PCBA process control method based on multi-modal perception and graph neural network

Abstract

The application discloses a PCBA process control method based on multi-modal perception and a graph neural network, comprising the following steps: collecting a multi-modal feature vector through a multi-modal sensor array, performing virtual-real synchronous calibration and mechanism model deviation analysis on a digital twin model, and obtaining a health residual feature; constructing a dynamic coupling graph based on the multi-modal feature vector and the health residual feature; inputting the dynamic coupling graph into a double-channel dynamic graph convolution network to perform state disturbance propagation and sensor credibility joint reasoning, and obtaining a comprehensive state representation vector; inputting the comprehensive state representation vector into a physical information neural network to perform a welding point forming quality prediction; when a quality deviation between a welding point forming quality index and a preset target quality index reaches a preset trigger threshold, performing reinforcement learning process parameter dynamic decision; and issuing a process parameter adjustment amount to a production line control system to realize adaptive control, significantly improve the flexibility of the production line, reduce the welding defect rate, and improve the yield.

Description

Technical Field

[0001] This application relates to the field of printed circuit board assembly process control technology, and in particular to a PCBA process control method based on multimodal perception and graph neural networks. Background Technology

[0002] Printed circuit board assembly (PCBA) is a core component of modern electronics manufacturing. Surface mount technology (SMT) production lines assemble electronic components through processes such as printing, placement, reflow soldering, and automated optical inspection. Reflow soldering, as a critical process for forming electrical connections, directly impacts solder joint quality through precise temperature profile control. However, SMT production lines are typically complex systems with multiple variables coupled and intertwined disturbances: fluctuations in ambient temperature and humidity, vibration and displacement caused by equipment aging, differences in incoming solder paste viscosity, and variations in preceding placement pressure all contribute to the final solder joints through a complex physical chain, leading to poor soldering consistency, high void ratios, and poor wetting, resulting in rework. Summary of the Invention

[0003] The main purpose of this application is to provide a PCBA process control method based on multimodal perception and graph neural network, which aims to solve the technical problems of poor welding consistency and high rework rate caused by environmental disturbances and equipment aging in a variable process environment.

[0004] To achieve the above objectives, this application proposes a PCBA process control method based on multimodal sensing and graph neural networks. The PCBA process control method based on multimodal sensing and graph neural networks includes: Multimodal feature vectors are collected during the operation of the surface mount technology production line by a multimodal sensor array, and the digital twin model of the production line is calibrated in a virtual-real synchronization manner and the mechanism model deviation analysis is performed based on the multimodal feature vectors to obtain health residual characteristics. Based on the multimodal feature vector and the health residual feature, a dynamic coupling graph is constructed, wherein the node set of the dynamic coupling graph includes physical device nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges. The dynamic coupling graph is input into a dual-channel dynamic graph convolutional network for joint inference of state perturbation propagation and sensor credibility to obtain a comprehensive state representation vector. The dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel is used to aggregate the neighborhood features of physical equipment nodes and workpiece nodes and infer the propagation path of state perturbation caused by changes in the state of upstream equipment to downstream process nodes. The second channel is used to aggregate the twin mapping edge features between each physical equipment node and the corresponding virtual reference node and to perform weighted correction on the original features of the physical equipment nodes by calculating the credibility weight of each sensor. The comprehensive state representation vector is input into the physical information neural network to predict the solder joint forming quality, thereby obtaining the solder joint forming quality index of the current batch of printed circuit boards. When the quality deviation between the weld joint forming quality index and the preset target quality index reaches the preset trigger threshold, the reinforcement learning controller is activated, and the reinforcement learning controller performs dynamic decision-making on the reinforcement learning process parameters to obtain the process parameter adjustment amount. The process parameter adjustment amounts are converted into executable control commands and sent to the production line control system to adjust the reflow soldering temperature profile and the chip mounter pressure parameters in real time, thereby achieving adaptive control of the PCBA process.

[0005] In one embodiment, the process of acquiring multimodal feature vectors during the operation of a surface mount technology production line using a multimodal sensor array, and performing virtual-real synchronous calibration and mechanism model deviation analysis on the digital twin model of the production line based on the multimodal feature vectors to obtain health residual characteristics, includes: A multimodal sensor array is deployed at key process nodes in the surface mount technology production line to collect multi-source heterogeneous data in real time during the production line operation. The multi-source heterogeneous data includes environmental temperature and humidity data, equipment vibration data, infrared thermal imaging data of the welding area, and equipment current data. The control command settings of each actuator are read in real time from the production line control system. The control command settings are time-paired with the multi-source heterogeneous data and feature extraction is performed to obtain a multimodal feature vector. A digital twin model of the production line is constructed. In the digital twin model, a corresponding virtual reference node is configured for each physical equipment node. The virtual reference node is embedded with a simplified mechanism model of the corresponding physical equipment. The simplified mechanism model includes a mounting force calculation model based on the pressure conversion relationship, a heat power calculation model based on Joule's law, and a wind speed calculation model based on the air volume and rotation speed relationship. The control command setpoint is input into the simplified mechanism model to determine the theoretical output reference value of each actuator at the current moment; Determine the residual sequence between the actual feedback value of the sensor and the theoretical output reference value at the same time, and perform exponential weighted moving average smoothing on the residual sequence to obtain the initial residual characteristics; A mapping function from physical space to virtual space is established, and the twin parameters of the digital twin model are calibrated online by minimizing the physical-virtual feature divergence to obtain the calibrated twin parameters; The theoretical output baseline value is corrected based on the calibrated twin parameters to obtain the corrected theoretical output baseline value, and the health residual characteristics are determined based on the corrected theoretical output baseline value.

[0006] In one embodiment, constructing a dynamic coupling graph based on the multimodal feature vector and the health residual features includes: Define physical device nodes, workpiece nodes, and virtual reference nodes configured for each physical device node, wherein the physical device nodes include pick-and-place machine nodes and independent temperature zone nodes of the reflow oven, and the workpiece nodes represent the current batch of printed circuit boards. The feature data at the corresponding position in the multimodal feature vector is used as the initial feature of the physical device node. The static information of the current batch of printed circuit boards is transformed by the embedding layer and used as the initial feature of the workpiece node. The health residual feature is used as the initial feature of the virtual reference node to generate an initial graph structure. The static information includes identification code, board size parameters and component distribution density data. Based on the material flow sequence of the surface mount technology production line, directed edges are constructed from upstream equipment nodes to workpiece nodes and from workpiece nodes to downstream equipment nodes as physical connection edges. These physical connection edges are used to describe the flow path of materials between printing, chip mounting, reflow soldering and automatic optical inspection processes. Obtain the physical location coordinates of each physical equipment node in the production line, calculate the Euclidean distance between any two physical equipment nodes, and when the Euclidean distance is less than a preset coupling distance threshold, construct an undirected edge between the two physical equipment nodes as an environmental coupling edge. The environmental coupling edge is used to describe the thermal radiation interference and vibration transmission coupling effect between the equipment. For each physical device node, a directed edge is constructed between it and the corresponding virtual reference node as a twin mapping edge, and the direction of the twin mapping edge is from the physical device node to the virtual reference node; A dynamic coupled graph is constructed based on the initial graph structure, the physical connection edges, the environmental coupling edges, and the twin mapping edges.

[0007] In one embodiment, constructing a dynamic coupling graph based on the initial graph structure, the physical connection edges, the environmental coupling edges, and the twin mapping edges includes: Collect virtual-real fusion state vector sequences from historical production cycles to construct a causal discovery dataset; A causal discovery algorithm based on conditional independence test is executed on the causal discovery dataset. The maximum lag order is set, the conditional independence test statistic for each entity device node pair is calculated, non-causal edges are deleted through directed separation criterion, and an undirected causal skeleton graph is constructed. A fraction-optimized causal structure learning algorithm is executed on the undirected causal skeleton graph to optimize the data fitting loss under the acyclicity constraint, outputting a directional causal graph, and using the directional causal graph as a causal structure prior. The thermal influence coefficient and vibration influence coefficient are determined based on environmental temperature and humidity data and equipment current data. The weighted sum of the thermal influence coefficient and the vibration influence coefficient is used as the initial dynamic weight of the environmental coupling edge. The causal structure prior is converted into a causal adjacency matrix, and the causal adjacency matrix is ​​fused with the initial dynamic weights of the environmental coupling edge using Hadamard to obtain the causal-enhanced environmental coupling edge. The initial graph structure is integrated with the physical connection edges, the causal-enhanced environmental coupling edges, and the twin mapping edges to obtain a dynamic coupling graph.

[0008] In one embodiment, the step of inputting the dynamic coupling graph into a dual-channel dynamic graph convolutional network for state perturbation propagation and joint inference of sensor credibility to obtain a comprehensive state representation vector includes: A dual-channel dynamic graph convolutional network is constructed, wherein the dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel adopts three stacked dynamic graph convolutional layers. Each dynamic graph convolutional layer consists of three sub-modules: neighborhood feature aggregation, node feature update, and temporal state propagation. The second channel adopts a multi-head graph attention mechanism for feature aggregation. Each attention head independently learns the attention weights between the physical device node and the virtual reference node. The dynamic coupling graph is input into the first and second channels of the dual-channel dynamic graph convolutional network, respectively. In the first channel, the neighborhood features of the physical device nodes and workpiece nodes in the dynamic coupling graph are aggregated by stacked dynamic graph convolutional layers, and the node hidden state is updated by gated loop unit to obtain the perturbation state embedding vector of each process node. The perturbation state embedding vector is used to characterize the cumulative process perturbation intensity and propagation path characteristics caused by the change of the state of the upstream physical device to the downstream process node. In the second channel, a multi-head graph attention mechanism is used to calculate the attention coefficient of the twin mapping edge between each entity device node and the corresponding virtual reference node in the dynamic coupling graph. The attention coefficient is used as the confidence weight of the sensor corresponding to the entity device node, and the original features of the entity device node are weighted and corrected based on the confidence weight to obtain the corrected state vector. The perturbation state embedding vector and the corrected state vector are concatenated along the feature dimension to generate a comprehensive state representation vector.

[0009] In one embodiment, in the first channel, neighborhood features of entity device nodes and workpiece nodes in the dynamically coupled graph are aggregated through stacked dynamic graph convolutional layers, and the node hidden states are updated through gated recurrent units to obtain the perturbation state embedding vector of each process node, including: In the first dynamic graph convolutional layer, for each entity device node and workpiece node in the dynamic coupling graph, the features of the upstream entity device node and downstream workpiece node in the corresponding first-order neighborhood are aggregated to obtain the aggregated neighborhood feature vector. The aggregated neighborhood feature vector is concatenated with the current feature of the corresponding node, input into the gated loop unit, the hidden state of the node is updated, and the updated node feature vector of the first layer is output. The updated node feature vector of the first layer is used as the input of the second dynamic graph convolutional layer to aggregate the feature information of nodes in the second-order neighborhood and obtain the expanded neighborhood feature vector. The expanded neighborhood feature vector is concatenated with the updated node feature vector of the first layer, and then input into the gated loop unit to update the hidden state of the node, and the updated node feature vector of the second layer is output. The updated node feature vector of the second layer is used as the input of the third dynamic graph convolutional layer to aggregate the feature information of nodes in the third-order neighborhood and obtain the deep neighborhood feature vector. The deep neighborhood feature vector is concatenated with the updated node feature vector of the second layer, and then input into the gated loop unit for the third time to update the hidden state of the node, so as to obtain the perturbation state embedding vector of each process node.

[0010] In one embodiment, in the second channel, a multi-head graph attention mechanism is used to calculate the attention coefficients of the twin mapping edges between each entity device node and its corresponding virtual reference node in the dynamic coupling graph. These attention coefficients are then used as the confidence weights of the sensors corresponding to the entity device nodes. Based on these confidence weights, the original features of the entity device nodes are weighted and corrected to obtain a corrected state vector, including: In the second channel, the twin mapping edge between each physical device node and the corresponding virtual reference node is extracted from the dynamic coupling graph, and the original feature vector of the physical device node and the theoretical baseline feature vector of the virtual reference node are obtained. The original feature vector of the physical device node is concatenated with the theoretical baseline feature vector of the virtual reference node to generate a joint feature representation of the twin mapping edge; Configure multiple parallel attention heads, where each attention head independently learns a set of query transformation matrices and key transformation matrices; The joint feature representation is linearly transformed by the query transformation matrix and the key transformation matrix respectively to obtain the query vector and key vector corresponding to each attention head; The scaled dot product attention score between the query vector and the key vector is determined, and after processing with a normalized exponential function, the initial attention coefficients corresponding to each attention head are obtained. The initial attention coefficients corresponding to each attention head are concatenated and dimensionality reduced to obtain the fused attention coefficients. The fused attention coefficients are used as the confidence weights of the sensors corresponding to the physical device nodes, and the original feature vectors of the physical device nodes are weighted and corrected based on the confidence weights to obtain the corrected state vectors.

[0011] In one embodiment, the step of inputting the comprehensive state representation vector into a physical information neural network to predict solder joint formation quality and obtain the solder joint formation quality index of the current batch of printed circuit boards includes: A physical information neural network is constructed, wherein the physical information neural network includes an input layer, multiple physical constraint hidden layers and a multi-task output layer connected in sequence. The physical constraint hidden layers embed partial differential equations of the physical mechanism of solder joint formation, and the physical mechanism partial differential equations include heat conduction equations, fluid dynamics equations and phase transition dynamics equations. The input layer maps the comprehensive state representation vector to an initial feature tensor of a preset dimension, and the initial feature tensor is passed to the first physical constraint hidden layer. In the physical constraint hidden layer, the temperature field distribution characteristics of the solder joint during the reflow soldering process are calculated based on the heat conduction equation, the fluid velocity field and pressure field characteristics of the solder paste in the molten state are calculated based on the fluid dynamics equation, and the phase transition ratio characteristics of the solder from solid to liquid and back to solid are calculated based on the phase transition dynamics equation. The temperature field distribution characteristics, fluid velocity field and pressure field characteristics, and phase transition ratio characteristics are used as physical constraint characteristics and fused with the initial feature tensor to obtain the fused physical enhancement feature tensor. The fused physical enhancement feature tensor is input to the multi-task output layer, wherein the multi-task output layer contains multiple parallel fully connected sub-networks, each of which corresponds to a solder joint forming quality sub-index, wherein the solder joint forming quality sub-index includes solder joint void rate, solder joint shear strength, solder joint wetting angle and solder joint lateral offset. Regression prediction is performed on the physical enhancement feature tensor through each fully connected sub-network to obtain the predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle and solder joint lateral offset for each solder joint on the current batch of printed circuit boards. The predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset are spliced ​​and weighted to generate the solder joint forming quality index of the current batch of printed circuit boards.

[0012] In one embodiment, when the quality deviation between the weld joint forming quality index and the preset target quality index reaches a preset trigger threshold, the reinforcement learning controller is activated, and dynamic decision-making of reinforcement learning process parameters is performed through the reinforcement learning controller to obtain the process parameter adjustment amount, including: The quality deviation is determined based on the weld joint forming quality index and the preset target quality index; The quality deviation is compared with a preset trigger threshold, and the reinforcement learning controller is activated when the quality deviation is greater than the preset trigger threshold. Construct a reinforcement learning state space, which includes a comprehensive state representation vector, quality bias, and health residual features; The actor network and critic network are constructed using a safety-constrained deep deterministic strategy gradient algorithm. The current state is sampled from the reinforcement learning state space and input into the actor network to obtain the process parameter adjustment action; The process parameter adjustment actions are input into the critic network for value evaluation, and the action value function value is obtained. The policy parameters of the actor network are updated based on the action value function value until the process parameter adjustment action output by the actor network satisfies the health constraints and minimizes the quality deviation, thus obtaining the optimized process parameter adjustment action. The process parameter adjustment amount is determined based on the optimized process parameter adjustment action, wherein the process parameter adjustment amount includes the reflow soldering temperature profile correction value and the pick-and-place machine pressure parameter correction value.

[0013] Furthermore, to achieve the above objectives, this application also proposes a collaborative design device for building orientation that integrates multi-objective optimization with urban microclimate coupling. This collaborative design device for building orientation coupling includes: The acquisition module is used to acquire multimodal feature vectors during the operation of the surface mount technology production line through a multimodal sensor array, and to perform virtual-real synchronous calibration and mechanism model deviation analysis on the digital twin model of the production line based on the multimodal feature vectors to obtain health residual characteristics. The construction module is used to construct a dynamic coupling graph based on the multimodal feature vector and the health residual feature, wherein the node set of the dynamic coupling graph includes physical device nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges. The inference module is used to input the dynamic coupling graph into a dual-channel dynamic graph convolutional network to perform joint inference of state perturbation propagation and sensor credibility, and obtain a comprehensive state representation vector. The dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel is used to aggregate the neighborhood features of physical equipment nodes and workpiece nodes and infer the propagation path of state perturbation caused by changes in the state of upstream equipment to downstream process nodes. The second channel is used to aggregate the twin mapping edge features between each physical equipment node and the corresponding virtual reference node and to perform weighted correction on the original features of the physical equipment nodes by calculating the credibility weight of each sensor. The prediction module is used to input the comprehensive state representation vector into the physical information neural network to predict the solder joint forming quality and obtain the solder joint forming quality index of the current batch of printed circuit boards. The decision module is used to activate the reinforcement learning controller when the quality deviation between the weld joint forming quality index and the preset target quality index reaches a preset trigger threshold, and to make dynamic decisions on the reinforcement learning process parameters through the reinforcement learning controller to obtain the process parameter adjustment amount. The adjustment module is used to convert the process parameter adjustment amount into executable control commands and send them to the production line control system to adjust the reflow soldering temperature profile and the placement machine pressure parameters in real time, so as to achieve adaptive control of the PCBA process.

[0014] The proposed technical solutions, including one or more, effectively block the propagation of erroneous data to the decision-making layer by introducing virtual reference nodes and their embedded simplified equipment mechanism models, and by using health residual feature calculation and dynamic correction mechanism of sensor credibility weights in a dual-channel dynamic graph convolutional network. This enables the production line to maintain the accuracy and stability of state perception when facing equipment aging characteristics drift, sudden changes in operating conditions caused by multi-product switching, and complex electromagnetic / thermal environmental interference, significantly improving the overall robustness of the flexible production line. By introducing a causal-enhanced dynamic coupling graph and using first-channel perturbation propagation inference to quantify the cascading impact of upstream process state fluctuations on downstream workpieces, and by combining physical information neural networks, the prediction of the solder joint forming process is realized. Furthermore, by using a reinforcement learning controller to dynamically adjust process parameters, adaptive control of the PCBA process is achieved, which can significantly reduce the soldering defect rate and improve the yield. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an embodiment of the PCBA process control method based on multimodal perception and graph neural networks provided in this application. Figure 2 This is a schematic diagram of the modular structure of a building orientation collaborative design device that integrates multi-objective optimization and urban microclimate coupling, as described in an embodiment of this application.

[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0021] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, or a building orientation collaborative design device that integrates multi-objective optimization and urban microclimate coupling. The following description uses a building orientation collaborative design device that integrates multi-objective optimization and urban microclimate coupling as an example to illustrate this embodiment and the following embodiments.

[0022] Based on this, embodiments of this application provide a PCBA process control method based on multimodal perception and graph neural networks, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the PCBA process control method based on multimodal perception and graph neural networks of this application.

[0023] In this embodiment, the PCBA process control method based on multimodal perception and graph neural networks includes steps S10~S60: Step S10: Collect multimodal feature vectors during the operation of the surface mount technology production line using a multimodal sensor array, and perform virtual-real synchronous calibration and mechanism model deviation analysis on the digital twin model of the production line based on the multimodal feature vectors to obtain health residual characteristics.

[0024] It should be noted that a multimodal sensor array refers to the deployment of various types of sensors, such as temperature and humidity sensors, vibration sensors, current sensors, and infrared thermal imaging sensors, at key process nodes in a surface mount technology production line. These sensors can collect multi-dimensional data in real time during the production line operation and form multimodal feature vectors.

[0025] A surface mount technology (SMT) production line refers to an automated production line that uses SMT technology to assemble printed circuit boards. It includes multiple continuous processes, such as printing, placement, reflow soldering, and automated optical inspection.

[0026] Multimodal feature vectors are feature vector representations with a unified dimension formed after preprocessing, feature extraction, and time alignment of collected multi-source heterogeneous data.

[0027] A production line digital twin model is a digital mirror model constructed in virtual space that is highly consistent with the physical production line in terms of geometric structure, physical characteristics, and behavioral logic. It consists of a geometric layer, a physical layer, a behavioral layer, and virtual reference nodes.

[0028] Simultaneous virtual-real calibration is a process of optimizing digital twin model parameters online by establishing a mapping function from physical space to virtual space, aiming to minimize the difference between physically measured characteristics and virtual model output characteristics. Mechanism model deviation analysis refers to the process of quantifying the difference between the actual state and ideal health state of equipment by comparing sensor measured values ​​with theoretical model output values ​​based on a simplified mechanism model embedded in virtual reference nodes.

[0029] Health residual features refer to the residual features after smoothing and twin parameter correction, which are used to quantify the degree of health deviation of each device node at the current moment.

[0030] In practical implementation, various types of sensors, such as temperature, humidity, vibration, current, and infrared thermal imaging sensors, are deployed at key process nodes of the surface mount technology production line to collect multi-dimensional data in real time during the production line operation. The collected multi-source heterogeneous data undergoes preprocessing, including data cleaning, noise reduction, and normalization, to eliminate outliers and dimensional differences. Feature extraction algorithms, such as principal component analysis and independent component analysis, are used to extract representative features from the preprocessed data. Simultaneously, time alignment is performed to ensure that data collected by different sensors correspond consistently across time. Finally, these processed data are integrated into a unified-dimensional multimodal feature vector. Subsequently, using the constructed mapping function from physical space to virtual space, the multimodal feature vectors are input into the production line digital twin model. By optimizing the model parameters online, the output features of the virtual model are made as close as possible to the measured physical features, achieving synchronous calibration between the virtual and physical worlds. Furthermore, based on the simplified mechanism model embedded in the virtual reference node, the measured values ​​of the sensors are compared with the theoretical output values ​​of the model to quantify the difference between the actual state and the ideal health state of the equipment, obtaining preliminary residual features. Finally, the preliminary residual features are smoothed to eliminate the influence of random noise, and the accuracy of the residual features is further improved through twin parameter correction, thereby obtaining health residual features that can quantify the degree of health deviation of each equipment node at the current moment.

[0031] In one feasible implementation, step S10 may include: deploying a multimodal sensor array at key process nodes of the surface mount technology production line to collect multi-source heterogeneous data during the production line operation in real time, wherein the multi-source heterogeneous data includes environmental temperature and humidity data, equipment vibration data, infrared thermal imaging data of the welding area, and equipment current data; reading control command setpoints of each actuator in real time from the production line control system, performing time pairing of the control command setpoints with the multi-source heterogeneous data and extracting features to obtain a multimodal feature vector; constructing a digital twin model of the production line, in which a corresponding virtual reference node is configured for each physical equipment node, and the virtual reference node embeds a simplified mechanism model of the corresponding physical equipment, the simplified mechanism model including a pressure conversion mechanism based on air pressure. The system employs a mounting force calculation model, a heat power calculation model based on Joule's law, and a wind speed calculation model based on the relationship between airflow and rotational speed. The control command setpoints are input into the simplified mechanism model to determine the theoretical output benchmark value of each actuator at the current moment. A residual sequence between the actual sensor feedback value and the theoretical output benchmark value at the same moment is determined, and the residual sequence is smoothed using an exponentially weighted moving average to obtain initial residual characteristics. A mapping function from physical space to virtual space is established, and the twin parameters of the digital twin model are calibrated online by minimizing the physical-virtual feature divergence to obtain calibrated twin parameters. Based on the calibrated twin parameters, the theoretical output benchmark value is corrected to obtain a corrected theoretical output benchmark value, and health residual characteristics are determined based on the corrected theoretical output benchmark value.

[0032] It should be noted that, in this embodiment, the key process nodes are the printing process, the placement process, the reflow soldering process, and the automatic optical inspection process. The sensors deployed at these locations can comprehensively capture the operating status of the production line. The multimodal sensor array consists of a temperature and humidity sensor, a triaxial vibration accelerometer, an infrared thermal imager, and a Hall current sensor. Among them, the temperature and humidity sensor is deployed at the air inlet and outlet of each temperature zone of the reflow oven to monitor the ambient temperature and humidity and their rate of change; the triaxial vibration accelerometer is deployed at the root of the nozzle of each placement head of the placement machine, with a sampling frequency of 2000Hz, to capture the high-frequency vibration characteristics of placement impact; the infrared thermal imager is deployed 30 cm directly above the outlet of the reflow oven, with a frame rate of 50fps, to collect the temperature field distribution of the entire board at the moment the PCB board leaves the oven; the Hall current sensor is deployed at the spindle drive motor of the placement machine, the reflow heating wire circuit, and the cooling fan motor, with a sampling frequency of 100Hz, to monitor the equipment current fluctuation.

[0033] All sensors are triggered by the same hardware signal, such as a synchronization pulse generated by an FPGA, to achieve microsecond-level synchronous acquisition. Each data point is then timestamped with a uniform microsecond-level time stamp, thus obtaining multi-source heterogeneous data during production line operation, including environmental temperature and humidity data, equipment vibration data, infrared thermal imaging data of the soldering area, and equipment current data. Simultaneously, the control instruction registers in the production line's programmable logic controller (PLC) are read in real-time via the industrial Ethernet protocol. This allows for the parsing and acquisition of the Z-axis pressure setting, placement speed setting, and material handling height setting for each placement head of the pick-and-place machine, as well as the temperature setting, chain speed setting, and fan speed setting for each zone of the reflow oven. These control instruction settings are also timestamped with high precision and time-aligned with the sensor data to ensure that the control instructions and sensor feedback values ​​at the same moment are matched.

[0034] To improve data quality, it is necessary to preprocess and extract features from the raw data, including ambient temperature and humidity data, equipment vibration data, infrared thermal imaging data of the welding area, and equipment current data. For example, wavelet threshold denoising and Kalman filtering are performed on the collected equipment vibration data to extract time-domain statistics and frequency domain main frequency energy ratio, resulting in processed vibration features. Moving average filtering is performed on the collected ambient temperature and humidity data to extract the rate of change and second-order difference features, resulting in processed temperature and humidity features. Non-uniformity correction and super-resolution reconstruction are performed on the collected infrared thermal imaging data of the welding area to extract the average temperature, temperature gradient, and cooling rate feature vectors of the solder pad area, resulting in processed thermal imaging features. Fundamental amplitude, harmonic distortion rate, and peak starting current are extracted from the collected equipment current data, resulting in processed current features. The processed vibration features, processed temperature and humidity features, processed thermal imaging features, and processed current features are then stitched together after timestamp alignment to generate a multimodal feature vector.

[0035] A digital twin model of the production line, comprising geometric, physical, and behavioral layers, is constructed. First, based on the computer-aided design drawings of the production line, a 3D model with a geometric accuracy of 0.1mm is built, and a quadratic error metric algorithm is used for lightweight processing to ensure a real-time simulation frame rate of no less than 30fps. A layered simulation kernel is then built upon this lightweight geometric model: the heat conduction layer uses the finite volume method to discretize the solution domain; the mechanical dynamics layer uses the modal superposition method for order reduction; and the fluid layer uses the lattice Boltzmann method to simulate the airflow within the reflow oven. These three layers are bidirectionally coupled through boundary conditions to form a multiphysics kernel. Furthermore, the actual PLC program is analyzed and converted into an equivalent state machine model to achieve software-in-the-loop simulation of the control logic, generating the digital twin model of the production line. In the digital twin model, a corresponding virtual reference node is configured for each physical device node, such as a pick-and-place machine or a reflow soldering zone. The virtual reference node embeds a simplified mechanistic model of the device. For example, the simplified mechanistic model for the pick-and-place head is a linear model based on air pressure-to-pressure conversion, calculating the theoretical placement force based on the pressure setpoint; the simplified mechanistic model for the reflow soldering heating wire is a thermal power model based on Joule's law, calculating the theoretical heating current based on the temperature setpoint and the PID controller output; and the simplified mechanistic model for the cooling fan is an empirical model based on airflow and rotation speed, calculating the theoretical wind speed based on the rotation speed setpoint. The parameters of these simplified mechanistic models are initialized through calibration experiments during the equipment commissioning phase.

[0036] The control command setpoints read in real time are input into the simplified mechanism model to obtain the theoretical output reference values ​​of each actuator at the current moment. The residuals between the actual feedback values ​​of the sensors and the theoretical output reference values ​​at the same moment are calculated, and the residual sequence is smoothed by exponential weighted moving average to obtain the initial residual characteristics.

[0037] To compensate for the deviation between the model and the actual production line, a mapping function from physical space to virtual space is established. By minimizing the divergence between the multimodal feature vectors of the physical space and the output features of the virtual reference node (e.g., KL divergence or Euclidean distance), the twin parameters of the digital twin model, such as the thermal conductivity correction factor and damping coefficient, are optimized online. The optimized twin parameters are then used to correct the theoretical output baseline value, resulting in a corrected theoretical output baseline value. Based on this, the residuals are recalculated, ultimately yielding more accurate health residual characteristics.

[0038] Simultaneously, the health index of each sensor is calculated. For example, based on residual statistics or model confidence, a gating vector is constructed. The multimodal feature vector of the physical space and the health residual features of the virtual reference node are adaptively fused through the gating vector to generate a virtual-real fusion state vector. When the sensor health index is lower than a preset threshold or the health residual features exceed a preset range, the measured weight of that dimension is automatically reduced and the theoretical weight of the virtual reference node is increased.

[0039] Step S20: Based on the multimodal feature vector and the health residual feature, construct a dynamic coupling graph, wherein the node set of the dynamic coupling graph includes physical device nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges.

[0040] It should be noted that a dynamic coupling graph is a graph structure model that can reflect the complex interactive relationships between equipment, workpieces and virtual models in real time during the operation of a production line. In this model, the node set includes physical equipment nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges.

[0041] Physical device nodes represent digital nodes of physically existing, independently functional equipment units in a surface mount technology production line. For example, a pick-and-place machine node refers to each pick-and-place machine as a whole, encompassing all its placement heads, drive shafts, and other sub-components. Similarly, independent temperature zone nodes in a reflow oven refer to each independent temperature control zone of the reflow oven as a separate physical device node, rather than treating the entire oven as a single node. This is because different temperature zones have independent effects on soldering quality and are thermally coupled to each other. Feature data at corresponding locations is extracted from multimodal feature vectors as the initial features for the physical device nodes. For pick-and-place machine nodes, this includes vibration features, current features, and placement pressure-related features; for temperature zone nodes, it includes temperature features, heating current features, and temperature and humidity features.

[0042] Workpiece nodes represent digital nodes of a batch of printed circuit boards currently flowing on the production line. Each workpiece node corresponds to a batch of PCBs being processed. Its initial characteristics are obtained by converting the static information of the current batch of printed circuit boards through an embedding layer, including identification codes, board size parameters, and component distribution density data.

[0043] A virtual reference node is a virtual digital node configured for each physical device node, embedding a simplified mechanism model. It represents the theoretical benchmark of the device under ideal health conditions. The health residual characteristics serve as the initial characteristics of the virtual reference node, reflecting the degree of deviation between the actual state and the theoretical benchmark.

[0044] Physical connection edges are directed edges constructed based on the actual physical flow sequence of materials on the production line, describing the path of materials flowing from upstream equipment to downstream equipment. Environmental coupling edges are undirected edges constructed based on the physical distance between equipment and real-time environmental data, describing environmental coupling interference between equipment, such as heat radiation and vibration transmission. Twin mapping edges are directed edges connecting each physical equipment node to its corresponding virtual reference node, describing the mapping relationship between the entity and its virtual mirror.

[0045] In the specific implementation, based on the feature data of the corresponding positions of each physical device node in the multimodal feature vector, initial features are assigned to the physical device nodes. For example, vibration features, current features, and mounting pressure-related features are assigned to the pick-and-place machine node, and temperature features, heating current features, and temperature and humidity features are assigned to the temperature zone node. For the workpiece node, the static information of the printed circuit board batch currently flowing on the production line, such as identification codes, board size parameters, and component distribution density data, is converted by the embedding layer and used as its initial features. The virtual reference node uses health residual features as its initial features to reflect the degree of deviation between the actual state and the theoretical benchmark, and then constructs... Physical connection edges are formed by determining the path of materials from upstream to downstream equipment based on the actual physical flow sequence of materials on the production line, thus creating directed edges. When constructing environmental coupling edges, the environmental coupling interference between equipment, such as heat radiation and vibration transmission, is analyzed based on the physical distance between equipment and real-time environmental data, thus constructing undirected edges. For twin mapping edges, each physical equipment node is connected to its corresponding virtual reference node to form a directed edge, describing the mapping relationship between the entity and the virtual mirror, thereby completing the construction of a dynamic coupling graph. This graph can reflect the complex interaction relationships between equipment, workpieces, and virtual models during the operation of the production line in real time.

[0046] In one feasible implementation, step S20 may include: defining physical device nodes, workpiece nodes, and virtual reference nodes configured for each physical device node, wherein the physical device nodes include pick-and-place machine nodes and independent temperature zone nodes of reflow ovens, and the workpiece nodes represent the current batch of printed circuit boards; using the feature data at the corresponding position in the multimodal feature vector as the initial features of the physical device node, using the static information of the current batch of printed circuit boards after being transformed by the embedding layer as the initial features of the workpiece node, and using the health residual features as the initial features of the virtual reference node to generate an initial graph structure, wherein the static information includes identification codes, board size parameters, and component distribution density data; based on the material flow sequence of the surface mount technology production line, constructing a path from upstream device nodes to workpiece nodes and from workpiece nodes... Directed edges pointing from nodes to downstream device nodes serve as physical connection edges, which describe the material flow path between printing, surface mount, reflow soldering, and automated optical inspection processes. The physical coordinates of each physical device node in the production line are obtained, and the Euclidean distance between any two physical device nodes is calculated. When the Euclidean distance is less than a preset coupling distance threshold, an undirected edge is constructed between the two physical device nodes as an environmental coupling edge, which describes the coupling effect of thermal radiation interference and vibration transmission between devices. A directed edge is constructed between each physical device node and its corresponding virtual reference node, serving as a twin mapping edge, with the direction of the twin mapping edge pointing from the physical device node to the virtual reference node. A dynamic coupling graph is constructed based on the initial graph structure, the physical connection edges, the environmental coupling edges, and the twin mapping edges.

[0047] It should be noted that by defining different types of nodes, such as physical equipment nodes, workpiece nodes, and virtual reference nodes, and assigning them corresponding initial features, the state information of each component in the production line can be comprehensively and accurately described, thereby generating an initial graph structure. Physical equipment nodes cover key equipment, such as pick-and-place machine nodes and independent temperature zone nodes of reflow ovens. Their initial features are extracted from multimodal feature vectors, ensuring the richness and accuracy of the features. Workpiece nodes correspond to the batch of printed circuit boards being processed. Their initial features are obtained after embedding layer transformation and include key information such as identification codes, board size parameters, and component distribution density data. Virtual reference nodes use health residual features as initial features.

[0048] In practical implementation, based on the material flow sequence of the surface mount technology production line, a path planning algorithm is used to automatically construct physical connection edges to describe the flow path of materials between printing, chip mounting, reflow soldering, and automatic optical inspection processes, which is the main channel for the transmission of process disturbances.

[0049] When constructing environmental coupling edges, the physical location coordinates between devices and real-time environmental data are used. Distance calculation and interference analysis algorithms are used to determine the environmental coupling relationship between devices. When the Euclidean distance between two physical device nodes is less than the preset coupling distance threshold, an undirected edge is constructed between them to describe the coupling effect of thermal radiation interference and vibration transmission between devices.

[0050] For the construction of twin mapping edges, a directed edge is built between each physical device node and its corresponding virtual reference node. The direction of this edge is from the physical device node to the virtual reference node, which is used to describe the mapping relationship between the physical device and its mirror image in the virtual space. In this way, twin mapping edges not only connect the physical world and the virtual world, but also enable the model to reflect the state changes of the physical device in real time and transmit these changes to the virtual reference node, thereby realizing dynamic interaction and synchronous updates between the virtual and the real.

[0051] By introducing a causal reasoning mechanism, the environmental coupling edges are causally enhanced, and then a dynamic coupling graph is constructed by combining the initial graph structure, physical connection edges, and twin mapping.

[0052] In one feasible implementation, the construction of a dynamic coupling graph based on the initial graph structure, the physical connection edges, the environmental coupling edges, and the twin mapping edges includes: collecting virtual-real fusion state vector sequences from historical production cycles to construct a causal discovery dataset; executing a causal discovery algorithm based on conditional independence testing on the causal discovery dataset, setting a maximum lag order, calculating the conditional independence test statistic for each entity device node pair, deleting non-causal edges using a directed separation criterion, and constructing an undirected causal skeleton graph; and executing a causal structure learning algorithm based on score optimization on the undirected causal skeleton graph to optimize the data fitting under acyclicity constraints. The loss is combined to output a directed causal graph, which is then used as the causal structure prior. Based on environmental temperature and humidity data and equipment current data, thermal influence coefficients and vibration influence coefficients are determined, and the weighted sum of these coefficients is used as the initial dynamic weights of the environmental coupling edges. The causal structure prior is converted into a causal adjacency matrix, and the causal adjacency matrix is ​​fused with the initial dynamic weights of the environmental coupling edges using Hadamard fusion to obtain causally enhanced environmental coupling edges. The initial graph structure is then integrated with the physical connection edges, the causally enhanced environmental coupling edges, and the twin mapping edges to obtain a dynamic coupling graph.

[0053] It should be noted that the virtual-real fusion state vector sequence is formed by fusing multimodal feature vectors and health residual features from the historical generation cycle. It records the interaction state information between equipment, workpieces, and virtual models at different historical moments in the production line. By processing the virtual-real fusion state vector sequence into a format suitable for causal discovery, including standardization of each variable and determination of the time lag window, a causal discovery dataset is obtained. This dataset contains the temporal dependencies and potential causal relationships between various variables during the production line operation.

[0054] Causal discovery algorithms based on conditional independence tests can be the Peter-Clark algorithm or its variants. This algorithm calculates conditional independence test statistics for each pair of entity device nodes in the causal discovery dataset by setting a maximum lag order. It then uses a directed separation criterion to remove non-causal edges, thereby constructing an undirected causal skeleton graph. This process can uncover the potential causal relationship structure between variables in production line operation. The conditional independence test statistic refers to calculating partial correlation coefficients or mutual information statistics for any two variables, given a subset of other variables, to test their conditional independence. If the statistic value is less than a preset threshold, the two variables are considered independent under that condition, and there is no direct causal relationship. Therefore, the corresponding edges are removed, resulting in an undirected causal skeleton graph.

[0055] Fraction-based causal structure learning algorithms can employ the NOTEARS algorithm or its variants, transforming causal structure learning into a continuous optimization problem. Under acyclicity constraints, this algorithm finds the optimal causal structure by optimizing the data fitting loss function, such as the Bayesian information criterion or the Akaike information criterion, outputting a directional causal graph, which serves as the prior of the causal structure.

[0056] The thermal influence coefficient is determined based on environmental temperature and humidity data, that is, the intensity of thermal radiation interference is calculated according to the temperature difference and distance between equipment, as shown in the following formula: in, This represents the thermal influence coefficient between device i and device j at time t. and Let d represent the positive deviations of the surface temperatures of devices i and i from the ambient temperature, respectively. ij Let represent the Euclidean distance between device i and device j. It is a numerical stability constant. It is the thermal conductivity coefficient.

[0057] The vibration influence coefficient is determined based on equipment current data, that is, the vibration transmission influence is calculated based on the vibration intensity reflected by equipment current fluctuations, as shown in the following formula: in, This represents the vibration influence coefficient between device i and device j at time t. and Let represent the effective current values ​​of device i and device j at time t, respectively. It is the basic vibration transmission coefficient between equipment i and equipment j. Let represent the Euclidean distance between device i and device j. It is also a numerical stability constant.

[0058] The calculated thermal influence coefficient and vibration influence coefficient are weighted and summed to obtain the initial dynamic weight of the environmental coupling edge. This weight can comprehensively reflect the degree of influence of thermal radiation interference and vibration transmission coupling effect between devices on the environmental coupling edge.

[0059] The causal structure prior is transformed into a causal adjacency matrix, which can intuitively represent the causal relationship direction between each entity device node. The causal adjacency matrix is ​​then fused with the initial dynamic weights of the environmental coupling edge using Hadamard, i.e., corresponding element-wise multiplication, to obtain the causally enhanced environmental coupling edge. This fusion process organically combines the causal structure information with the dynamic weights of the environmental coupling, so that the environmental coupling edge can not only reflect the physical interference relationship between devices, but also reflect the potential causal relationship between various variables in the production line operation, thereby enhancing the expressive power and accuracy of the environmental coupling edge.

[0060] By integrating the initial graph structure with physical connection edges, causal-enhanced environmental coupling edges, and twin mapping edges, a dynamic coupled graph is obtained. In the dynamic graph structure, physical connection edges carry material flow information, environmental coupling edges carry environmental interference information after causal enhancement, and twin mapping edges carry the mapping relationship between entity nodes and virtual reference nodes.

[0061] Step S30: Input the dynamic coupling graph into a dual-channel dynamic graph convolutional network to perform state perturbation propagation and sensor credibility joint inference to obtain a comprehensive state representation vector. The dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel is used to aggregate the neighborhood features of physical equipment nodes and workpiece nodes and infer the propagation path of state perturbation caused by changes in the state of upstream equipment to downstream process nodes. The second channel is used to aggregate the twin mapping edge features between each physical equipment node and the corresponding virtual reference node and to perform weighted correction on the original features of the physical equipment nodes by calculating the credibility weight of each sensor.

[0062] It should be noted that the dual-channel dynamic graph convolutional network is a graph neural network architecture that includes two parallel processing channels. These channels perform feature extraction and inference for different types of edges and nodes in the dynamically coupled graph, and finally fuse the outputs of the two channels into a comprehensive state representation.

[0063] State disturbance propagation refers to the process in the first channel where process disturbances, such as abnormal patch pressure and temperature fluctuations, are simulated through the information aggregation mechanism of graph neural networks and transmitted from upstream equipment to downstream equipment along physical connection edges and environmental coupling edges. Ultimately, the cumulative disturbance intensity and propagation path characteristics experienced by each equipment node are quantified.

[0064] Sensor credibility joint inference refers to the process in the second channel of dynamically calculating the credibility weight of each sensor by comparing the measured characteristics of the physical device node with the theoretical baseline characteristics of the corresponding virtual reference node, and then using the weight to perform weighted correction on the measured data.

[0065] The dual-channel dynamic graph convolutional network comprises a first channel and a second channel. The first channel focuses on the physical connection edges between physical equipment nodes and workpiece nodes. It aggregates neighborhood features through dynamic graph convolution operations and infers the perturbation propagation path of upstream equipment state changes on downstream process nodes based on historical data and real-time state inference. This channel employs an attention mechanism to dynamically allocate neighborhood node weights, strengthening the dependencies between key processes. A gated recurrent unit (GRU) is introduced to capture temporal features, ensuring the temporal consistency of the perturbation propagation path. The second channel focuses on the twin mapping edges between physical equipment nodes and virtual reference nodes. By calculating the deviation between sensor measurements and the health residual features of the virtual reference node, a sensor reliability assessment model is constructed. This model, based on a Bayesian inference framework, dynamically updates sensor weights using historical fault data. Finally, it performs reliability-weighted correction on the original features of physical equipment nodes, suppressing the interference of abnormal sensor data on state representation. The outputs of the two channels are fused into a comprehensive state representation vector through a concatenation operation. This retains the temporal dynamics of physical perturbation propagation while incorporating the calibration capability of virtual space for sensor reliability, providing a robust state-aware foundation for subsequent process control decisions.

[0066] Step S40: Input the comprehensive state representation vector into the physical information neural network to predict the solder joint forming quality and obtain the solder joint forming quality index of the current batch of printed circuit boards.

[0067] It should be noted that Physical Information Neural Network (PINN) is a deep learning method that embeds physical mechanism equations, such as partial differential equations, as prior knowledge into the neural network architecture. Unlike purely data-driven black-box models, PINN not only fits the data during training but also forces the network output to satisfy physical laws, thereby improving the model's generalization ability and interpretability. The Physical Information Neural Network in this embodiment includes an input layer, multiple physically constrained hidden layers, and a multi-task output layer connected in sequence. The physically constrained hidden layers include a heat conduction equation constraint layer for calculating the temperature field distribution characteristics, a fluid dynamics equation constraint layer for calculating the velocity and pressure field characteristics, and a phase change dynamics equation constraint layer for calculating the phase change ratio characteristics.

[0068] Understandably, the solder joint forming quality index of the current batch of printed circuit boards is a numerical value that quantitatively characterizes the quality of solder joints on the printed circuit board and is used to evaluate the quality of the soldering process. It includes solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset.

[0069] In one feasible implementation, step S40 may include: constructing a physical information neural network, wherein the physical information neural network includes an input layer, multiple physical constraint hidden layers, and a multi-task output layer connected in sequence, wherein the physical constraint hidden layers embed a physical mechanism partial differential equation for solder joint formation, the physical mechanism partial differential equation including a heat conduction equation, a fluid dynamics equation, and a phase transition dynamics equation; mapping the comprehensive state representation vector to an initial feature tensor of a preset dimension through the input layer, and passing the initial feature tensor to the first physical constraint hidden layer; in the physical constraint hidden layer, calculating the temperature field distribution characteristics of the solder joint during the reflow soldering process based on the heat conduction equation, calculating the fluid velocity field and pressure field characteristics of the solder paste in the molten state based on the fluid dynamics equation, calculating the phase transition ratio characteristics of the solder from solid to liquid and back to solid based on the phase transition dynamics equation, and transmitting the temperature field distribution characteristics, fluid velocity field, and pressure field characteristics to the input layer. The phase transition ratio feature is used as a physical constraint feature and fused with the initial feature tensor to obtain a fused physical enhancement feature tensor. The fused physical enhancement feature tensor is input to the multi-task output layer, which contains multiple parallel fully connected sub-networks. Each fully connected sub-network corresponds to a solder joint forming quality sub-index, which includes solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset. Regression prediction is performed on the physical enhancement feature tensor through each fully connected sub-network to obtain the predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset for each solder joint on the current batch of printed circuit boards. The predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset are concatenated and weighted to generate the solder joint forming quality index for the current batch of printed circuit boards.

[0070] It should be noted that the Physical Information Neural Network (PIN) is a deep learning model that embeds physical mechanism equations as prior knowledge into a neural network architecture. It consists of an input layer, multiple physical constraint hidden layers, and a multi-task output layer connected sequentially. The input layer receives a comprehensive state representation vector from a dual-channel graph network and maps it to an initial feature tensor of uniform dimension through a linear transformation. The physical constraint hidden layers contain parallel computational modules for multiple physical mechanism equations, including heat conduction equation modules, fluid dynamics equation modules, and phase transition dynamics equation modules. These modules calculate the corresponding physical field features and fuse them with data features. For example, the heat conduction equation constraint layer calculates temperature field distribution features, the fluid dynamics equation constraint layer calculates velocity and pressure field features, and the phase transition dynamics equation constraint layer calculates phase transition ratio features. The multi-task output layer contains multiple parallel fully connected sub-networks, each responsible for predicting a specific weld joint forming quality sub-indicator.

[0071] The thermal conduction equation constraint refers to the fact that during reflow soldering, heat conduction in the PCB board, pads, and solder follows Fourier's law of thermal conductivity. The fluid dynamics equation constraint refers to the fact that after the solder paste melts in reflow soldering, it becomes a fluid, and its flow behavior is controlled by the Navier-Stokes equations, affecting the wetting angle and void formation of the solder joint. The phase transformation kinetics equation constraint refers to the fact that the solder undergoes a solid-to-liquid-to-solid phase transformation during reflow soldering; the growth of intermetallic compounds follows the laws of phase transformation kinetics, directly affecting the shear strength of the solder joint.

[0072] In each physical constraint hidden layer, data features are fused with physical features calculated from various physical equations to generate a physical augmentation feature tensor. This can be achieved by using a linear transformation after concatenation, or by injecting physical features into the main network stream using residual connections. This allows the network to learn data-driven features while continuously being guided by physical laws, achieving a deep fusion of data and knowledge.

[0073] The multi-task output layer contains multiple parallel fully connected sub-networks. Each sub-network is responsible for predicting a specific solder joint formation quality sub-indicator, including solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset. Among them, solder joint voids refer to the voids formed inside the solder joint due to the failure of gases generated by flux volatilization in the solder paste to escape in time during the reflow soldering process. The solder joint void rate is the proportion of the void area inside the solder joint to the total area of ​​the solder joint, reflecting the solder joint density, as shown in the following formula: in, This represents the void ratio of solder joints, where N is the total number of solder joints. Let be the area of ​​the cavity inside the i-th solder joint. Let be the total area of ​​the i-th solder joint.

[0074] The shear strength of a solder joint is determined by the thickness and morphology of the intermetallic compound layer. If the intermetallic compound layer is too thin, the bonding strength will be insufficient; if it is too thick, the brittleness will increase. The shear strength of a solder joint is the maximum shear force that the solder joint can withstand, reflecting its mechanical strength.

[0075] The solder joint wetting angle is the ability of the solder to spread on the surface of the solder pad after melting. It is determined by surface tension and interfacial reaction. An excessively large wetting angle indicates poor wetting, while an excessively small wetting angle may indicate excessive solder spreading. The solder joint wetting angle is the contact angle between the solder and the solder pad, reflecting the wetting performance.

[0076] Lateral offset of solder joints is the deviation of component placement during the surface mount process, or the self-alignment offset of components caused by the surface tension of molten solder during reflow soldering. It is the positional deviation between the component pin and the center of the pad, reflecting the mounting accuracy.

[0077] The system integrates the predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset for each solder joint on the current batch of printed circuit boards into a comprehensive quality index through weighted summation. The weight of each sub-index is dynamically adjusted according to process control requirements. For example, the weight of shear strength can be increased for high-reliability products, while the wetting angle control is emphasized for high-frequency signal transmission scenarios. The final solder joint forming quality index is output in numerical form, typically ranging from 0 to 100 points, with higher scores indicating better soldering quality. This index also supports multi-dimensional analysis and can display the distribution of each sub-index through a visual interface, helping engineers quickly locate process defect types. For example, when the predicted void rate exceeds a preset threshold, the system automatically triggers reflow soldering temperature profile optimization suggestions; if the lateral offset remains abnormal, it prompts a check of the pick-and-place machine calibration status. The calculation process of the quality index adopts a dynamic error compensation mechanism, automatically correcting model drift by comparing with historical batch data to ensure the stability of the evaluation results. In the multi-task output layer, each sub-network is equipped with an online learning module, which continuously updates the prediction parameters based on newly acquired welding data, enabling the model to adapt to the process characteristics of different PCB models. The final output comprehensive quality index is not only used for the current batch inspection, but can also serve as input to the digital twin system to drive the virtual production line to perform closed-loop optimization of process parameters.

[0078] Step S50: When the quality deviation between the weld joint forming quality index and the preset target quality index reaches the preset trigger threshold, the reinforcement learning controller is activated, and the reinforcement learning controller performs dynamic decision-making on the reinforcement learning process parameters to obtain the process parameter adjustment amount.

[0079] It should be noted that the reinforcement learning controller is a decision-making framework based on the interaction between the agent and the environment. It learns the optimal control strategy through a trial-and-error mechanism. In this embodiment, the controller uses the quality deviation between the comprehensive quality index and the preset target as the state input and uses a deep deterministic policy gradient algorithm to generate the process parameter adjustment amount.

[0080] Reinforcement learning-based dynamic decision-making for process parameters refers to a process that uses a reinforcement learning framework as its core, the current production line status as input, process parameter adjustments as output, and quality improvement as the reward objective. It achieves adaptive process control through the collaborative optimization of policy networks and value networks.

[0081] In its implementation, the controller comprises an actor network and a critic network. The actor network receives the current state and outputs process adjustments in the continuous action space, such as reflow soldering temperature and chip placement pressure. The critic network evaluates the value of the actions output by the actor network and optimizes the actor network's parameters through gradient backpropagation. During training, an experience replay mechanism is used to store historical interaction data, and random sampling is performed to update the network, breaking data correlation. Simultaneously, a target network stabilization training process is introduced to prevent policy oscillations. The reinforcement learning controller also integrates a safety constraint module. When the predicted adjustment exceeds the physical limits of the equipment or process specifications, a boundary protection mechanism is automatically triggered to limit the actions within a safe range. For example, if the temperature adjustment output by the actor network exceeds the maximum allowable value of the reflow oven, the system will correct it to the maximum safe temperature and record the abnormal event for engineers to analyze. Furthermore, the controller supports dynamic reward function design, which can flexibly adjust the reward weight according to the quality requirements of different production stages. For example, in the sample production stage, the focus is on solder joint shear strength, while in the mass production stage, priority is given to controlling void ratio and wetting angle. The total reward value is calculated through weighted summation to guide the agent to learn a strategy that conforms to actual production goals.

[0082] In one feasible implementation, step S50 may include: determining a quality deviation based on the solder joint forming quality index and a preset target quality index; comparing the quality deviation with a preset trigger threshold, and activating a reinforcement learning controller when the quality deviation is greater than the preset trigger threshold; constructing a reinforcement learning state space, the reinforcement learning state space including a comprehensive state representation vector, quality deviation, and health residual features; constructing an actor network and a critic network using a safety-constrained deep deterministic policy gradient algorithm; sampling the current state from the reinforcement learning state space, inputting the current state into the actor network to obtain a process parameter adjustment action; inputting the process parameter adjustment action into the critic network for value evaluation to obtain an action value function value; updating the policy parameters of the actor network based on the action value function value until the process parameter adjustment action output by the actor network satisfies the health constraint conditions and minimizes the quality deviation, obtaining an optimized process parameter adjustment action; determining the process parameter adjustment amount based on the optimized process parameter adjustment action, wherein the process parameter adjustment amount includes a reflow soldering temperature profile correction value and a pick-and-place machine pressure parameter correction value.

[0083] It should be noted that the preset target quality index is a comprehensive threshold calculated by weighting the standard values ​​of weld formation quality indicators such as weld void rate, weld shear strength, weld wetting angle, and weld lateral offset. The quality deviation is the difference between the actual measured weld formation quality index and the preset target quality index, used to quantify the degree of deviation between the current welding quality and the ideal state. When the quality deviation exceeds the preset trigger threshold, it indicates a significant abnormality in the current welding process, requiring immediate activation of the reinforcement learning controller for dynamic adjustment.

[0084] The reinforcement learning state space is the input set of the reinforcement learning algorithm, containing three dimensions: a comprehensive state representation vector, quality deviation, and health residual features. The comprehensive state representation vector is extracted by a dual-channel graph network, integrating multimodal data such as temperature, pressure, and fluid velocity during the welding process. Quality deviation is calculated by comparing the actual weld joint formation quality index with the preset target value, reflecting the degree of deviation between the current process and the ideal state. Health residual features capture unmodeled dynamic or potential fault information of the system by comparing predicted values ​​with actual measured values. The construction of the state space must ensure that the data of each dimension are synchronized on the time scale, and normalization is used to eliminate dimensional differences, such as mapping temperature values ​​to the [0,1] interval and converting quality deviation into a percentage form. To improve training efficiency, a sliding window mechanism is used to dynamically update the state space, retaining only the data from the most recent N time steps, while a masking mechanism is introduced to handle missing values.

[0085] The Safety-Constrained Deep Deterministic Policy Gradient Algorithm (SDPRP) is a reinforcement learning algorithm that introduces a safety constraint mechanism on top of the traditional DPRP. This algorithm constructs an actor network and a critic network. The actor network is responsible for adjusting process parameters based on the current state, while the critic network evaluates the value of these actions. During training, the algorithm not only considers minimizing quality deviation but also ensures, through a safety constraint module, that the output process parameter adjustments do not exceed the physical limits of the equipment or process specifications. For example, if the reflow soldering temperature adjustment value output by the actor network exceeds the maximum allowable temperature of the equipment, the safety constraint module automatically corrects it to the maximum safe temperature, thus ensuring the safety of the production process. Simultaneously, the critic network provides optimization directions for the actor network by evaluating the value of actions, continuously updating the actor network's policy parameters through gradient backpropagation until the process parameter adjustments output by the actor network satisfy the health constraints and minimize quality deviation.

[0086] Health constraints encompass various requirements, including equipment physical limits, process specifications, and soldering quality stability. For example, adjusting the pressure parameters of a pick-and-place machine must ensure that it does not cause physical damage to components, while also guaranteeing the consistency of solder joint formation quality. By introducing health constraints, reinforcement learning controllers can optimize process parameters while ensuring the stability and safety of the production process, avoiding quality problems or equipment failures caused by excessive parameter adjustments.

[0087] The setting of health constraints needs to be combined with specific process requirements and equipment characteristics, and dynamically adjusted through experimental verification and expert experience. For example, for high-precision PCBs, stricter lateral offset constraints can be set; for high-temperature sensitive components, the adjustment range of reflow soldering temperature needs to be limited. Furthermore, multi-objective collaborative optimization is supported, balancing conflicts between different constraints through weighted summation or analytic hierarchy process (AHP) to ensure the overall optimality of process parameter adjustment schemes. To further improve the adaptability of the controller, health constraints can integrate historical fault data and process knowledge bases, dynamically predicting potential risks of parameter adjustments through machine learning models, achieving proactive constraint management. For example, when it is detected that a certain type of component is prone to solder joint cracks within a specific temperature range, the constraint boundary of that temperature range will be automatically tightened, and alternative process schemes will be recommended.

[0088] During reinforcement learning training, health constraints are integrated into the loss function design through a penalty function mechanism. Actions that violate the constraints are rewarded with high negative values, guiding the agent to learn strategies that conform to process specifications. Simultaneously, constraint relaxation techniques are employed to gradually loosen the safety boundaries, exploring better combinations of process parameters while ensuring training stability. For example, strict safety constraints are set at the beginning of training, and as training progresses, the constraints are gradually relaxed, allowing the agent to explore a wider parameter space within a safe range, thereby finding the global optimum.

[0089] The performance of reinforcement learning controllers also depends on the proper setting of hyperparameters, such as the learning rate, discount factor, and experience replay buffer size. These hyperparameters need to be fine-tuned experimentally to determine the optimal combination. For example, an excessively high learning rate may cause oscillations during training, while an excessively low learning rate will prolong the convergence time. The discount factor determines the agent's emphasis on future rewards and needs to be set appropriately according to the time scale of the process control.

[0090] To enhance the generalization capability of the reinforcement learning controller, domain randomization techniques are employed to augment the training data. By simulating different equipment states, process parameter fluctuations, and environmental disturbances, the model adapts to the varied conditions of actual production. For example, parameters such as the temperature uniformity of the reflow oven and the repeatability of the pick-and-place machine are randomly adjusted in the simulation environment to generate diverse training samples, enhancing the model's robustness to equipment differences and environmental changes. Simultaneously, a transfer learning mechanism is introduced, using pre-trained model parameters as initial values ​​for fine-tuning under new equipment or process scenarios. This accelerates model convergence and reduces reliance on large amounts of labeled data. For instance, when a new pick-and-place machine is installed on the production line, transfer learning can quickly adapt the control strategy to the new equipment, achieving high control accuracy with only a small amount of measured data.

[0091] Furthermore, to address the dynamic uncertainties in actual production, the reinforcement learning controller also integrates an online adaptive module, which can dynamically adjust hyperparameters or network structure based on real-time feedback data. For example, when a continuous increase in quality deviation is detected, the learning rate is automatically increased to accelerate policy updates; if the deviation in action value function estimation exceeds a threshold, a network layer adjustment mechanism is triggered to optimize the model's expressive power by adding or removing hidden layer nodes.

[0092] In terms of hardware implementation, the reinforcement learning controller adopts an edge computing architecture, deploying the core algorithm on a local industrial computer or embedded device to ensure real-time response speed. For example, in the scenario of reflow soldering temperature control, the controller needs to complete state perception, decision calculation and action execution within milliseconds to avoid temperature fluctuations affecting the solder joint quality.

[0093] Step S60: Convert the process parameter adjustment amount into an executable control command and send it to the production line control system to adjust the reflow soldering temperature curve and the placement machine pressure parameters in real time to achieve adaptive control of the PCBA process.

[0094] It should be noted that the conversion process of process parameter adjustment values ​​into executable control commands must strictly follow the equipment communication protocol and control logic. For example, the reflow soldering temperature profile correction value needs to be decomposed into the setpoint adjustment values ​​of multiple temperature zones and sent to the temperature control module via the Modbus TCP protocol; the placement machine pressure parameter correction value needs to be converted into the torque command of the servo drive and updated synchronously at the millisecond level using the EtherCAT bus.

[0095] To ensure the reliability of control commands, a dual verification mechanism is adopted: during the command generation stage, CRC verification is used to ensure data integrity; during the command issuance stage, the execution status is verified through the device response confirmation mechanism, and if no valid feedback is received, the retransmission mechanism is automatically triggered.

[0096] In addition, the control system also integrates an anomaly handling module. When an abnormal equipment status is detected, such as a temperature sensor malfunction or pressure exceeding the limit, parameter adjustment is immediately suspended and switched to a safe mode, while an alarm message is sent.

[0097] In multi-device collaborative control scenarios, the timing consistency of each subsystem's actions is ensured based on a time synchronization protocol. For example, after the pick-and-place machine completes component pickup, the reflow soldering temperature profile is then stepped up to avoid soldering defects caused by timing misalignment.

[0098] This embodiment provides a PCBA process control method based on multimodal perception and graph neural networks. By introducing virtual reference nodes and their embedded simplified equipment mechanism models, and using health residual feature calculation and dynamic correction mechanism of sensor credibility weights in a dual-channel dynamic graph convolutional network, the propagation of erroneous data to the decision layer is effectively blocked. This allows the production line to maintain the accuracy and stability of state perception when facing equipment aging characteristics drift, sudden changes in operating conditions caused by multi-product switching, and complex electromagnetic / thermal environmental interference, significantly improving the overall robustness of the flexible production line. By introducing a causal-enhanced dynamic coupling graph and using first-channel perturbation propagation inference to quantify the cascading impact of upstream process state fluctuations on downstream workpieces, and combining it with a physical information neural network, the prediction of the solder joint forming process is realized. Through reinforcement learning controller for dynamic process parameter adjustment, adaptive control of the PCBA process is achieved, which can significantly reduce the soldering defect rate and improve the yield.

[0099] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S30 includes steps S301 to S305: Step S301: Construct a dual-channel dynamic graph convolutional network, wherein the dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel adopts three stacked dynamic graph convolutional layers. Each dynamic graph convolutional layer consists of three sub-modules: neighborhood feature aggregation, node feature update, and temporal state propagation. The second channel adopts a multi-head graph attention mechanism for feature aggregation. Each attention head independently learns the attention weights between the physical device node and the virtual reference node.

[0100] It should be noted that the first channel of the dual-channel dynamic graph convolutional network is the disturbance propagation channel, which handles the physical connection edges and environmental coupling edges between physical equipment nodes and workpiece nodes, capturing the propagation law of process disturbances along the physical structure of the production line. This requires multi-hop information aggregation and temporal memory, so stacked dynamic graph convolutional layers are used. In this embodiment, the first channel uses three stacked dynamic graph convolutional layers. Each dynamic graph convolutional layer collects the state information of adjacent nodes, such as physical quantities like temperature and pressure, through a neighborhood feature aggregation module; the node feature update module combines the current node state with the aggregated neighborhood information and updates the node feature representation through nonlinear transformation; the temporal state propagation module captures the changing trend of process state over time by introducing a gated loop unit, ensuring the temporal continuity of disturbance propagation.

[0101] The second channel, the feature enhancement channel, processes the twin mapping edges between each physical device node and its corresponding virtual reference node. By comparing measured values ​​with theoretical baseline values, it assesses sensor reliability and corrects abnormal data. This requires precise one-to-one relationship modeling, hence the use of a multi-head graph attention mechanism. By independently learning the attention weights between physical device nodes and virtual reference nodes under each attention head, it dynamically allocates the degree of influence of different nodes on the current process state. For example, it assigns higher weights to critical device nodes to highlight their impact on process stability.

[0102] The dual-channel design preserves the temporal dynamics of the production line's physical structure while enhancing the expressive power of key features by processing physical disturbance propagation and feature importance weighting in parallel.

[0103] Step S302: Input the dynamic coupling graph into the first and second channels of the dual-channel dynamic graph convolutional network respectively.

[0104] It should be noted that during the input phase, the first channel receives features of physical device nodes and workpiece nodes, as well as physical connection edges and environmental coupling edges. Through three stacked dynamic graph convolutional layers, neighborhood features are aggregated layer by layer, and the hidden state of nodes is updated through gated recurrent units, enabling the node state to propagate along the time axis. The output is the perturbation state embedding vector of each physical device node, realizing multi-hop propagation modeling of process perturbations and capturing the temporal cumulative effect.

[0105] The second channel receives the features of each physical device node and its corresponding virtual reference node, as well as the twin mapping edges between them. A multi-head graph attention mechanism is employed to calculate the attention coefficient between the physical node and the virtual node, which serves as the real-time confidence weight for the sensor corresponding to that physical node. Based on this weight, the original features of the physical node are weighted and corrected, outputting the corrected state vector and confidence weight for each physical device node. This dynamically quantifies sensor confidence, suppresses anomalous data, and improves input quality.

[0106] Step S303: In the first channel, the neighborhood features of the physical device nodes and workpiece nodes in the dynamic coupling graph are aggregated by stacked dynamic graph convolutional layers, and the node hidden state is updated by a gated loop unit to obtain the perturbation state embedding vector of each process node. The perturbation state embedding vector is used to characterize the cumulative process perturbation intensity and propagation path characteristics caused by the change in the state of the upstream physical device to the downstream process node.

[0107] It should be noted that the stacked dynamic graph convolutional layers, by aggregating neighborhood features layer by layer, can effectively capture the propagation patterns of equipment state changes in the production line. For example, in the reflow soldering process, when the temperature in the upstream temperature zone fluctuates, the neighborhood feature aggregation module of the first channel collects temperature change information from adjacent temperature zones. The node feature update module combines the current temperature zone state with the aggregated information and generates a node representation containing heat conduction features through nonlinear transformation. The temporal state propagation module uses gated recurrent units (GRUs) to model the trend of temperature disturbance over time, such as tracking the propagation process of temperature fluctuations from the heating zone to the cooling zone. Through the three-layer stacked structure, the first channel can extract process disturbance features within a multi-hop range and ultimately generate a disturbance state embedding vector for each process node. This vector not only contains disturbance intensity information but also records the propagation path features of the disturbance in the physical structure of the production line, providing the reinforcement learning controller with accurate criteria for judging the source and scope of disturbance.

[0108] In one feasible implementation, step S303 may include: in the first dynamic graph convolutional layer, for each entity device node and workpiece node in the dynamic coupling graph, aggregating the features of the upstream entity device node and downstream workpiece node in the corresponding first-order neighborhood to obtain an aggregated neighborhood feature vector; concatenating the aggregated neighborhood feature vector with the current features of the corresponding node itself, inputting it into the gated recurrent unit, updating the hidden state of the node, and outputting the updated node feature vector of the first layer; using the updated node feature vector of the first layer as the input of the second dynamic graph convolutional layer to aggregate the features of the nodes in the second-order neighborhood. The information is used to obtain an expanded neighborhood feature vector; the expanded neighborhood feature vector is concatenated with the node feature vector updated in the first layer, and then input again into the gated recurrent unit to update the hidden state of the node, outputting the node feature vector updated in the second layer; the node feature vector updated in the second layer is used as the input of the third dynamic graph convolutional layer to aggregate the feature information of nodes in the third-order neighborhood to obtain a deep neighborhood feature vector; the deep neighborhood feature vector is concatenated with the node feature vector updated in the second layer, and then input for the third time into the gated recurrent unit to update the hidden state of the node, obtaining the perturbation state embedding vector of each process node.

[0109] It should be noted that by aggregating neighborhood features layer by layer and updating node states, the three-layer stacked dynamic graph convolutional layer can systematically capture the propagation pattern of equipment state changes in the production line.

[0110] The first dynamic graph convolutional layer aggregates feature information from upstream and downstream nodes within a first-order neighborhood to initially construct local process disturbance features. The second layer extends to second-order neighborhood aggregation, fusing equipment state change information over a larger range. The third layer aggregates third-order neighborhood features to ultimately form deep neighborhood features covering multi-hop propagation paths. The neighborhood feature vectors after each aggregation layer are concatenated with the current node features and, after nonlinear transformation by a gated recurrent unit, generate hidden node states containing temporal memory. This hierarchical aggregation mechanism allows the disturbance state embedding vector to simultaneously characterize the disturbance intensity and its propagation path characteristics within the production line's physical structure. For example, in the reflow soldering process, it can accurately locate the propagation trajectory of temperature fluctuations from the heating zone through the insulation zone to the cooling zone, providing multi-dimensional disturbance analysis basis for the reinforcement learning controller.

[0111] Step S304: In the second channel, a multi-head graph attention mechanism is used to calculate the attention coefficient of the twin mapping edge between each entity device node and the corresponding virtual reference node in the dynamic coupling graph. The attention coefficient is used as the confidence weight of the sensor corresponding to the entity device node. Based on the confidence weight, the original features of the entity device node are weighted and corrected to obtain the corrected state vector.

[0112] It should be noted that the multi-head graph attention mechanism can dynamically evaluate the reliability of sensor data by independently learning the association weights between the physical device node and the virtual reference node under each attention head. For example, in the patch process, if the measured value of a pressure sensor deviates significantly from the theoretical benchmark value, the mechanism will automatically reduce the weight of the node corresponding to that sensor, while increasing the weight allocation of other reliable sensors.

[0113] In practice, each attention head calculates the similarity score between the features of the entity node and the features of the virtual reference node, and generates an attention coefficient matrix after Softmax normalization. This matrix not only reflects the degree of anomaly in the sensor data, but also ensures the robustness of weight allocation through independent computation by multiple heads. Even if a single attention head is disturbed by noise, the effective computation of other heads can still guarantee the accuracy of the overall evaluation.

[0114] In the attention-based weighted correction process, the original features of the physical equipment nodes, such as physical quantities like temperature and pressure, are multiplied by their corresponding weights and summed to generate a corrected state vector. For example, for the pressure parameters of a high-precision pick-and-place machine, if the sensor's measurement value is too high due to electromagnetic interference, the mechanism will reduce the weight of that sensor and increase the weight of other normal sensors in the same process, ultimately outputting a comprehensive feature vector that more closely approximates the actual process state. This dynamic weighting mechanism effectively suppresses the interference of abnormal data on process control while retaining the influence weights of key equipment nodes on process stability, providing a more reliable state input for the subsequent reinforcement learning controller.

[0115] In one feasible implementation, step S304 may include: in the second channel, extracting the twin mapping edges between each physical device node and its corresponding virtual reference node from the dynamic coupling graph, and obtaining the original feature vector of the physical device node and the theoretical baseline feature vector of the virtual reference node; concatenating the original feature vector of the physical device node and the theoretical baseline feature vector of the virtual reference node to generate a joint feature representation of the twin mapping edges; configuring multiple parallel attention heads, wherein each attention head independently learns a set of query transformation matrices and key transformation matrices; performing linear transformations on the joint feature representations through the query transformation matrices and key transformation matrices respectively to obtain the query vector and key vector corresponding to each attention head; determining the scaling dot product attention score between the query vector and the key vector, and obtaining the initial attention coefficients corresponding to each attention head after processing with a normalized exponential function; concatenating and dimensionality-reducing the initial attention coefficients corresponding to each attention head to obtain the fused attention coefficients; using the fused attention coefficients as the confidence weights of the sensors corresponding to the physical device nodes, and performing weighted correction on the original feature vectors of the physical device nodes based on the confidence weights to obtain the corrected state vectors.

[0116] It should be noted that the use of multiple parallel attention heads in the multi-head graph attention mechanism allows for the capture of complex relationships between physical device nodes and virtual reference nodes from different dimensions. Each attention head performs a linear transformation on the joint feature representation of the twin mapping edges using independent query transformation matrices and key transformation matrices, resulting in specific query and key vectors. This allows for the calculation of initial attention coefficients with different emphases. The query vector is an abstract representation of the physical device node features in a specific dimension, used to capture query patterns associated with the virtual reference node; the key vector reflects the response patterns of the virtual reference node features in the corresponding dimension.

[0117] By calculating the scaled dot product attention score between the query vector and the key vector, the association strength between physical device nodes and virtual reference nodes under a specific attention head can be quantified. The formula for calculating the scaled dot product attention score is as follows: in, This represents the scaled dot product attention score of the i-th attention head at time t. This represents the query vector of entity device node i at time t for the k-th attention head. This represents the key vector of the virtual reference node i at time t for the k-th attention head. is the dimension of the key vector.

[0118] After processing with a normalized exponential function, the initial attention coefficients are compressed to the [0,1] interval, preserving relative differences while avoiding numerical overflow. The initial attention coefficients generated by different attention heads are integrated into a high-dimensional vector through a concatenation operation, and then fused through dimensionality reduction, such as a fully connected layer transformation, to synthesize the correlation information captured by different attention heads and obtain the fused attention coefficients.

[0119] The fused attention coefficient matrix not only reflects the reliability distribution of sensor data but also enhances the model's adaptability to complex process environments through a multi-head attention mechanism. For example, in wave soldering, when multiple temperature sensors experience data fluctuations due to solder splashing, different attention heads can focus on dimensions such as heat conduction paths and equipment aging characteristics. The weighted distribution after fusion reduces the impact of abnormal sensors while retaining the accurate representation of the process state by normal sensors. The final corrected state vector has the same dimensions as the original feature vector, but each component has been dynamically adjusted according to the sensor reliability. For example, the weight of a key temperature sensor is reduced from 0.9 to 0.6, while the weights of other sensors in the same area are increased from 0.8 to 0.85, ensuring that the comprehensive feature vector more closely reflects the actual process conditions. This weighted correction mechanism provides the reinforcement learning controller with a more robust state input, effectively improving the robustness of the process control system.

[0120] When using the fused attention coefficients as the confidence weights of the sensors corresponding to the physical device nodes to perform weighted correction on the original feature vectors of the physical device nodes, the original feature value corresponding to each sensor is multiplied by its confidence weight, and then the weighted results of all sensors are summed to obtain the corrected state vector.

[0121] The corrected state vector retains the core process characteristics of key equipment nodes while reducing the weight of anomalous data, allowing the reinforcement learning controller to focus more on changes in the actual process state when making decisions. For example, in the reflow soldering process, when the temperature sensor in the cooling zone experiences a data jump due to environmental interference, the correction mechanism reduces the weight of that sensor from 0.85 to 0.62, while increasing the weight of other normal sensors in the same process from 0.78 to 0.83. The final output comprehensive feature vector then compresses the temperature component fluctuation range from ±5℃ in the original data to ±1.2℃. This dynamic weight allocation mechanism, through independent calculation by multiple attention heads, can identify and isolate anomalous data from different sources, such as pressure sensor drift caused by mechanical vibration and current signal distortion caused by electromagnetic interference—typical industrial noise scenarios.

[0122] Step S305: Concatenate the perturbation state embedding vector and the corrected state vector along the feature dimension to generate a comprehensive state representation vector.

[0123] It should be noted that by concatenating the perturbation state embedding vector with the corrected state vector, the comprehensive state representation vector can simultaneously integrate the state propagation patterns of production line equipment and the reliability information of sensor data. This dual-channel feature fusion mechanism retains the temporal propagation features captured by the dynamic graph convolutional layer and integrates the real-time data reliability evaluated by the multi-head graph attention mechanism, forming a three-dimensional representation of the process state. For example, in the reflow soldering process, when the temperature sensor in the heating zone experiences measurement value drift due to component aging, the dynamic graph convolutional layer records the propagation trajectory of the abnormal temperature along the production line to the insulation zone, generating a perturbation state embedding vector containing propagation path features; simultaneously, the multi-head graph attention mechanism reduces the weight allocation of abnormal sensors by comparing data from other normal sensors in the same process, generating a corrected state vector. The comprehensive state representation vector formed by concatenating the two can both locate the propagation range of temperature anomalies through the perturbation state embedding vector and quantify the actual temperature deviation of each region through the corrected state vector, providing the reinforcement learning controller with a dual decision-making basis including the abnormal propagation path and the actual state deviation.

[0124] In practice, the concatenation operation directly connects the numerical sequences of two vectors along the feature dimension. For example, concatenating a 128-dimensional perturbation state embedding vector with a 64-dimensional corrected state vector generates a 192-dimensional comprehensive state representation vector. This vector retains the original physical quantity values ​​while organically fusing multimodal information through vector dimension expansion. This feature fusion method avoids the information loss caused by simple weighting or threshold judgment in traditional methods. For example, in the wave soldering process, when the flux spraying pressure sensor distorts its data due to pipeline blockage, the comprehensive state representation vector can reflect the influence path of abnormal pressure on the solder wave shape through the perturbation state embedding vector. At the same time, the corrected state vector provides the true pressure value after excluding abnormal data, enabling the reinforcement learning controller to distinguish between equipment failure and process fluctuations, avoiding erroneous control actions caused by errors from a single data source.

[0125] In this embodiment, by introducing a dual-channel graph neural network architecture, the perturbation propagation reasoning and sensor credibility assessment are decoupled, which not only ensures the accuracy of state perception but also realizes data self-correction. Through the splicing and fusion of dual-channel features, the comprehensiveness and reliability of state representation are significantly improved.

[0126] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the PCBA process control method based on multimodal perception and graph neural networks in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0127] This application also provides a collaborative design device for building orientation that integrates multi-objective optimization with urban microclimate coupling. Please refer to [link / reference]. Figure 2The building orientation collaborative design device, which integrates multi-objective optimization with urban microclimate coupling, includes: The acquisition module 10 is used to acquire multimodal feature vectors during the operation of the surface mount technology production line through a multimodal sensor array, and to perform virtual-real synchronous calibration and mechanism model deviation analysis on the digital twin model of the production line based on the multimodal feature vectors to obtain health residual characteristics.

[0128] The construction module 20 is used to construct a dynamic coupling graph based on the multimodal feature vector and the health residual feature. The node set of the dynamic coupling graph includes physical device nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges.

[0129] The inference module 30 is used to input the dynamic coupling graph into a dual-channel dynamic graph convolutional network to perform joint inference of state perturbation propagation and sensor credibility to obtain a comprehensive state representation vector. The dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel is used to aggregate the neighborhood features of physical equipment nodes and workpiece nodes and infer the propagation path of state perturbation caused by changes in the state of upstream equipment to downstream process nodes. The second channel is used to aggregate the twin mapping edge features between each physical equipment node and the corresponding virtual reference node and to perform weighted correction on the original features of the physical equipment nodes by calculating the credibility weight of each sensor.

[0130] The prediction module 40 is used to input the comprehensive state representation vector into the physical information neural network to predict the solder joint forming quality and obtain the solder joint forming quality index of the current batch of printed circuit boards.

[0131] The decision module 50 is used to activate the reinforcement learning controller when the quality deviation between the weld joint forming quality index and the preset target quality index reaches a preset trigger threshold, and to make dynamic decisions on the reinforcement learning process parameters through the reinforcement learning controller to obtain the process parameter adjustment amount.

[0132] The adjustment module 60 is used to convert the process parameter adjustment amount into executable control commands and send them to the production line control system to adjust the reflow soldering temperature curve and the placement machine pressure parameters in real time, so as to realize the adaptive control of the PCBA process.

[0133] The building orientation collaborative design device integrating multi-objective optimization and urban microclimate coupling provided in this application adopts the PCBA process control method based on multimodal perception and graph neural networks in the above embodiments, which can solve the technical problems of poor welding consistency and high rework rate caused by environmental disturbances and equipment aging in variable process environments. Compared with the prior art, the beneficial effects of the building orientation collaborative design device integrating multi-objective optimization and urban microclimate coupling provided in this application are the same as the beneficial effects of the PCBA process control method based on multimodal perception and graph neural networks provided in the above embodiments, and other technical features in the building orientation collaborative design device integrating multi-objective optimization and urban microclimate coupling are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0134] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the PCBA process control method based on multimodal perception and graph neural networks as described above.

[0135] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A PCBA process control method based on multimodal sensing and graph neural networks, characterized in that, The method includes: Multimodal feature vectors are collected during the operation of the surface mount technology production line by a multimodal sensor array, and the digital twin model of the production line is calibrated in a virtual-real synchronization manner and the mechanism model deviation analysis is performed based on the multimodal feature vectors to obtain health residual characteristics. Based on the multimodal feature vector and the health residual feature, a dynamic coupling graph is constructed, wherein the node set of the dynamic coupling graph includes physical device nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges. The dynamic coupling graph is input into a dual-channel dynamic graph convolutional network for joint inference of state perturbation propagation and sensor credibility to obtain a comprehensive state representation vector. The dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel is used to aggregate the neighborhood features of physical equipment nodes and workpiece nodes and infer the propagation path of state perturbation caused by changes in the state of upstream equipment to downstream process nodes. The second channel is used to aggregate the twin mapping edge features between each physical equipment node and the corresponding virtual reference node and to perform weighted correction on the original features of the physical equipment nodes by calculating the credibility weight of each sensor. The comprehensive state representation vector is input into the physical information neural network to predict the solder joint forming quality, thereby obtaining the solder joint forming quality index of the current batch of printed circuit boards. When the quality deviation between the weld joint forming quality index and the preset target quality index reaches the preset trigger threshold, the reinforcement learning controller is activated, and the reinforcement learning controller performs dynamic decision-making on the reinforcement learning process parameters to obtain the process parameter adjustment amount. The process parameter adjustment amounts are converted into executable control commands and sent to the production line control system to adjust the reflow soldering temperature profile and the chip mounter pressure parameters in real time, thereby achieving adaptive control of the PCBA process.

2. The method as described in claim 1, characterized in that, The process involves acquiring multimodal feature vectors during the operation of a surface mount technology production line using a multimodal sensor array, and then performing virtual-real synchronous calibration and mechanism model deviation analysis on the digital twin model of the production line based on these multimodal feature vectors to obtain health residual characteristics, including: A multimodal sensor array is deployed at key process nodes in the surface mount technology production line to collect multi-source heterogeneous data in real time during the production line operation. The multi-source heterogeneous data includes environmental temperature and humidity data, equipment vibration data, infrared thermal imaging data of the welding area, and equipment current data. The control command settings of each actuator are read in real time from the production line control system. The control command settings are time-paired with the multi-source heterogeneous data and feature extraction is performed to obtain a multimodal feature vector. A digital twin model of the production line is constructed. In the digital twin model, a corresponding virtual reference node is configured for each physical equipment node. The virtual reference node is embedded with a simplified mechanism model of the corresponding physical equipment. The simplified mechanism model includes a mounting force calculation model based on the pressure conversion relationship, a heat power calculation model based on Joule's law, and a wind speed calculation model based on the air volume and rotation speed relationship. The control command setpoint is input into the simplified mechanism model to determine the theoretical output reference value of each actuator at the current moment; Determine the residual sequence between the actual feedback value of the sensor and the theoretical output reference value at the same time, and perform exponential weighted moving average smoothing on the residual sequence to obtain the initial residual characteristics; A mapping function from physical space to virtual space is established, and the twin parameters of the digital twin model are calibrated online by minimizing the physical-virtual feature divergence to obtain the calibrated twin parameters; The theoretical output baseline value is corrected based on the calibrated twin parameters to obtain the corrected theoretical output baseline value, and the health residual characteristics are determined based on the corrected theoretical output baseline value.

3. The method as described in claim 1, characterized in that, The construction of a dynamic coupling graph based on the multimodal feature vectors and the health residual features includes: Define physical device nodes, workpiece nodes, and virtual reference nodes configured for each physical device node, wherein the physical device nodes include pick-and-place machine nodes and independent temperature zone nodes of the reflow oven, and the workpiece nodes represent the current batch of printed circuit boards. The feature data at the corresponding position in the multimodal feature vector is used as the initial feature of the physical device node. The static information of the current batch of printed circuit boards is transformed by the embedding layer and used as the initial feature of the workpiece node. The health residual feature is used as the initial feature of the virtual reference node to generate an initial graph structure. The static information includes identification code, board size parameters and component distribution density data. Based on the material flow sequence of the surface mount technology production line, directed edges are constructed from upstream equipment nodes to workpiece nodes and from workpiece nodes to downstream equipment nodes as physical connection edges. These physical connection edges are used to describe the flow path of materials between printing, chip mounting, reflow soldering and automatic optical inspection processes. Obtain the physical location coordinates of each physical equipment node in the production line, calculate the Euclidean distance between any two physical equipment nodes, and when the Euclidean distance is less than a preset coupling distance threshold, construct an undirected edge between the two physical equipment nodes as an environmental coupling edge. The environmental coupling edge is used to describe the thermal radiation interference and vibration transmission coupling effect between the equipment. For each physical device node, a directed edge is constructed between it and the corresponding virtual reference node as a twin mapping edge, and the direction of the twin mapping edge is from the physical device node to the virtual reference node; A dynamic coupled graph is constructed based on the initial graph structure, the physical connection edges, the environmental coupling edges, and the twin mapping edges.

4. The method as described in claim 3, characterized in that, The construction of a dynamic coupling graph based on the initial graph structure, the physical connection edges, the environmental coupling edges, and the twin mapping edges includes: Collect virtual-real fusion state vector sequences from historical production cycles to construct a causal discovery dataset; A causal discovery algorithm based on conditional independence test is executed on the causal discovery dataset. The maximum lag order is set, the conditional independence test statistic for each entity device node pair is calculated, non-causal edges are deleted through directed separation criterion, and an undirected causal skeleton graph is constructed. A fraction-optimized causal structure learning algorithm is executed on the undirected causal skeleton graph to optimize the data fitting loss under the acyclicity constraint, outputting a directional causal graph, and using the directional causal graph as a causal structure prior. The thermal influence coefficient and vibration influence coefficient are determined based on environmental temperature and humidity data and equipment current data, and the weighted sum of the thermal influence coefficient and the vibration influence coefficient is used as the initial dynamic weight of the environmental coupling edge. The causal structure prior is converted into a causal adjacency matrix, and the causal adjacency matrix is ​​fused with the initial dynamic weights of the environmental coupling edge using Hadamard to obtain the causal-enhanced environmental coupling edge. The initial graph structure is integrated with the physical connection edges, the causal-enhanced environmental coupling edges, and the twin mapping edges to obtain a dynamic coupling graph.

5. The method as described in claim 1, characterized in that, The process of inputting the dynamic coupling graph into a dual-channel dynamic graph convolutional network for state perturbation propagation and joint inference of sensor credibility yields a comprehensive state representation vector, including: A dual-channel dynamic graph convolutional network is constructed, wherein the dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel adopts three stacked dynamic graph convolutional layers. Each dynamic graph convolutional layer consists of three sub-modules: neighborhood feature aggregation, node feature update, and temporal state propagation. The second channel adopts a multi-head graph attention mechanism for feature aggregation. Each attention head independently learns the attention weights between the physical device node and the virtual reference node. The dynamic coupling graph is input into the first and second channels of the dual-channel dynamic graph convolutional network, respectively. In the first channel, the neighborhood features of the physical device nodes and workpiece nodes in the dynamic coupling graph are aggregated by stacked dynamic graph convolutional layers, and the node hidden state is updated by gated loop unit to obtain the perturbation state embedding vector of each process node. The perturbation state embedding vector is used to characterize the cumulative process perturbation intensity and propagation path characteristics caused by the change of the state of the upstream physical device to the downstream process node. In the second channel, a multi-head graph attention mechanism is used to calculate the attention coefficient of the twin mapping edge between each entity device node and the corresponding virtual reference node in the dynamic coupling graph. The attention coefficient is used as the confidence weight of the sensor corresponding to the entity device node, and the original features of the entity device node are weighted and corrected based on the confidence weight to obtain the corrected state vector. The perturbation state embedding vector and the corrected state vector are concatenated along the feature dimension to generate a comprehensive state representation vector.

6. The method as described in claim 5, characterized in that, In the first channel, neighborhood features of entity device nodes and workpiece nodes in the dynamically coupled graph are aggregated through stacked dynamic graph convolutional layers, and the hidden state of nodes is updated through a gated recurrent unit to obtain the perturbation state embedding vector of each process node, including: In the first dynamic graph convolutional layer, for each entity device node and workpiece node in the dynamic coupling graph, the features of the upstream entity device node and downstream workpiece node in the corresponding first-order neighborhood are aggregated to obtain the aggregated neighborhood feature vector. The aggregated neighborhood feature vector is concatenated with the current feature of the corresponding node, input into the gated loop unit, the hidden state of the node is updated, and the updated node feature vector of the first layer is output. The updated node feature vector of the first layer is used as the input of the second dynamic graph convolutional layer to aggregate the feature information of nodes in the second-order neighborhood and obtain the expanded neighborhood feature vector. The expanded neighborhood feature vector is concatenated with the updated node feature vector of the first layer, and then input into the gated loop unit to update the hidden state of the node, and the updated node feature vector of the second layer is output. The updated node feature vector of the second layer is used as the input of the third dynamic graph convolutional layer to aggregate the feature information of nodes in the third-order neighborhood and obtain the deep neighborhood feature vector. The deep neighborhood feature vector is concatenated with the updated node feature vector of the second layer, and then input into the gated loop unit for the third time to update the hidden state of the node, so as to obtain the perturbation state embedding vector of each process node.

7. The method as described in claim 5, characterized in that, In the second channel, a multi-head graph attention mechanism is used to calculate the attention coefficient of the twin mapping edge between each entity device node and its corresponding virtual reference node in the dynamic coupling graph. The attention coefficient is used as the confidence weight of the sensor corresponding to the entity device node. Based on the confidence weight, the original features of the entity device node are weighted and corrected to obtain the corrected state vector, including: In the second channel, the twin mapping edge between each physical device node and the corresponding virtual reference node is extracted from the dynamic coupling graph, and the original feature vector of the physical device node and the theoretical baseline feature vector of the virtual reference node are obtained. The original feature vector of the physical device node is concatenated with the theoretical baseline feature vector of the virtual reference node to generate a joint feature representation of the twin mapping edge; Configure multiple parallel attention heads, where each attention head independently learns a set of query transformation matrices and key transformation matrices; The joint feature representation is linearly transformed by the query transformation matrix and the key transformation matrix respectively to obtain the query vector and key vector corresponding to each attention head; The scaled dot product attention score between the query vector and the key vector is determined, and after processing with a normalized exponential function, the initial attention coefficients corresponding to each attention head are obtained. The initial attention coefficients corresponding to each attention head are concatenated and dimensionality reduced to obtain the fused attention coefficients. The fused attention coefficients are used as the confidence weights of the sensors corresponding to the physical device nodes, and the original feature vectors of the physical device nodes are weighted and corrected based on the confidence weights to obtain the corrected state vectors.

8. The method as described in claim 1, characterized in that, The step of inputting the comprehensive state representation vector into a physical information neural network to predict solder joint formation quality and obtain the solder joint formation quality index of the current batch of printed circuit boards includes: A physical information neural network is constructed, wherein the physical information neural network includes an input layer, multiple physical constraint hidden layers and a multi-task output layer connected in sequence. The physical constraint hidden layers embed partial differential equations of the physical mechanism of solder joint formation, and the physical mechanism partial differential equations include heat conduction equations, fluid dynamics equations and phase transition dynamics equations. The input layer maps the comprehensive state representation vector to an initial feature tensor of a preset dimension, and the initial feature tensor is passed to the first physical constraint hidden layer. In the physical constraint hidden layer, the temperature field distribution characteristics of the solder joint during the reflow soldering process are calculated based on the heat conduction equation, the fluid velocity field and pressure field characteristics of the solder paste in the molten state are calculated based on the fluid dynamics equation, and the phase transition ratio characteristics of the solder from solid to liquid and back to solid are calculated based on the phase transition dynamics equation. The temperature field distribution characteristics, fluid velocity field and pressure field characteristics, and phase transition ratio characteristics are used as physical constraint characteristics and fused with the initial feature tensor to obtain the fused physical enhancement feature tensor. The fused physical enhancement feature tensor is input to the multi-task output layer, wherein the multi-task output layer contains multiple parallel fully connected sub-networks, each of which corresponds to a solder joint forming quality sub-index, wherein the solder joint forming quality sub-index includes solder joint void rate, solder joint shear strength, solder joint wetting angle and solder joint lateral offset. Regression prediction is performed on the physical enhancement feature tensor through each fully connected sub-network to obtain the predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle and solder joint lateral offset for each solder joint on the current batch of printed circuit boards. The predicted values ​​of solder joint void rate, solder joint shear strength, solder joint wetting angle, and solder joint lateral offset are spliced ​​and weighted to generate the solder joint forming quality index of the current batch of printed circuit boards.

9. The method as described in claim 1, characterized in that, When the quality deviation between the weld joint forming quality index and the preset target quality index reaches a preset trigger threshold, the reinforcement learning controller is activated. The reinforcement learning controller then performs dynamic decision-making on the reinforcement learning process parameters to obtain the process parameter adjustment amount, including: The quality deviation is determined based on the weld joint forming quality index and the preset target quality index; The quality deviation is compared with a preset trigger threshold, and the reinforcement learning controller is activated when the quality deviation is greater than the preset trigger threshold. Construct a reinforcement learning state space, which includes a comprehensive state representation vector, quality bias, and health residual features; The actor network and critic network are constructed using a safety-constrained deep deterministic strategy gradient algorithm. The current state is sampled from the reinforcement learning state space and input into the actor network to obtain the process parameter adjustment action; The process parameter adjustment actions are input into the critic network for value evaluation, and the action value function value is obtained. The policy parameters of the actor network are updated based on the action value function value until the process parameter adjustment action output by the actor network satisfies the health constraints and minimizes the quality deviation, thus obtaining the optimized process parameter adjustment action. The process parameter adjustment amount is determined based on the optimized process parameter adjustment action, wherein the process parameter adjustment amount includes the reflow soldering temperature profile correction value and the pick-and-place machine pressure parameter correction value.

10. A PCBA process control device based on multimodal perception and graph neural networks, characterized in that, The device includes: The acquisition module is used to acquire multimodal feature vectors during the operation of the surface mount technology production line through a multimodal sensor array, and to perform virtual-real synchronous calibration and mechanism model deviation analysis on the digital twin model of the production line based on the multimodal feature vectors to obtain health residual characteristics. The construction module is used to construct a dynamic coupling graph based on the multimodal feature vector and the health residual feature, wherein the node set of the dynamic coupling graph includes physical device nodes, workpiece nodes and virtual reference nodes, and the edge set includes physical connection edges, environmental coupling edges and twin mapping edges. The inference module is used to input the dynamic coupling graph into a dual-channel dynamic graph convolutional network to perform joint inference of state perturbation propagation and sensor credibility, and obtain a comprehensive state representation vector. The dual-channel dynamic graph convolutional network includes a first channel and a second channel. The first channel is used to aggregate the neighborhood features of physical equipment nodes and workpiece nodes and infer the propagation path of state perturbation caused by changes in the state of upstream equipment to downstream process nodes. The second channel is used to aggregate the twin mapping edge features between each physical equipment node and the corresponding virtual reference node and to perform weighted correction on the original features of the physical equipment nodes by calculating the credibility weight of each sensor. The prediction module is used to input the comprehensive state representation vector into the physical information neural network to predict the solder joint forming quality and obtain the solder joint forming quality index of the current batch of printed circuit boards. The decision module is used to activate the reinforcement learning controller when the quality deviation between the weld joint forming quality index and the preset target quality index reaches a preset trigger threshold, and to make dynamic decisions on the reinforcement learning process parameters through the reinforcement learning controller to obtain the process parameter adjustment amount. The adjustment module is used to convert the process parameter adjustment amount into executable control commands and send them to the production line control system to adjust the reflow soldering temperature profile and the placement machine pressure parameters in real time, so as to achieve adaptive control of the PCBA process.

Images