SMT production line process parameter adaptive optimization method based on digital twinning

Abstract

This invention discloses an adaptive optimization method for SMT production line process parameters based on digital twins, belonging to the field of surface mount technology (SMT) data control technology. It includes constructing a digital twin to map SMT production line elements, reconstructing a model to calculate the SMT production line process status, deriving an SMT production line capacity optimization scheme, dynamically correcting process parameters, eliminating model overfitting faults caused by batch variations in incoming materials, and optimizing the joint simulation of equipment status and process parameters. This invention upgrades SMT production line process optimization from traditional reactive optimization to predictive optimization, reducing quality defects at the source and achieving a fundamental upgrade in the optimization model. Through a four-dimensional hierarchical digital twin and reduced-order model virtual sensing technology, it achieves real-time acquisition of key physical quantities that traditional methods cannot measure, such as thermal field distribution, stencil deformation, and dynamic gaps. This eliminates decision-making blind spots in process optimization, improves the accuracy and timeliness of optimization decisions, and thus fills the data dimension gap in process optimization.

Description

Technical Field

[0001] This invention relates to the field of surface mount technology (SMT) data control technology, and in particular to an adaptive optimization method for SMT production line process parameters based on digital twins. Background Technology

[0002] Surface Mount Technology (SMT) is a core process in modern electronics manufacturing, encompassing multiple complex processes such as solder paste printing, component placement, reflow soldering, and wave soldering. It involves dozens of strongly coupled process parameters, and the parameter configuration directly determines the soldering yield, production efficiency, and manufacturing cost of electronic products.

[0003] Traditional SMT production line process parameter optimization methods mainly rely on the manual experience of senior process engineers, offline pilot production testing, and statistical process control (SPC). However, these methods have the following insurmountable technical drawbacks in actual production applications: 1. Lack of data dimensions leads to blind spots in process optimization. Traditional methods can only collect equipment operating parameters and post-processing inspection results from SPI and AOI. The equipment status and process results are severely disconnected, and key physical quantities that determine the process effect, such as thermal field distribution, stencil deformation, and dynamic gap between the stencil and PCB, cannot be obtained. There are fundamental data blind spots in the optimization logic.

[0004] 2. Multiple objectives cannot be coordinated, and it is easy to get stuck in local optima. There are strong coupling characteristics between SMT process parameters. For example, parameter adjustment to solve the "tombstone" defect may cause the "solder ball" defect. Traditional methods cannot simultaneously take into account the coordinated optimization of multiple objectives such as yield, production efficiency, energy consumption, equipment and consumable life. They can only achieve local optima of single indicators and cannot meet the diversified needs of flexible manufacturing.

[0005] 3. The response is delayed and cannot adapt to batch drift. Traditional methods rely on post-event feedback from detection equipment such as SPI and AOI. By this time, quality defects have already occurred, and it is impossible to correct the process in real time. At the same time, changes in the characteristics of incoming materials and environmental temperature and humidity between batches will cause process parameters to drift. Traditional methods cannot achieve real-time adaptation and are prone to batch quality defects. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the prior art by proposing an adaptive optimization method for SMT production line process parameters based on digital twins.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The adaptive optimization method for SMT production line process parameters based on digital twins includes the following steps: S1: Construct a digital twin to map SMT production line elements, and construct a four-dimensional hierarchical digital twin of the SMT production line. The four-dimensional hierarchical digital twin includes a geometric twin layer, a physical twin layer, a behavioral twin layer and a rule twin layer coupled in sequence, so as to realize the real-time mapping of all elements of the physical entity of the SMT production line and the digital twin. S2: Reconstruction model to calculate the process status of SMT production line. Based on the reduced-order model, virtual sensing mechanism and key physical field reconstruction model are constructed. The unmeasurable key physical parameters are calculated in real time through measurable process and equipment data, filling the blind spot of physical measurement and obtaining full-dimensional process status data of SMT production line. S3: Derive the SMT production line capacity optimization scheme, conduct multi-objective virtual experiments based on the digital twin, take the production quality, efficiency, energy consumption and equipment life of the SMT production line as the collaborative optimization objectives, perform Pareto optimization through multi-objective evolutionary algorithm, output the non-dominated optimal solution set of process parameters, and after virtual verification, send the feasible scheme to the physical production line for execution. S4: Dynamically correct process parameters. Based on the real-time virtual-real synchronization of the physical production line and the digital twin, the process is simulated and predicted in advance through a reduced-order physical model. Potential quality defects are predicted and feedforward compensation instructions are generated and sent to the physical production line equipment to realize the dynamic correction of the process in real time. S5: Eliminate model overfitting faults caused by changes in incoming material batches, construct an inheritable product process twin asset library, encapsulate the process data and models of mass-produced products into reusable twin sub-models, and solve the cold start problem of new product introduction and model overfitting problem caused by changes in incoming material batches through similarity retrieval and domain adaptive migration algorithm. S6: Optimize the joint simulation of equipment status and process parameters. Embed the equipment consumable degradation model in the digital twin to realize the joint simulation optimization of equipment health status and process parameters, balance production continuity and equipment maintenance costs, and output the optimal maintenance window suggestion.

[0008] As a preferred embodiment, in step S1, the geometric twin layer is a high-precision 3D model of all equipment and tooling in the SMT production line, covering the three-dimensional geometric models of the placement motion mechanism, reflow oven structure, scraper mechanism, PCB support pin array, stencil, and carrier. The physical twin layer collects the operating physical parameters of the SMT production line equipment in real time through an IoT sensor network, including the torque and speed of the scraper servo motor, the temperature and hot air parameters of the reflow oven temperature zone, stencil tension data, equipment vibration data, and PCB positioning data, and maps the collected data to the digital twin in real time to achieve synchronization between the physical entity and the digital model. The behavioral twin layer constructs a dynamic mapping model of process parameters and physical response, which is used to output the corresponding physical response results of stencil deformation, solder paste filling volume, device temperature curve, and placement accuracy change in real time after inputting the process parameter adjustment amount. The rule twin layer transforms the IPC electronic component acceptability standard, enterprise SMT process specifications, and process expert experience into algorithm-recognizable constraints and embeds them into the decision logic of the digital twin.

[0009] As a preferred embodiment, in step S2, the key physical field reconstruction model includes a reflow oven dynamic thermal field reconstruction model, and the construction steps of the reflow oven dynamic thermal field reconstruction model are as follows: S21: Obtain the basic thermal field distribution model of the reflow oven chamber through CFD fluid thermodynamics simulation; S22: By integrating real-time measurement data from multi-node thermocouple arrays and infrared thermal imaging sensors within the furnace, a reduced-order thermodynamic model is constructed to achieve real-time reconstruction of dynamic thermal field cloud maps across the entire furnace area. This enables visualization and monitoring of the real-time temperature curves of each device on the PCB board and automatic identification of cold and hot spot regions in the thermal field.

[0010] As a preferred embodiment, in step S2, the virtual sensing mechanism calculates the values ​​of unmeasurable parameters in real time based on a reduced-order model and combined with real-time acquired measurable process and equipment data for parameters that cannot be directly measured by physical sensors, such as the dynamic gap between the stencil and the PCB, the real-time temperature inside the BGA solder joint, and the deformation displacement of the carrier.

[0011] As a preferred embodiment, in step S3, the optimization objectives of the multi-objective virtual experiment include solder paste printing yield, mounting accuracy, reflow soldering defect rate, production efficiency, equipment energy consumption, and consumable wear. The optimization variables include squeegee pressure, speed, and demolding parameters in the solder paste printing stage; mounting coordinates, angles, and pressure parameters in the mounting stage; and temperature zone temperature and chain speed parameters in the reflow soldering stage. The constraints are the process specifications and equipment capability boundaries in the regular twin layer. The multi-objective evolutionary algorithm adopts the non-dominated sorting genetic algorithm NSGA-II with an elite strategy, and conducts multiple sets of virtual printing, virtual mounting, and virtual reflow soldering simulation experiments in parallel. It outputs a Pareto non-dominated optimal solution set that meets the constraints. Each solution set corresponds to a set of process parameter schemes and corresponding optimization objective values. The selected process parameter schemes are first fully virtual verified in the digital twin. After confirming that there are no quality or equipment risks, they are then sent to the physical production line equipment for execution.

[0012] As a preferred embodiment, in step S4, the data from the sensors and detection equipment of the physical production line are synchronously mapped to the digital twin to achieve millisecond-level real-time virtual-physical synchronization. The advanced simulation prediction is used to predict potential quality defects such as mounting offset, excessive solder paste thickness, and temperature curve deviation from the process range. The feedforward compensation command is generated and sent to the corresponding equipment before the defect occurs to complete the real-time correction of process parameters and realize the prevention and correction of process defects in advance.

[0013] As a preferred embodiment, in step S5, the twin sub-models stored in the twin asset library include PCB design data, optimal process parameter combinations, defect compensation models, stencil design and wear compensation data, equipment adaptation parameters, and quality traceability data for the corresponding products. In the new product introduction stage, the system extracts key features of the product to be manufactured, such as PCB thickness, device package type, pin spacing, device height, and panelization method. Through a physical similarity retrieval algorithm, it matches the historical product with the highest similarity in the twin asset library, migrates the corresponding twin sub-model, and generates the initial process plan for the new product. In response to changes in incoming material batches, the system uses a domain adaptive algorithm to quickly update the model boundary conditions of the digital twin using the measured inspection data of the first PCB, adjusts the process parameter compensation model, and avoids the model overfitting to historical batch data.

[0014] As a preferred embodiment, in step S6, the equipment consumable degradation model includes a steel mesh tension decay model, a scraper wear model, a carrier deformation degradation model, a nozzle wear model, and a reflow soldering heating tube aging model, which are used to predict the remaining service life (RUL) of key consumables and components in real time. The co-simulation optimization takes the health status of consumables as a constraint condition for process parameter optimization. During the parameter optimization process, process quality and consumable service life are considered simultaneously. Before the consumable status reaches the critical threshold, the production benefits and costs after consumable replacement are simulated through virtual simulation to balance production continuity and maintenance costs, and the optimal maintenance window suggestion is output.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention upgrades SMT production line process optimization from traditional reactive optimization to predictive optimization, completing the leap from passive correction to proactive design, reducing the generation of quality defects from the source, and achieving a fundamental upgrade of the optimization mode. Through four-dimensional hierarchical digital twin and reduced-order model virtual sensing technology, it realizes real-time acquisition of key physical quantities that cannot be measured by traditional methods, such as thermal field distribution, stencil deformation, and dynamic gaps. This eliminates the decision blind spot in process optimization, improves the accuracy and timeliness of optimization decisions, and thus makes up for the data dimension shortcomings of process optimization.

[0016] 2. This invention transfers the process parameter trial and error process from the physical production line to the digital twin environment, completely avoiding material waste and downtime losses caused by physical trial and error. At the same time, it outputs multiple optimal solutions through Pareto optimization, breaking through the limitations of traditional methods in optimizing single-point indicators, and achieving multi-objective collaborative optimization of yield, efficiency, energy consumption, and equipment life, thereby meeting diverse production needs.

[0017] 3. This invention shortens the new product introduction cycle and significantly improves the initial yield of new products through a reusable twin asset library and similarity transfer algorithm. Through domain adaptive algorithm, it can quickly adapt to changes in incoming material batches, avoid model overfitting to historical data, and improve the adaptive ability of the process system to operating condition fluctuations. It solves the industry pain points of cold start and overfitting. Furthermore, it incorporates the consumable degradation model into the process control closed loop, taking into account the service life of consumables and the health status of equipment while optimizing process parameters. It accurately predicts maintenance windows, balances production continuity and maintenance costs, and avoids process out-of-control and batch quality defects caused by consumable failure, thereby achieving joint optimization of equipment health and process.

[0018] 4. This invention addresses the challenges of traditional processes such as mixed-assembly board reflow soldering, wave soldering, FPC flexible mounting, and nitrogen reflow soldering. Through a regionalized, point-to-point, and feedforward adaptive control strategy, it effectively addresses extreme operating conditions such as heat capacity differences, shadowing effects, expansion and contraction nonlinearity, and strong coupling of multiple parameters, thus broadening the process adaptability range and improving the adaptability to extreme and complex operating conditions. Attached Figure Description

[0019] Figure 1 This is a flowchart of the adaptive optimization method for SMT production line process parameters based on digital twins proposed in this invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0022] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] Reference Figure 1 The adaptive optimization method for SMT production line process parameters based on digital twins includes the following steps: S1: Construct a digital twin to map SMT production line elements. Construct a four-dimensional hierarchical digital twin of the SMT production line. The four-dimensional hierarchical digital twin includes a geometric twin layer, a physical twin layer, a behavioral twin layer, and a rule twin layer that are coupled in sequence, so as to realize the real-time mapping of all elements of the physical entity of the SMT production line and the digital twin. Furthermore, in step S1, the geometric twin layer is a high-precision 3D model of all equipment and tooling in the SMT production line, covering the three-dimensional geometric models of the chip placement mechanism, reflow oven structure, scraper mechanism, PCB support pin array, stencil, and carrier. The construction process of the geometric twin layer is as follows: Using 3D modeling software, a high-precision 1:1 3D model of all equipment and tooling in the SMT production line is constructed, covering the geometric features of all elements such as the X / Y / Z motion axes of the pick-and-place machine, the nozzle mechanism, the squeegee mechanism of the printer, the multi-temperature zone structure of the reflow oven, the PCB support pin array, the stencil, and the carrier. The modeling accuracy is no less than ±0.01mm, completely restoring the spatial layout and mechanical structure of the production line. The physical twin layer is constructed by collecting the operating physical parameters of SMT production line equipment in real time through IoT sensor networks, including the torque and speed of the scraper servo motor, the temperature and hot air parameters of the reflow oven temperature zone, the stencil tension data, the equipment vibration data, the PCB positioning data, and the real-time detection data of SPI and AOI. The collected data is then mapped to the digital twin in real time to achieve state synchronization between the physical entity and the digital model. All collected data is synchronously mapped to the digital twin via industrial Ethernet with a delay of no more than 10ms. The behavioral twin layer constructs a dynamic mapping model between process parameters and physical responses. This model, after inputting process parameter adjustments, outputs in real-time the corresponding physical response results, including stencil deformation, solder paste filling volume, device temperature profiles, and changes in mounting accuracy. The construction process is as follows: Based on physical mechanisms and experimental data, a dynamic mapping model of process parameters and physical responses is constructed, including: For the solder paste printing stage, a dynamic response model is established for squeegee pressure, squeegee speed, demolding distance, stencil deformation, solder paste filling volume, and solder paste printing thickness; for the component placement stage, a mapping model is established for placement coordinates, placement pressure, nozzle vacuum, placement offset, and component damage probability; for the reflow soldering stage, a dynamic response model is established for the set temperature of the temperature zone, conveyor speed, and the real-time temperature curves of each component on the PCB. These models can achieve the following: after inputting process parameter adjustments, the corresponding physical response results are output in real time, providing a model foundation for subsequent virtual experiments and predictions. The rule twin layer transforms IPC electronic component acceptability standards, enterprise SMT process specifications, and process expert experience into constraints that can be recognized by the algorithm, embedding them into the decision logic of the digital twin. Its construction process is as follows: The IPC-A-610 electronic component acceptability standard, the company's internal SMT process specifications, and the expert experience of senior process engineers are transformed into algorithm-recognizable constraints and decision rules, which are then embedded into the decision logic of the digital twin. These include the upper and lower limits of solder paste thickness, the allowable threshold for placement offset, the range requirements of reflow soldering temperature profiles, and the safety boundaries of equipment operation, ensuring that all process optimization solutions meet compliance requirements.

[0025] S2: Reconstruction model to calculate the process status of SMT production line. Based on the reduced-order model, virtual sensing mechanism and key physical field reconstruction model are constructed. The unmeasurable key physical parameters are calculated in real time through measurable process and equipment data, filling the blind spot of physical measurement and obtaining full-dimensional process status data of SMT production line. Furthermore, in step S2, the virtual sensing mechanism, based on a reduced-order model and combined with real-time acquired measurable process and equipment data, calculates the values ​​of unmeasurable parameters in real time for parameters that cannot be directly measured by physical sensors, such as the dynamic gap between the stencil and the PCB, the real-time temperature inside the BGA solder joints, and the deformation and displacement of the carrier. The sampling frequency is no less than 100Hz. Its construction process is as follows: For key parameters that cannot be directly measured by physical sensors, such as the dynamic gap between the stencil and the PCB, the real-time temperature inside the solder joints of BGA devices, and the deformation displacement of the carrier, a corresponding reduced-order physical model is constructed. Combined with the measurable data collected in real time by the physical twin layer in step 1, the values ​​of the unmeasurable parameters are calculated in real time. For example, for the dynamic gap between the stencil and the PCB, a reduced-order model of the elastic deformation of the stencil is established. Inputting real-time collected data on squeegee pressure, squeegee position, stencil tension, and support pin layout and height, the deformation displacement of the stencil at any position during the printing process is calculated in real time, thus obtaining the dynamic gap value between the stencil and the PCB. The sampling frequency reaches 100Hz, completely filling the measurement blind spot of traditional methods. The key physical field reconstruction model includes the reflow oven dynamic thermal field reconstruction model. The construction steps of the reflow oven dynamic thermal field reconstruction model are as follows: S21: Obtain the basic thermal field distribution model of the reflow oven chamber through CFD fluid thermodynamics simulation; S22: By integrating real-time measurement data from multi-node thermocouple arrays and infrared thermal imaging sensors within the furnace, a reduced-order thermodynamic model is constructed to achieve real-time reconstruction of dynamic thermal field cloud maps across the entire furnace area. This enables visualization and monitoring of the real-time temperature curves of each device on the PCB board and automatic identification of cold and hot spot regions in the thermal field. In the above process, firstly, a full-scale fluid thermodynamic simulation of the reflow oven is performed using FluentCFD simulation software to obtain a basic thermal field distribution model within the oven. Then, real-time temperature data collected by a multi-node thermocouple array arranged within the oven and an infrared thermal imaging sensor inside the oven are fused with the CFD simulation model to construct a reduced-order thermodynamic model with computational efficiency meeting real-time requirements. This enables real-time reconstruction of the dynamic thermal field cloud map of the entire oven area, with a thermal field resolution of one temperature node per square centimeter. Based on the reconstructed thermal field cloud map, the real-time temperature curves of each BGA, QFP, and other devices on the PCB board can be visualized and continuously monitored, automatically identifying cold and hot spot areas in the thermal field, providing full-dimensional thermal field data for reflow soldering parameter optimization. Based on the above, the thermal field order reduction model is constructed using eigenorthogonal decomposition combined with Galerkin projection, and its functional expression is:

[0026] in, The furnace / PCB thermal field temperature is represented as a reduced-order model (ROM) reconstruction (unit: °C). Spatial coordinates (unit: mm). Time (unit: seconds); time constant It is a least-squares fit based on sensor data, representing the time coefficient vector (dimension: The formula describes the dynamic change of the thermal field over time, and its update is:

[0027] In the formula, The interpolation matrix of the basis functions at the sensor locations (dimension: number of sensors) ), The POD basis function is the sensor's measured temperature vector (unit: °C, dimension: number of sensors × 1). It is constructed based on the CFD snapshot matrix X. It was before The offline construction formula for the left singular vector is:

[0028] In the formula, The CFD snapshot matrix (dimension: number of spatial grids × number of snapshots N) is constructed from offline CFD simulation results. The order of reduction is usually r = 20~50 to ensure energy retention > 99%; Based on the above, , Offline thermal field simulation of the reflow oven was performed using CFD software such as Fluent to obtain temperature snapshots (N=100~500 sets) under different operating conditions, and then the results were calculated using the POD algorithm (SVD decomposition). Based on the installation coordinates of the thermocouples / infrared thermal imaging sensors inside the furnace, the basis functions are... Spatial interpolation is performed to obtain; Real-time data is collected using a multi-node thermocouple array inside the furnace (measurement accuracy ±0.1℃) and an infrared thermal imager (sampling frequency ≥10Hz). Based on real time It can be calculated online using the least squares formula.

[0029] S3: Derive the SMT production line capacity optimization scheme, conduct multi-objective virtual experiments based on digital twins, take the production quality, efficiency, energy consumption and equipment life of the SMT production line as the collaborative optimization objectives, perform Pareto optimization through multi-objective evolutionary algorithm, output the non-dominated optimal solution set of process parameters, and after virtual verification, send the feasible scheme to the physical production line for execution. Furthermore, in step S3, the optimization objectives of the multi-objective virtual experiment include solder paste printing yield, mounting accuracy, reflow soldering defect rate, production efficiency, equipment energy consumption, and consumable wear. The optimization variables include squeegee pressure, speed, and demolding parameters in the solder paste printing process; mounting coordinates, angle, and pressure parameters in the mounting process; and temperature zone and chain speed parameters in the reflow soldering process. The constraints are the process specifications and equipment capability boundaries in the regular twin layer. The multi-objective evolutionary algorithm adopts the non-dominated sorting genetic algorithm NSGA-II with elitist strategy, and conducts multiple sets of virtual printing, virtual mounting and virtual reflow soldering simulation experiments in parallel. It outputs Pareto non-dominated optimal solution sets that meet the constraints. Each solution set corresponds to a set of process parameter schemes and corresponding optimization objective values. Based on the above, the fitness of the multi-objective Pareto optimization adopts the non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy. The fitness function is a multi-objective vector, and its functional expression is as follows: The specific form of each sub-objective function is as follows: Yield target The (maximization) formula is: in, The defect weights (satisfying w1+w2+w3+w4=1) are determined by expert experience or the AHP method. The probability of defects such as bridges, tin beads, monuments, and cavities (dimensionless, range 0~1). efficiency goals The (maximization) formula is: in, Cycle time for a single PCB production unit (unit: seconds); Energy consumption target (Minimize, negate to maximize) The formula is: in, Real-time power (unit: W) for printing press, pick-and-place machine, and reflow oven; Equipment life target The formula for maximizing (i.e. minimizing) consumable waste is: in, The weights for the wear of the steel mesh and scraper (satisfying k1+k2=1). Wear amount of stencil produced for a single PCB (unit: μm). This represents the maximum permissible wear of the steel mesh (unit: μm). Scraper wear rate for a single PCB (unit: μm). The maximum permissible wear of the scraper (unit: μm); Its defect probability The logistic regression model based on digital twin simulation is as follows:

[0030] In the formula, These are the logistic regression coefficients, obtained from historical defect data; Based on the above, The Analytic Hierarchy Process (AHP) was used to invite process experts to score the defect type and the importance of consumables in pairs to determine the defect type. Based on historical production data (process parameters + defect detection results), a logistic regression model was trained using maximum likelihood estimation. Lifespan specifications or historical data from consumable suppliers; These are the optimization variables for the NSGA-II algorithm, randomly generated or obtained through crossover mutation within the constraints. Output through virtual printing / mounting / reflow soldering simulation using digital twins.

[0031] The selected process parameter scheme is first verified in the digital twin to complete the entire process virtual verification. After confirming that there are no quality or equipment risks, it is then sent to the physical production line equipment for execution. Based on the above, before adjusting the process parameters on the physical production line, a multi-objective virtual experiment is first conducted in the digital twin environment. The collaborative optimization objectives are solder paste printing yield, mounting accuracy, reflow soldering defect rate, production efficiency, equipment energy consumption, and consumable wear. The optimization variables are squeegee pressure, squeegee speed, and demolding parameters in the printing process, mounting coordinates, mounting angle, and mounting pressure in the mounting process, and the set temperature and chain speed of each temperature zone in the reflow soldering process. The process specifications and equipment safety boundaries in the regular twin layer are used as constraints. The non-dominated sorting genetic algorithm (NSGA-II) with an elite strategy is used to conduct thousands of virtual printing, virtual mounting, and virtual reflow soldering simulation experiments in parallel. The key quality indicators such as bridging probability, solder ball probability, tombstoning defect rate, and solder joint void rate under different parameter combinations are quickly calculated, as well as the corresponding production efficiency, energy consumption, and consumable wear data.

[0032] After the simulation is completed, the system outputs a set of Pareto non-dominated optimal solutions that satisfy the constraints. Each solution set corresponds to a set of process parameter schemes, and the optimization objective values ​​for each scheme are labeled. A typical output scheme is as follows: Option A: Overall welding yield is 99.82%, production efficiency is the same as the benchmark option, but equipment energy consumption is 9.7% higher than the benchmark. Option B: Overall welding yield is 99.51%, production efficiency is 5.2% higher than the benchmark, and equipment energy consumption is the same as the benchmark; Option C: Overall welding yield is 99.65%, equipment energy consumption is reduced by 8.3% compared to the benchmark, and production efficiency is on par with the benchmark.

[0033] Process engineers can select the optimal solution based on the priority requirements of the current production order. The system first performs a full-process virtual verification of the selected solution in the digital twin to confirm that there are no quality risks, no equipment overruns, etc., and then sends the parameter solution to the corresponding equipment on the physical production line through the industrial bus for execution, completely avoiding material waste and downtime losses caused by physical trial and error.

[0034] S4: Dynamically correct process parameters. Based on the real-time virtual-real synchronization of the physical production line and the digital twin, the process is simulated and predicted in advance through a reduced-order physical model. Potential quality defects are predicted and feedforward compensation instructions are generated and sent to the physical production line equipment to realize the dynamic correction of the process in real time. Furthermore, in step S4, the sensor and detection equipment data of the physical production line are synchronously mapped to the digital twin with a delay of no more than 10ms to achieve millisecond-level real-time virtual-physical synchronization. The step size of the advanced simulation prediction is no more than 10ms and the prediction time is no less than 5 seconds. This is used to predict potential quality defects such as placement offset, excessive solder paste thickness, and temperature curve deviation from the process range. The feedforward compensation instruction is generated and sent to the corresponding equipment before the defect occurs to complete the real-time correction of process parameters and realize the prevention and correction of process defects in advance. Based on the above, the mounting offset advance prediction and feedforward control adopt a state-space model combined with Kalman filtering for advance prediction, and its functional form is as follows:

[0035] in, The mounting offset state vector at time t (containing Offset, angular offset (unit: μm / °). The process noise and measurement noise (following a zero-mean Gaussian distribution, with covariance matrix ) are... The formula for calculating the feedforward compensation is:

[0036] in, This is the feedforward compensation amount (X / Y axis coordinate correction amount for the pick-and-place machine, unit: μm). For feedforward control gain matrix (designed via LQR or pole placement). To be ahead of time Time prediction offset ; The Kalman filter prediction update formula is:

[0037] In the formula, To control the input vector (including nozzle movement speed, placement pressure, and vacuum level, units: mm / s / N / kPa). This is the state-space system matrix (obtained through system identification). This is the offset vector measured by the vision system of the pick-and-place machine (unit: μm / °). The offset state estimated by Kalman filtering (unit: μm / °). For the predicted and estimated covariance matrix, Here is the Kalman gain matrix. For the dimension of the identity matrix and Consistent (n×n), used for matrix dimension matching, ensuring the validity of matrix operations, and has no physical dimensions.

[0038] Based on the above, Through system identification (such as least squares method), based on the step response data of the pick-and-place machine (input control quantity) Output measured offset If the measurement dimension and the state dimension are inconsistent (e.g., only X / Y offset is measured, but the angle is not measured) when the offline identification is obtained, the matrix dimension and elements are adjusted according to the actual measurement items, and the matrix coefficients are calibrated by system identification (least square method) to ensure that the mapping relationship is consistent with the measured data. Based on the linear quadratic regulator (LQR) design, the weight matrix is ​​determined by process requirements (offset allowable threshold); It is directly generated based on the state dimension n, without the need for additional measurement, and is a standard matrix for mathematical operations; This was obtained based on statistical analysis of historical noise data. Data is collected using the pick-and-place machine's vision positioning system (accuracy ±0.5μm) (updated every 10 components placed). Read data in real time from the encoder and pressure sensor of the pick-and-place machine servo controller; Calculated online using the Kalman filter formula.

[0039] Based on the above, real-time data from various sensors and SPI / AOI inspection equipment on the physical production line are synchronously mapped to the digital twin with a delay of no more than 10ms, achieving millisecond-level real-time virtual-physical synchronization between the physical production line and the digital twin. Based on the reduced-order physical model constructed in step 2, the digital twin performs advanced simulation prediction of the process within the next 5 seconds with a simulation step size of 10ms, predicting potential quality defects such as placement coordinate offset exceeding the process window, solder paste thickness exceeding the standard, and reflow soldering temperature curve deviating from the process range.

[0040] When the digital twin predicts an impending process defect, it immediately triggers feedforward control logic to automatically generate corresponding parameter compensation instructions. These instructions are then sent to the corresponding physical equipment before the defect occurs, enabling real-time correction of the process parameters. Typical application scenarios include: When the digital twin detects the solder paste printing data of the first PCB through SPI and combines it with the stencil deformation model to predict that a specific printing area of ​​the next PCB will have insufficient solder paste thickness, it immediately sends a compensation command to the printing machine to adjust the squeegee pressure and demolding speed of the corresponding area. When the digital twin predicts that nozzle wear will cause the placement offset of the next component to exceed the threshold by using nozzle vacuum data and placement accuracy feedback, it will immediately adjust the compensation value of the placement coordinates to correct the placement position of the next component. When the digital twin predicts that thermal fluctuations in the reflow oven will cause insufficient heating rates for large-size BGA devices in a certain area of ​​the PCB, it immediately adjusts the power of the hot air motor and the temperature setpoint for the corresponding temperature zone to ensure that the device temperature profile meets the process requirements.

[0041] Through the above mechanism, this invention achieves a fundamental leap in SMT quality control from post-inspection and rework to in-process correction and prevention.

[0042] S5: Eliminate model overfitting faults caused by changes in incoming material batches, construct an inheritable product process twin asset library, encapsulate the process data and models of mass-produced products into reusable twin sub-models, and solve the cold start problem of new product introduction and model overfitting problem caused by changes in incoming material batches through similarity retrieval and domain adaptive migration algorithm. Furthermore, in step S5, the twin sub-models stored in the twin asset library include the PCB design data, optimal process parameter combination, defect compensation model, stencil design and wear compensation data, equipment adaptation parameters, and quality traceability data of the corresponding products. During the new product introduction phase, the system extracts key physical features of the product to be manufactured, such as PCB thickness, component package type, pin pitch, component height, and BGA distribution. Through a physical similarity retrieval algorithm, it matches the three most similar historical products in the twin asset library. The process parameters, compensation logic, and constraint rules in the corresponding twin sub-model are then transferred to the digital twin of the new product as the initial process solution for the new product. This eliminates the need for manual blind tuning, thereby improving the initial yield rate of the new product and shortening the introduction cycle. To address the issue of process parameter drift caused by changes in incoming material batches, the system uses a domain adaptive algorithm to quickly update the boundary conditions of the virtual model of the digital twin using the SPI and AOI measured data of the first PCB, adjusts the compensation model of the process parameters, avoids the algorithm model from overfitting historical batch data, improves the adaptability to fluctuations in incoming materials and environmental changes, and prevents sudden quality defects caused by changes in incoming materials. Based on the above, the similarity retrieval of the twin asset database adopts a multi-feature weighted fusion method, and the overall similarity function is:

[0043] in, The total similarity is (dimensionless, ranging from 0 to 1, with greater similarity being closer to 1). Similarity weights (satisfying) The similarity of each sub-sub is learned from historical transfer effects, and the formula for calculating the similarity is as follows: PCB geometric similarity The formula is:

[0044] in, The thickness of the PCB for new / stock products (unit: mm). The PCB area of ​​new / stock products (unit: mm²). Quantity of new / stocked products (dimensionless). The normalized maximum values ​​for thickness, area, and number of panels; Component similarity The formula (based on Jaccard coefficients) is as follows:

[0045] in, For the first Jaccard similarity coefficient (dimensionless) for similar components. The total number of component types (dimensionless). Process similarity The formula is:

[0046] in, This is an indicator function (1 for the same, 0 for different). The parameter fine-tuning formula for domain adaptive migration (based on Bayesian update of the first measured data) is as follows: In the formula, This is the vector of initial process parameters for the new product after migration and fine-tuning. The optimal process parameter vector for the most similar library products. This is the measured quality data of the first PCB (solder paste thickness, mounting offset, etc.). For new product quality data predicted by the product model in the library, The migration gain matrix (obtained from linear regression of historical migration data); Based on the above, the library products and Extract from the twin asset library (linked to MES system, PCB design files, BOM); Based on similarity and transfer effect data from over 100 historical new product introductions, it was obtained through offline training using reinforcement learning or linear regression; new product... Extract from the new product's Gerber file, BOM, and production work order; Quality data of the first PCB was acquired via SPI and a pick-and-place machine vision system. Obtained through online calculation.

[0047] S6: Optimize the joint simulation of equipment status and process parameters. Embed the equipment consumable degradation model in the digital twin to realize the joint simulation optimization of equipment health status and process parameters, balance production continuity and equipment maintenance costs, and output the optimal maintenance window suggestion.

[0048] Furthermore, in step S6, the simulation optimization process is as follows: Embedding the remaining useful life (RUL) prediction models of key consumables and core components in the SMT production line into the digital twin, including stencil tension decay model, squeegee wear model, carrier deformation degradation model, nozzle wear model, and reflow soldering heating tube aging model; by collecting real-time equipment operation data, consumable usage time, and process quality feedback data, the status of the degradation model is updated in real time to accurately predict the remaining useful life of consumables and core components; Based on the above, the dynamic gap between the stencil and the PCB is constructed using a thin-plate elasticity model combined with tension attenuation compensation, and the function form is as follows: in, For stencils and PCBs in Position and dynamic gap at time t (unit: μm). The initial static gap (gap without scraper pressure, in the state of a new steel mesh, unit: μm). The instantaneous deformation of the steel mesh caused by the scraper pressure (unit: μm). Cumulative deformation (unit: μm) caused by stencil tension decay, and stencil deformation caused by scraper pressure. Using the thin plate small deflection bending equation:

[0049] In the formula, This represents the pressure distribution of the scraper at position (x,y) (unit: N / mm²). The length and width (in mm) of the effective printing area of ​​the stencil. The bending stiffness of the steel mesh (unit: N·mm). The calculation formula is:

[0050] In the formula, This refers to the elastic modulus of the steel mesh material (e.g., 193 GPa for stainless steel, unit: GPa). The thickness of the steel mesh (unit: mm). The Poisson's ratio of the steel mesh material (e.g., 0.3 for stainless steel). Deformation due to tension decay An exponential degradation model is adopted:

[0051] in, Let t be the tension of the steel mesh at time t (unit: N / cm). Initial tension of the new steel mesh (unit: N / cm). The stencil tension degradation rate (unit: 1 / printing times). The maximum cumulative deformation (unit: μm) when the tension completely fails. Based on the above, The gap between the stencil and the PCB was obtained by scanning with a laser displacement sensor (accuracy ±0.1μm) without scraper pressure. Material manuals or factory inspection reports from the stencil supplier; Extract from the stencil design file (Gerber); It was obtained by calibration using a tension meter (accuracy ±0.1N / cm) and a laser displacement sensor during the installation of the new steel mesh; The pressure can be calculated by combining the torque sensor (accuracy ±0.01N・m) of the servo motor of the printing press doctor blade with the pressure distribution model, or by directly measuring it by integrating a thin-film pressure sensor on the doctor blade. Data is collected in real time using a steel mesh tension sensor (sampling frequency ≥ 1Hz). It is based on historical tension data and is updated online via maximum likelihood estimation (MLE).

[0052] During process parameter optimization, the digital twin incorporates the health status of consumables as a core constraint into the optimization logic, enabling joint diagnosis and collaborative optimization of equipment health and process parameters. Typical application scenarios include: • When the tension of the stencil decays to 80% of the critical threshold, the digital twin automatically limits the upper limit of the squeegee pressure when optimizing solder paste printing parameters to prevent the stencil tension from decaying further rapidly. At the same time, it simulates the solder paste printing yield under different squeegee pressure parameters and outputs the optimal parameter scheme that balances printing quality and stencil life. When the stencil tension approaches the critical threshold, the digital twin automatically conducts a virtual stencil replacement simulation, simulating the yield improvement, production efficiency changes, downtime costs and consumable costs after replacing the stencil with a new one. It automatically balances the cost-benefit of maintaining continuous production and downtime for stencil replacement, and accurately provides the optimal maintenance window suggestion to avoid process control failure and batch quality defects caused by consumable failure. Based on the above, the remaining useful life (RUL) prediction of the steel mesh adopts an exponential-linear combined degradation model (taking into account both early rapid decay and later linear wear), and its functional form is as follows:

[0053] in, Let t be the tension of the steel mesh at time t (unit: N / cm). Initial tension of the new steel mesh (unit: N / cm). The exponential degradation rate (describes the rapid relaxation of the initial tension of the stencil, unit: 1 / print count) Linear degradation rate (describes the linear decrease in tension caused by wear of the stencil in the later stages, unit: N / (cm·print times)). The number of print runs (dimensionless). To measure noise (which follows a zero-mean Gaussian distribution with variance...) ); The formula for calculating RUL is:

[0054] in, The critical threshold for steel mesh tension (usually 100%) (Unit: N / cm) Remaining lifespan of the stencil (unit: number of prints); Degradation parameters The online update uses Bayesian estimation:

[0055] In the formula, It is the probability density function; Based on the above, The new steel mesh was calibrated using a tension meter during installation (the average value was taken after 5 repeated measurements). Based on historical enterprise data, the threshold at which the defect rate significantly increases when the statistical tension is below this value; prior distribution. Based on historical full life cycle data of the same type of steel mesh, conjugate priors (such as normal-inverse gamma distribution) are used to determine the priors; After every 100 PCBs are printed, an automatic tension detection device (integrated into the printing press) measures the tension once. Based on accumulated tension data Online Bayesian updates are performed using the Markov Chain Monte Carlo (MCMC) method. First arrival time was calculated by numerically solving the degradation model.

[0056] In this embodiment, for various extremely complex scenarios in SMT production, targeted adaptive optimization is achieved through digital twins, specifically including: High heat capacity difference mixed assembly board reflow soldering scenario: Based on the thermal field digital twin, the device areas with different heat capacities on the PCB are divided into zones. For the different heat capacity characteristics of large-size BGA and small-size 0201 devices, the power and chain speed of the reflow oven temperature zone are coordinated and controlled. At the same time, the low-temperature area on the PCB is identified by infrared thermal imaging in the oven. Based on the above, real-time hot air adjustment in the stencil area enables selective thermal compensation, ensuring that devices with different heat capacities are all within the optimal reflow temperature curve range, thus solving the problems of tombstoning, cold welding, and void defects in mixed-assembly boards.

[0057] Wave soldering scenario: Based on visual recognition, the layout of components and pin orientation on the PCB are obtained. The flux spraying coverage effect and wave contact state during the wave soldering process are simulated through digital twins to achieve closed-loop control of flux spraying and dynamically adjust the spraying flow rate and spraying area. At the same time, based on the simulation results, the wave height and the frequency and amplitude of the turbulence wave are dynamically adjusted to optimize the thermal field zoning distribution in the preheating zone and eliminate soldering defects caused by component shadows.

[0058] FPC flexible circuit board production scenario: Based on the FPC expansion and contraction prediction model of digital twin, combined with the measured data of previous printing and testing, the local expansion, contraction and deformation of FPC in different process stages are predicted, realizing independent local coordinate correction in the printing and placement stages; during the printing and placement process, based on the real-time feedback of pressure sensors, the force-position hybrid control of the Z-axis is realized to avoid FPC deformation under pressure; at the same time, the deformation degradation model of the carrier is incorporated into the digital twin to realize real-time compensation of carrier tolerance and solve the problem of insufficient FPC placement accuracy.

[0059] Nitrogen reflow soldering low void ratio control scenario: A coupled model of oxygen concentration, temperature, and solder joint void ratio is constructed in a digital twin to achieve coordinated control of oxygen concentration and reflow temperature curve; Based on real-time collected oxygen concentration data in the furnace and solder joint void ratio detection data of PCB devices, the optimal nitrogen flow rate and temperature parameters are optimized in reverse through a Bayesian optimization algorithm. At the same time, independent zone control of nitrogen flow rate is implemented for different temperature zones of the reflow oven, reducing nitrogen consumption while ensuring low solder joint void ratio.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for adaptive optimization of SMT production line process parameters based on digital twinning, characterized in that, Includes the following steps: S1: Construct a digital twin to map SMT production line elements, and construct a four-dimensional hierarchical digital twin of the SMT production line. The four-dimensional hierarchical digital twin includes a geometric twin layer, a physical twin layer, a behavioral twin layer and a rule twin layer coupled in sequence, so as to realize the real-time mapping of all elements of the physical entity of the SMT production line and the digital twin. S2: Reconstruction model to calculate the process status of SMT production line. Based on the reduced-order model, virtual sensing mechanism and key physical field reconstruction model are constructed. The unmeasurable key physical parameters are calculated in real time through measurable process and equipment data, filling the blind spot of physical measurement and obtaining full-dimensional process status data of SMT production line. S3: Derive the SMT production line capacity optimization scheme, conduct multi-objective virtual experiments based on the digital twin, take the production quality, efficiency, energy consumption and equipment life of the SMT production line as the collaborative optimization objectives, perform Pareto optimization through multi-objective evolutionary algorithm, output the non-dominated optimal solution set of process parameters, and after virtual verification, send the feasible scheme to the physical production line for execution. S4: Dynamically correct process parameters. Based on the real-time virtual-real synchronization of the physical production line and the digital twin, the process is simulated and predicted in advance through a reduced-order physical model. Potential quality defects are predicted and feedforward compensation instructions are generated and sent to the physical production line equipment to realize the dynamic correction of the process in real time. S5: Eliminate model overfitting faults caused by changes in incoming material batches, construct an inheritable product process twin asset library, encapsulate the process data and models of mass-produced products into reusable twin sub-models, and solve the cold start problem of new product introduction and model overfitting problem caused by changes in incoming material batches through similarity retrieval and domain adaptive migration algorithm. S6: Optimize the joint simulation of equipment status and process parameters. Embed the equipment consumable degradation model in the digital twin to realize the joint simulation optimization of equipment health status and process parameters, balance production continuity and equipment maintenance costs, and output the optimal maintenance window suggestion.

2. The digital-twin-based SMT production line process parameter self-adaptive optimization method according to claim 1, characterized in that, In step S1, the geometric twin layer is a high-precision 3D model of all equipment and tooling in the SMT production line, covering the three-dimensional geometric models of the placement motion mechanism, reflow oven structure, scraper mechanism, PCB support pin array, stencil and carrier. The physical twin layer collects the operating physical parameters of the SMT production line equipment in real time through an IoT sensor network, including scraper servo motor torque and speed, reflow oven temperature and hot air parameters, stencil tension data, equipment vibration data, and PCB positioning data, and maps the collected data to the digital twin in real time to achieve synchronization between the physical entity and the digital model. The behavioral twin layer constructs a dynamic mapping model of process parameters and physical response, which is used to output the corresponding physical response results of stencil deformation, solder paste filling volume, device temperature curve, and placement accuracy change in real time after inputting the process parameter adjustment amount. The rule twin layer transforms the IPC electronic component acceptability standard, enterprise SMT process specifications, and process expert experience into algorithm-recognizable constraints and embeds them into the decision logic of the digital twin.

3. The digital-twin-based SMT production line process parameter self-adaptive optimization method according to claim 1, characterized in that, In step S2, the key physical field reconstruction model includes a reflow oven dynamic thermal field reconstruction model, and the construction steps of the reflow oven dynamic thermal field reconstruction model are as follows: S21: Obtain the basic thermal field distribution model of the reflow oven chamber through CFD fluid thermodynamics simulation; S22: By integrating real-time measurement data from multi-node thermocouple arrays and infrared thermal imaging sensors within the furnace, a reduced-order thermodynamic model is constructed to achieve real-time reconstruction of dynamic thermal field cloud maps across the entire furnace area. This enables visualization and monitoring of the real-time temperature curves of each device on the PCB board and automatic identification of cold and hot spot regions in the thermal field.

4. The digital-twin-based SMT production line process parameter self-adaptive optimization method according to claim 1, characterized in that, In step S2, the virtual sensing mechanism calculates the values ​​of unmeasurable parameters in real time based on a reduced-order model and combined with real-time acquired measurable process and equipment data for parameters that cannot be directly measured by physical sensors, such as the dynamic gap between the stencil and the PCB, the real-time temperature inside the BGA solder joints, and the deformation and displacement of the carrier.

5. The adaptive optimization method for SMT production line process parameters based on digital twins according to claim 1, characterized in that, In step S3, the optimization objectives of the multi-objective virtual experiment include solder paste printing yield, mounting accuracy, reflow soldering defect rate, production efficiency, equipment energy consumption, and consumable wear. The optimization variables include squeegee pressure, speed, and demolding parameters in the solder paste printing process; mounting coordinates, angles, and pressure parameters in the mounting process; and temperature zone temperature and chain speed parameters in the reflow soldering process. The constraints are the process specifications and equipment capability boundaries in the regular twin layer. The multi-objective evolutionary algorithm adopts the non-dominated sorting genetic algorithm NSGA-II with an elite strategy to conduct multiple sets of virtual printing, virtual mounting, and virtual reflow soldering simulation experiments in parallel, and outputs a Pareto non-dominated optimal solution set that meets the constraints. Each solution set corresponds to a set of process parameter schemes and corresponding optimization objective values. The selected process parameter schemes are first fully virtual verified in the digital twin. After confirming that there are no quality or equipment risks, they are then sent to the physical production line equipment for execution.

6. The adaptive optimization method for SMT production line process parameters based on digital twins according to claim 1, characterized in that, In step S4, the data from the sensors and detection equipment of the physical production line are synchronously mapped to the digital twin to achieve millisecond-level real-time virtual-physical synchronization. The advanced simulation prediction is used to predict potential quality defects such as mounting offset, excessive solder paste thickness, and temperature curve deviation from the process range. The feedforward compensation command is generated and sent to the corresponding equipment before the defect occurs to complete the real-time correction of process parameters and realize the prevention and correction of process defects in advance.

7. The adaptive optimization method for SMT production line process parameters based on digital twins according to claim 1, characterized in that, In step S5, the twin sub-models stored in the twin asset library include PCB design data, optimal process parameter combinations, defect compensation models, stencil design and wear compensation data, equipment adaptation parameters, and quality traceability data for the corresponding products. In the new product introduction stage, the system extracts key features of the product to be manufactured, such as PCB thickness, device package type, pin spacing, device height, and panelization method. Through a physical similarity retrieval algorithm, it matches the historical product with the highest similarity in the twin asset library, transfers the corresponding twin sub-model, and generates the initial process plan for the new product. In response to changes in incoming material batches, the system uses a domain adaptive algorithm to quickly update the model boundary conditions of the digital twin using the measured inspection data of the first PCB, adjusts the process parameter compensation model, and avoids the model overfitting to historical batch data.

8. The adaptive optimization method for SMT production line process parameters based on digital twins according to claim 1, characterized in that, In step S6, the equipment consumable degradation model includes a steel mesh tension decay model, a scraper wear model, a carrier deformation degradation model, a nozzle wear model, and a reflow soldering heating tube aging model. These models are used to predict the remaining service life (RUL) of key consumables and components in real time. The co-simulation optimization uses the health status of consumables as a constraint condition for process parameter optimization. During parameter optimization, process quality and consumable service life are considered simultaneously. Before the consumable status reaches a critical threshold, the production benefits and costs after consumable replacement are simulated through virtual simulation to balance production continuity and maintenance costs, and to output the optimal maintenance window suggestion.

Images