration noise Clip to the boundary of motion space ; in, For policy networks; Indicates the action Crop to inside, if ,but ;like ,but ;like ,but No change; Steps three and four: Calculate the reward function; the specific process is as follows: Step 3-41: Use the trained welding offset prediction network model Predicted coordinate offset values ​​after reflow soldering : ; in , The predicted coordinate offset of the placement point after reflow soldering; Step 342: Calculate the offset Euclidean distance : ; Step 3. Calculate the component size normalization factor. : ; Steps 3 and 4: Calculation of Efficiency Bonus Items : ; in, and These are the minimum and maximum values ​​of the suction rod's upward speed at the mounting point, respectively. and These are the minimum and maximum values ​​for the air blowing delay when placing the component at the mounting point; and These are the weights for speed and latency, respectively. Steps 3, 4, and 5: Calculate the dynamic displacement weights. : ; in, For components The allowable offset; Steps 3, 4, and 6: Calculate the total reward : ; Step 35: Training and updating the production process parameter compensation model to obtain a trained production process parameter compensation model; Step 36: Save the target network parameters This yields a well-trained production process parameter compensation model.

9. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 8, characterized in that: In step three and five, the production process parameter compensation model is trained and updated to obtain a trained production process parameter compensation model; the specific process is as follows: Step 3.51, Time Step Calculate the target action: for each state Input into the policy network Obtain the target action Execute the target action Get the state ; Step 352: Determine the current environment status Input the trained welding offset prediction network model Obtain the offset Euclidean distance ; Calculate total reward Calculation based on two Critic networks value, ; Based on total reward and calculate ,when or time step hour, ,otherwise Update the termination flag. ; in, For the goal value; Step 353, Storage Experience: Experience {the environment state before interaction} ,action ,award The new state after the action is performed Termination mark Stored in the experience replay buffer In the middle, if If yes, proceed to step 351; otherwise, proceed to step 354. Steps 3, 5, and 4: Replay from the experience buffer Random sampling Number of samples; Step 3.55: Calculate the two Critic loss functions. and Perform gradient descent to update the Critic network parameters. and Update the two Critic networks; the specific process is as follows: in, Represents the loss function Critic network parameters gradient, Represents the loss function Critic network parameters The gradient; Indicates sample goal value, ; Indicates sample Input the first Critic network calculation value; Indicates sample Input the second Critic network calculation value; Steps 3, 5, and 6: Policy Network Delay Update: If the current time step... Calculate the policy gradient Execute gradient ascent to update network parameters ; Policy gradient Represented as: in, Indicates the first Sample, Indicates the number of samples; This indicates calculating the gradient of action A; mod represents the modulo operation. Step 357, Target Network Soft Update: Update target Critic network parameters , ; Update target policy network parameters ; in, Indicates the soft update rate. Indicates the policy network parameters; Step 358, if time step If so, proceed to step 36; otherwise, proceed to step 351.

10. The surface mount process parameter feedforward adaptive control method based on solder paste printing detection according to claim 9, characterized in that: In step four, after the circuit board is printed with solder paste by the solder paste printer, the solder paste printing coordinate offset of the pads is detected by the solder paste inspection equipment. The initial pick-and-place machine production process parameters are then fed into a trained soldering offset prediction network model and a trained production process parameter compensation model to obtain adjusted production process parameters. The circuit board is then transferred to the pick-and-place machine to complete the surface soldering production of all mounting points on the current circuit board. The specific process is as follows: Step 4: After the solder paste is printed on the circuit board by the solder paste printer, it is sent to the solder paste inspection equipment (SPI) for inspection. The SPI equipment checks whether the solder paste at each mounting point on the current circuit board is misaligned. If the solder paste at each mounting point is not misaligned, a new circuit board is inspected; if the solder paste at each mounting point is misaligned, the SPI equipment obtains the result of the misalignment. The circuit board is then transferred to a pick-and-place machine for production. Step 42: Adjust the solder paste offset of each mounting point on the current circuit board. Process parameters and mounting coordinates Mounting height Placement of mounting components and air blowing delay Mounting speed of suction rod Length of various components ,width ,thickness Input a trained solder offset prediction network model, and the trained solder offset prediction network model outputs the coordinate offset of the corresponding mounting point after reflow soldering. ; based on Calculate the offset Euclidean distance : ; when Instead of performing step four and three, select the current production process parameters for surface welding production; when Proceed to step four three; Step 43: Place the current circuit board The placement position length direction coordinates in the initial production process parameters of the pick-and-place machine and width direction coordinates Mounting height Placement of mounting components and air blowing delay Mounting speed of suction rod and the length of various components ,width ,thickness The target policy network, which is input into the production process parameter compensation model, outputs the adjusted process parameter mounting coordinates after the model has been trained. Mounting height Placement of mounting elements and air blowing delay The upward speed of the suction rod at the mounting point ; Step 4: Place the adjusted process parameters onto the coordinate system. Mounting height Placement of mounting elements and air blowing delay The upward speed of the suction rod at the mounting point The circuit boards currently being produced by the pick-and-place machine are now being surface-mounted.

Images