High multilayer board lamination layer deviation real-time prediction and compensation control method

Abstract

The invention provides a high multilayer board lamination layer deviation real-time prediction and compensation control method, and the method comprises the steps: deploying a multi-point distributed sensing array in lamination equipment, collecting the data of temperature, pressure, medium flow, real-time layer deviation and the like of each stage, extracting coupling features significantly related to layer deviation response, constructing stage coupling state fingerprints, and carrying out the real-time prediction and compensation control of the lamination layer deviation of a high multilayer board. A lightweight graph neural network model is utilized to predict and generate a self-adaptive compensation instruction of a pressing stage according to the time sequence data, and coordinated control over pressure, temperature and pressure maintaining time is achieved; according to the method, the layer deviation residual error is judged in combination with real-time embedded optical measurement, and if abnormal triggering occurs, the compensation strategy is backtracked and optimized based on the historical knowledge graph, so that the layer deviation control precision and the production stability in the lamination process of the multi-layer board are remarkably improved, and intelligent dynamic compensation and quality closed-loop optimization driven by data are realized.

Description

Technical Field

[0001] This invention relates to the field of precision compensation control and process parameter optimization technology for high multilayer printed circuit board lamination, and particularly to a method for real-time prediction and compensation control of lamination layer deviation in high multilayer boards. Background Technology

[0002] In the current field of high-multilayer printed circuit board (PCB) lamination, layer misalignment control and process parameter optimization have become important research directions. In traditional multilayer PCB lamination processes, interlayer alignment accuracy is easily affected by multiple variables such as temperature, pressure, and resin flow. To ensure product reliability, the industry commonly employs control methods such as staged process parameter setting and setpoint compensation. Mainstream lamination compensation strategies are typically based on preset models or empirical parameter adjustments, relying heavily on feedback from single physical variables, such as temperature or pressure trajectory curves. They lack fine-grained responses to inherent process disturbances, and compensation strategies are mostly independent of the actual process parameter design during lamination, making it difficult to achieve deep synergy between process parameters and control strategies. This situation limits the improvement of lamination accuracy and makes it difficult to guarantee product consistency. Currently, common layer bias prediction and compensation schemes in existing technologies mainly fall into three categories: First, control methods based on PID regulation or conventional LSTM timing prediction map changes in process parameters to fixed compensation commands. However, this type of technology cannot dynamically sense transitions in physical states between stages, easily leading to the accumulation of layer bias residuals. Second, schemes utilizing multi-source sensor fusion, image comparison, or thermo-mechanical coupling simulation achieve a certain degree of data-driven optimization. However, these methods often rely on external quality assessment closed-loop or global modeling, requiring high equipment modification standards and exhibiting poor real-time performance and adaptability. Third, compensation is achieved for certain processes through methods such as mapping press voltage and current disturbances. However, this cannot fully describe the complex dynamic state of the pressing process and is difficult to support multi-stage adaptive optimization. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for real-time prediction and compensation control of lamination layer deviation in multilayer boards.

[0004] The technical solution of this invention is implemented as follows: a method for real-time prediction and compensation control of lamination layer deviation in high-multilayer boards, comprising: S1: Deploy a multi-point distributed sensor array in the high-multilayer board lamination equipment to synchronously collect the local temperature gradient, pressure transmission delay time, medium flow rate and real-time layer offset optical measurement values ​​of each lamination stage, forming a spatiotemporal sampling sequence with stage labels. S2: Perform causal sensitivity analysis on the spatiotemporal sampling sequence with stage labels, remove redundant parameters whose correlation with the layer bias response is lower than a preset threshold, and retain the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient as core coupling features. S3: Based on the core coupling features and the current pressing stage number, construct a stage coupling state fingerprint, wherein the stage coupling state fingerprint is a four-dimensional feature vector containing the temperature-pressure coupling phase difference, the interlayer slip accumulation, the resin flow resistance coefficient, and the current pressing stage number. S4: Construct a lightweight graph neural network model, where nodes represent pressing stages, and edge weights are determined based on the Euclidean distance of the stage coupling state fingerprint between adjacent stages and the rate of change of layer offset residuals. The training objective is to minimize the weighted deviation between the predicted layer offset and the measured layer offset in the next stage, and to constrain the output of the compensation command to meet the physical limiting and step response characteristics of the press actuator. S5: After each pressing stage is completed, a stage coupling state fingerprint is generated based on the data collected in the current stage. The stage coupling state fingerprint is input into the lightweight graph neural network model to obtain a compensation instruction set that adapts to the physical constraints and disturbance characteristics of the current stage. The compensation instruction set includes pressure gradient slope correction, temperature plateau offset, and pressure holding time increment. S6: Automatically map the compensation instruction set to the control word table of the programmable logic controller of the pressing equipment, and perform compensation operations for pressure gradient slope correction, temperature plateau offset and pressure holding time increment; S7: Within a preset time window before the end of each pressing stage, call the embedded optical measurement module to perform sub-pixel positioning of the board edge markers, calculate the actual layer offset residual, and determine whether the absolute value of the layer offset residual exceeds the preset residual threshold or whether the direction changes abruptly. S8: If the layer offset residual is determined to meet the triggering condition, the stage coupling state fingerprint combination of the previous two pressing stages is traced back, historical similar fingerprint clusters are retrieved, the corresponding compensation strategy fine-tuning rules are retrieved from the knowledge graph, and an enhanced compensation instruction with confidence label is generated.

[0005] The present invention provides a real-time prediction and compensation control method for lamination layer deviation in multilayer boards, which has the following beneficial effects: (1) This invention constructs a staged spatiotemporal data acquisition mechanism based on a multi-point distributed sensor array and introduces a causal sensitivity screening and "coupled state fingerprint" encoding strategy. This effectively overcomes the problems of traditional pressing control relying on preset models or empirical parameters, only single variable feedback, inability to dynamically perceive multi-variable coupled disturbances, and low modeling accuracy, response lag, and easy accumulation of layer deviation residuals due to parameter redundancy and unclear feature coupling. This invention retains only high causal correlation features such as temperature-pressure phase difference, interlayer slip integral and resin flow resistance coefficient, forming a four-dimensional state representation vector with clear physical meaning, which significantly improves the ability to identify key driving factors of layer deviation evolution. Combined with the information enhancement structure of stage labels, the system can accurately capture the dynamic transition characteristics between each process stage without global thermo-mechanical coupling simulation or external quality detection feedback, which greatly improves the real-time performance and robustness of process state identification and provides a reliable basis for subsequent compensation decisions. (2) This invention uses a lightweight graph neural network architecture to model the state transition relationship between the pressing stages. It uses the "coupled state fingerprint" as the node input and integrates the Euclidean distance and the rate of change of the layer offset residual between adjacent stages with the edge weights, thus realizing explicit modeling of the evolution trend of the process path. The network is trained with the goal of minimizing the weighted deviation between the predicted layer offset and the measured value, and embeds the physical constraints of the press actuator. On the basis of ensuring the feasibility of the control commands, it outputs a structured compensation instruction set containing three parameters: pressure gradient slope correction, temperature plateau offset, and pressure holding time increment. This instruction is directly mapped to the PLC control word table to realize closed-loop online optimization and control. Without increasing the hardware complexity, this invention significantly improves the adaptive response capability of the control system to local disturbances and the cross-stage collaborative adjustment accuracy. (3) This invention introduces a phased closed-loop verification and fingerprint recalibration mechanism, which further enhances the system's fault tolerance and strategy generalization ability: Before the end of each phase, the embedded optical module is called to perform sub-pixel level marker point positioning and evaluate the actual layer bias residual in real time. Once the residual exceeds the limit or the direction changes abruptly, the backtracking retrieval process is triggered. Based on historical similar fingerprint clusters, the optimal fine-tuning rules are matched from the knowledge graph to generate enhanced compensation instructions with confidence labels, forming an endogenous closed-loop optimization path of "perception-decision-verification-correction". This mechanism gets rid of the dependence on manual experience adjustment and offline simulation, and has good scalability and engineering deployment adaptability. It is especially suitable for high-layer and high-density PCB boards in precision lamination scenarios under complex working conditions, effectively ensuring the stability of product yield. Attached Figure Description

[0006] Figure 1 This is a flowchart of a real-time prediction and compensation control method for lamination layer deviation in multilayer boards according to the present invention. Figure 2This is a sub-flowchart of a real-time prediction and compensation control method for lamination layer deviation in multilayer boards according to the present invention. Figure 3 This is another sub-flowchart of the real-time prediction and compensation control method for the lamination layer deviation of a multilayer board according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for real-time prediction and compensation control of lamination layer deviation in high-multilayer boards, specifically including: S1: Deploy a multi-point distributed sensor array in the high-multilayer board lamination equipment to synchronously collect the local temperature gradient, pressure transmission delay time, medium flow rate and real-time layer offset optical measurement values ​​of each lamination stage, forming a spatiotemporal sampling sequence with stage labels. S2: Perform causal sensitivity analysis on the spatiotemporal sampling sequence with stage labels, remove redundant parameters whose correlation with the layer bias response is lower than a preset threshold, and retain the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient as core coupling features. S3: Based on the core coupling features and the current pressing stage number, construct a stage coupling state fingerprint, wherein the stage coupling state fingerprint is a four-dimensional feature vector containing the temperature-pressure coupling phase difference, the interlayer slip accumulation, the resin flow resistance coefficient, and the current pressing stage number. S4: Construct a lightweight graph neural network model, where nodes represent pressing stages, and edge weights are determined based on the Euclidean distance of the stage coupling state fingerprint between adjacent stages and the rate of change of layer offset residuals. The training objective is to minimize the weighted deviation between the predicted layer offset and the measured layer offset in the next stage, and to constrain the output of the compensation command to meet the physical limiting and step response characteristics of the press actuator. S5: After each pressing stage is completed, a stage coupling state fingerprint is generated based on the data collected in the current stage. The stage coupling state fingerprint is input into the lightweight graph neural network model to obtain a compensation instruction set that adapts to the physical constraints and disturbance characteristics of the current stage. The compensation instruction set includes pressure gradient slope correction, temperature plateau offset, and pressure holding time increment. S6: Automatically map the compensation instruction set to the control word table of the programmable logic controller of the pressing equipment, and perform compensation operations for pressure gradient slope correction, temperature plateau offset and pressure holding time increment; S7: Within a preset time window before the end of each pressing stage, call the embedded optical measurement module to perform sub-pixel positioning of the board edge markers, calculate the actual layer offset residual, and determine whether the absolute value of the layer offset residual exceeds the preset residual threshold or whether the direction changes abruptly. S8: If the layer offset residual is determined to meet the triggering condition, the stage coupling state fingerprint combination of the previous two pressing stages is traced back, historical similar fingerprint clusters are retrieved, the corresponding compensation strategy fine-tuning rules are retrieved from the knowledge graph, and an enhanced compensation instruction with confidence label is generated.

[0010] Step S1: Deploy a multi-point distributed sensor array in the high-multilayer board lamination equipment to synchronously collect local temperature gradients, pressure transmission delay times, medium flow rates, and real-time layer-biased optical measurements at each lamination stage, forming a spatiotemporal sampling sequence with stage labels. Specifically, this includes: S1.1: Perform sensor point planning for the high-multilayer lamination equipment. Based on the size of the lamination plate and the distribution characteristics of the thermal expansion coefficient of the material, perform spatial discretization analysis to generate a physical layout scheme for the multi-point distributed sensor array. The input conditions are the structural dimension parameters of the high-multilayer lamination equipment, the thermal expansion coefficient distribution curve of the lamination plate material, and the data on the lamination method of each layer. A spatial discretization analysis method (parameters: length, width, height of the pressing plate, thermal expansion coefficient distribution matrix) is used to achieve gridded segmentation of the pressing data spatial model, so as to accurately locate the sensor points in subsequent planning. Furthermore, by using a finite element mesh generation algorithm (parameters: three-dimensional geometric model, element type set to tetrahedral element, element size adjusted according to stress gradient in the pressed plate region), the discretization of temperature and stress distribution within the pressed plate is realized, and the geometric center coordinate set of each element is obtained; Furthermore, by using a thermo-mechanical coupled field calculation method (parameters: thermal expansion coefficient function, boundary conditions set as the four sides of the press plate are fixed and the middle is free to expand), the temperature gradient and corresponding stress value of each grid element are calculated, and a thermal stress distribution matrix is ​​generated; Furthermore, a feature sensitivity screening algorithm (parameters: temperature gradient threshold 0.5℃, stress gradient threshold 5MPa) is used to screen the points of the grid cells and obtain a matrix of candidate sensing points that meet the installation conditions. The Voronoi multi-region coverage optimization algorithm (parameters: candidate sensor point coordinate set, percentage of monitoring area to be covered set to 100%) transforms the results of the previous step into the layout coordinates of a multi-point distributed sensor array, thereby achieving the optimal arrangement of temperature, pressure, medium flow rate and optical measurement modules. For example, in a device with a pressing plate measuring 500mm long × 400mm wide × 5mm thick, the material is high Tg composite resin / fiberglass cloth, and the coefficient of thermal expansion distribution ranges as follows: to A three-dimensional geometric model was established, using tetrahedral elements with a minimum side length of 5mm, resulting in approximately 6500 mesh elements. After inputting boundary conditions, a thermo-mechanical coupling operation was performed, yielding a peak temperature gradient of 2.0℃ / mm in the central region and a peak stress gradient of 8MPa / mm in the corner regions. A temperature threshold of 0.5℃ and a stress threshold of 5MPa were applied for filtering, retaining approximately 320 candidate points. The Voronoi coverage optimization algorithm was used to reduce the candidate points to 32 actual installation points, covering the preset monitoring area. In this layout scheme, the temperature sensor is located in the area with the largest thermal gradient change rate, the pressure sensor covers the stress concentration area, and the optical measurement module is positioned at the edge where layer slippage may occur. Verification showed that during the pressing process, the temperature, pressure, and flow rate data measured by each sensor were within the accuracy range of ±0.5℃, ±1ms, and ±0.1mm / s, ensuring the effectiveness of the layout plan and high-quality data input. S1.2: Based on the physical layout scheme, install an industrial-grade infrared temperature sensor array, a piezoelectric pressure sensor, and a laser interferometric optical measurement module. Use the CAN bus protocol to realize the synchronous triggering configuration of multi-source sensing nodes to form a multi-point distributed sensing array with timestamp alignment capability. S1.3: Perform dynamic calibration on the deployed multi-point distributed sensor array, and use a sliding window mean filtering algorithm to eliminate sensor zero drift and compensate for environmental temperature and humidity disturbances to ensure that the measurement accuracy of local temperature gradient, pressure transmission delay time and medium flow rate reaches ±0.5℃ / ±1ms / ±0.1mm / s. S1.4: Based on the predefined compression stage division rules, the synchronous acquisition process is started, and the raw signals output by the multi-point distributed sensor array are subjected to anti-aliasing sampling and time series alignment processing to obtain the raw spatiotemporal data stream of local temperature gradient, pressure transmission delay time, medium flow rate and real-time layer-biased optical measurement values ​​of each compression stage. S1.5: Perform stage label embedding operation on the original spatiotemporal data stream, segment the data according to the stage number and construct a four-dimensional spatiotemporal matrix to generate a spatiotemporal sampling sequence with stage labels as the input dataset for the layer-partial modeling module. After the multi-point distributed sensing array completes anti-aliasing sampling and time series alignment processing, the original spatiotemporal data stream is called as the input data object. The input includes a multi-source sequence containing local temperature gradient, pressure transmission delay time, medium flow rate and real-time tomographic optical measurement values. The stage sequence number index matching algorithm (parameters: predefined stage division rules, timestamp alignment threshold) is used to realize the stage identification and segmentation classification of raw data, and to group data streams of the same stage into a unified data block; Furthermore, by using a stage label embedding encoding method (parameters: stage number, label mapping matrix), the stage identifier and label vectorization processing of each data block are realized, and a list of parameter data with stage labels is obtained. Furthermore, a feature channel arrangement algorithm (parameters: temperature channel number, pressure channel number, flow rate channel number, optical measurement channel number) is adopted to achieve orderly recombination of data from each channel and generate a feature sequence matrix, ensuring the structural consistency of temperature gradient, pressure delay, flow rate, and layer offset measurements. Furthermore, by using matrix dimension expansion and four-dimensional construction methods (parameters: number of sampling points in the time dimension, number of sensor positions in the spatial dimension, number of parameters in the feature dimension, and stage label values), a four-dimensional spatiotemporal matrix is ​​constructed, combining the four types of information—time, space, features, and stage labels—into a unified matrix representation and generating a complete four-dimensional data structure. By using stage label embedding and four-dimensional matrix construction, the original spatiotemporal data stream from the previous step is transformed into a spatiotemporal sampling sequence with stage labels, achieving the expected technical effect of providing structured and labeled high-precision input data for the layer-partial modeling module. For example, in a certain production batch, the input data is a sequence of raw signals simultaneously collected by 32 temperature sensors, 16 pressure sensors, 8 flow rate sensors, and 4 optical measurement channels, with a timestamp accuracy of 1ms. Using a stage number index matching algorithm, combined with a pre-defined four-stage pressing process rule, the sequence is divided into four stages: preheating, heating and pressurizing, high-temperature holding and pressurizing, and cooling and depressurizing. Using a stage label embedding encoding method, stage numbers 0, 1, 2, and 3 are mapped to unique thermal encoding label vectors of length 4 and appended to each data block. Channel arrangement is performed on the data from each channel to obtain feature matrices for temperature gradient (32-dimensional), pressure delay (16-dimensional), flow rate (8-dimensional), and layer offset measurement (4-dimensional). Based on a matrix dimension expansion method, the time dimension is set to 500 sampling points, the spatial dimension to the number of sensor positions, the feature dimension to 60 parameter values, and the stage dimension to a label length of 4, constructing a four-dimensional spatiotemporal matrix with dimensions of 500×60×4. In this matrix, any time index can be located to a specific spatial location, feature parameters, and compression stage label. The final output spatiotemporal sampling sequence with stage labels has been verified to significantly improve the accuracy and stability of compression layer bias prediction in subsequent modeling stages.

[0011] Step S2: Perform causal sensitivity analysis on the spatiotemporal sampling sequence with stage labels, remove redundant parameters whose correlation with the layer shift response is lower than a preset threshold, and retain the temperature-pressure coupling phase difference, interlayer slip accumulation, and resin flow resistance coefficient as core coupling features. Specifically, this includes: S2.1: Based on the time-series causal relationship model of process parameters and layer slip response in the pressing stage, a causal sensitivity analysis unit is constructed. The input is the local temperature gradient, pressure transmission delay time and medium flow rate in the spatiotemporal sampling sequence with stage labels. The causal strength index of each parameter and layer slip response is calculated by Granger causality test algorithm. The output is the initial causal strength vector of temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient, which provides a quantitative basis for feature screening. The data interface reading module is called to read the spatiotemporal sampling sequence with stage labels to obtain three types of raw sequence data: local temperature gradient, pressure transmission delay time, and medium flow rate, which are used as inputs for the causal sensitivity analysis unit. A temporal causal relationship modeling method (parameters: sampling time interval ≤ 1ms, sequence length ≥ 1000 points) is adopted to construct a Granger causality test calculation framework for the process parameters and layer bias response in the pressing stage, so as to achieve causal relationship fitting of the input sequence; Furthermore, the causal intensity index of each process parameter relative to the layer slip response is calculated using the Granger causality test algorithm (hysteresis order p=3-10, preferably 5, significance level α=0.01-0.1, preferably 0.05), and the initial causal intensity vectors of temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient are obtained. Furthermore, a numerical stability optimization method (parameter: regularization coefficient λ=0.001) is used to make the Granger causality test results more robust and resistant to noise, and a normalization transformation is performed on each component of the initial causality strength vector to facilitate a unified comparison in the subsequent feature selection process. The causal sensitivity analysis unit outputs the initial causal strength vector as a quantification basis, providing input for the feature threshold screening in S2.2. For example, in a high-multilayer PCB lamination production line, the sampling frequency is set to 1000Hz, and the local temperature gradient sequences of four temperature sensing nodes, the pressure transmission delay time sequences of four pressure sensing nodes, and the velocity sequences of three medium flow rate measurement points are simultaneously acquired. These three types of parameters, along with the synchronously acquired layer bias response sequences, are input into the Granger causality test algorithm, where the hysteresis order p is set to 5 to satisfy the maximum response delay of the local thermo-mechanical process during the lamination stage. The Granger test formula is as follows:

[0012] in, This is the statistic for the Granger causality test. For the sum of squared residuals of the constrained model, For the sum of squared residuals of the unconstrained model, For the number of restricted parameters, The total number of samples, This represents the total number of parameters in the model. The F-statistic for each parameter is calculated and compared with a critical value to determine significance. In this embodiment, the causal strength of the temperature-pressure coupling phase difference is 0.62, the causal strength of the interlayer slip accumulation is 0.57, and the causal strength of the resin flow resistance coefficient is 0.44, all exceeding the preset screening threshold of 0.3, providing stable and significant parameter basis for subsequent core coupling feature extraction. S2.2: Perform a threshold comparison operation on the initial causal intensity vector, where the input is the causal intensity index of temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient and a preset correlation threshold of 0.3. Based on the Boolean logic judgment algorithm, redundant parameters with causal intensity below the threshold are eliminated, and a feature screening result vector is generated to ensure that only process parameters with significant causal influence on the interlayer bias response are retained. S2.3: Based on the feature filtering result vector, perform the interlayer slip accumulation integral calculation, where the input is the pressure transmission delay time series and the medium flow rate time series data. The interlayer slip accumulation feature value is generated by the slip rate accumulation integral algorithm. The feature value is output as a key component of the core coupling feature to characterize the mechanical deformation accumulation effect in multi-stage pressing. S2.4: Perform phase difference extraction operation on the temperature-pressure coupling phase difference value, where the input is the synchronous measurement sequence of local temperature gradient and pressure transmission delay time, and the cross-correlation function calculation algorithm is used to generate the temperature-pressure coupling phase difference value feature quantity. The output of this feature quantity is used as the dynamic coupling characterization of the core coupling feature to quantify the temporal misalignment effect of thermal stress and mechanical pressure. The input conditions are the synchronous measurement sequence data of the local temperature gradient and pressure transmission delay time remaining after feature filtering by S2.2, with the time axis aligned and the sampling interval constant; A synchronous time window truncation method (parameter: window length corresponds to the effective sampling interval of the current compression stage) is adopted to achieve paired truncation processing of local temperature gradient and pressure transmission delay time series, ensuring the consistency of the time base for subsequent cross-correlation analysis; Furthermore, by using a cross-correlation function calculation algorithm (parameters: the delay interval is set to ±500 ms, and the normalization coefficient is the reciprocal of the maximum correlation value), the correlation between the two input sequences under different time delays is calculated, and the corresponding cross-correlation curve data is obtained. Furthermore, by using the extreme value search method (parameters: the search range is limited to the above-mentioned delay interval, and the step size is consistent with the sampling interval), the peak index of the cross-correlation coefficient curve is extracted, and the corresponding number of time delay samples is obtained; Furthermore, using a time sample number to physical time conversion algorithm (parameter: sampling frequency of 1 kHz), the peak index is converted into the phase lag time between temperature and pressure, and the phase difference is calculated based on the dominant frequency of the temperature and pressure waveforms during the pressing stage. Apply the following calculation formula:

[0013] in, This is the dominant frequency of the temperature and pressure signals during the pressing stage. This represents the time delay corresponding to the peak value of the waveform. By using the phase difference characteristic value obtained by coupling the phase lag time parameter with the main frequency as the core output of the temperature-pressure coupled phase difference value, the dynamic misalignment effect of thermal stress and mechanical pressure in the time domain is quantified. For example, during the heating and pressurization process in the compression stage, the sampling frequency is set to 1 kHz, the dominant frequency of the temperature gradient sequence is 0.8 Hz, and the dominant frequency of the pressure transfer delay time sequence is 0.8 Hz. The number of delay samples corresponding to the peak value is calculated using the cross-correlation function to be 250. Using the above formula, where... = s, that is, 0.25 s, =0.8 Hz, obtained ≈ The result, in radians, was stored as a core coupling feature and used as input for subsequent stages of coupling state fingerprint construction and compensation strategy model input. Validation showed that under this phase difference condition, the compensation strategy significantly improved the correction amount of layer offset, and the layer offset residual was reduced to within 3 μm. S2.5: Based on the resin flow resistance coefficient calculation process, the input is the measurement data of medium flow rate and interlayer medium thickness. The resin flow resistance coefficient feature value is generated by Darcy's law correction algorithm. The three types of feature values, namely temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient, are integrated into the core coupling feature set. This feature set is output as the basis for constructing the stage coupling state fingerprint, realizing feature dimension simplification and perturbation sensitive factors focus.

[0014] like Figure 2 As shown, step S3 involves constructing a stage coupling state fingerprint based on the core coupling features and the current pressing stage number. This stage coupling state fingerprint is a four-dimensional feature vector containing the temperature-pressure coupling phase difference, interlayer slip accumulation, resin flow resistance coefficient, and the current pressing stage number. Specifically, it includes: S3.1: Read the core coupling feature set output by step S2 to obtain the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient, and simultaneously obtain the current pressing stage number as the input basis for constructing the stage coupling state fingerprint. The core coupling feature set output from step S2 is subjected to structured data reading operations. A high-speed data bus protocol (parameters: CAN bus, with timestamp index) is used to achieve parallel acquisition of temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient. Furthermore, by using a stage sequence number synchronous retrieval algorithm (parameters: pressing stage identifier k, index matching mode), the current pressing stage number is obtained in real time, and the above physical characteristics are logically bound with the stage identifier to form the original input data framework; Furthermore, a missing value detection and mask filling algorithm (parameters: sliding window width w=5, linear interpolation mode) is adopted to realize the integrity verification and missing compensation of temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient, and obtain a core coupling feature set without missing values. Furthermore, a unit consistency verification algorithm (parameters: standard unit set {℃, mm / s, Pa·s, dimensionless stage number}) is used to determine the physical unit consistency of each feature quantity and generate a compliant numerical matrix. By using data type conversion processing, the result of the previous step is converted into a numerical vector, thus realizing the initial input conditions for constructing the stage-coupled state fingerprint; For example, during a high-multilayer PCB lamination process, the system reads the dataset output from step S2 via the CAN bus, where the temperature-pressure coupling phase difference value is... (Unit: rad), cumulative interlayer slip is (Unit: mm), resin flow resistance coefficient is (Unit: Pa·s), the compression stage number is After the data was read, the missing value detection module confirmed that all features were within the valid acquisition period, and the sliding window width was set to... No padding occurred. The unit consistency check results show that all parameters conform to the predefined standard unit set, and the data type conversion module merges them into a four-dimensional vector. As the input basis for constructing the stage-coupled state fingerprint, zero-mean normalization will be performed in S3.2. This process significantly improves the integrity and consistency of the input features, provides a stable feature input source for the graph neural network model, and effectively supports the dynamic generation capability of the multi-stage compression compensation strategy; S3.2: Based on the core coupling features obtained in S3.1, a zero-mean normalization algorithm is performed on the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient to eliminate the dimensional differences between different features and generate a normalized feature vector. S3.3: The standardized feature vector generated in S3.2 is concatenated with the current pressing stage number to construct a four-dimensional vector of the initial stage coupling state fingerprint; S3.4: Perform range boundary verification on the initial stage coupled state fingerprint four-dimensional vector constructed in S3.3 to ensure that each feature value is within the preset physical feasible range and output the verified fingerprint vector; Based on the initial stage coupled state fingerprint four-dimensional vector constructed by S3.3, a physical parameter limit retrieval method is adopted (parameter: preset physical boundary set, where the temperature-pressure coupling phase difference range is set to ±). The range of arc and interlayer slip accumulation is set to 0 to 0.5 mm, the range of resin flow resistance coefficient is set to 0.1 to 5.0 Pa·s, and the range of stage number is set to 1 to N. This enables interval matching verification of each feature component to obtain the boundary pairing result containing each feature. Furthermore, by using the interval reduction algorithm (parameters: upper and lower bounds and the current feature value), the over-boundary determination of the feature value is realized, and the over-boundary state identifier vector is obtained, where each element corresponds to a Boolean value indicating whether a feature component exceeds the physically feasible interval; Furthermore, by using a clipping algorithm (parameters: over-boundary state identifier vector, physical boundary set), physical boundary substitution operation is performed on the feature components determined to be over-boundary, replacing the over-boundary values ​​with the corresponding upper or lower boundary values, and generating a feature vector that has been clipped. Furthermore, by using a numerical consistency verification method (parameters: the feature vector after amplitude limiting and the statistical distribution characteristics of the standardized feature vector), a global consistency verification of the feature values ​​after amplitude limiting is achieved to ensure the feasibility of the combination of each component under the equipment operating conditions, and a consistency verification identifier is output. By using physical parameter limit retrieval and amplitude truncation processing, the initial stage coupling state fingerprint four-dimensional vector in step S3.3 is transformed into a verification pass vector that satisfies the preset physical feasible range and equipment operation constraints, thereby realizing the input security and stability of the stage coupling state fingerprint in subsequent model inference. For example, in the third stage of the high-multilayer PCB lamination process, the initial stage coupling state fingerprint four-dimensional vector is [1.8 rad, 0.62 mm, 6.1 Pa·s, 3]. The physical boundary set is set as the phase difference ± The parameters are: radian, cumulative slip (0-0.5 mm), impedance coefficient (0.1-5.0 Pa·s), and stage number (1-10). A physical parameter limit retrieval method is used to perform interval matching verification on this vector. The results show that the phase difference is within the range, the cumulative slip exceeds the upper limit, the impedance coefficient exceeds the upper limit, and the stage number is within the range. A truncation algorithm is then executed, replacing the cumulative slip of 0.62 mm with 0.5 mm and the impedance coefficient of 6.1 Pa·s with 5.0 Pa·s, resulting in a corrected vector [1.8 rad, 0.5 mm, 5.0 Pa·s, 3]. Next, a numerical consistency check method is used, with the equipment operating conditions corresponding to stage number 3 as a reference, to verify the feasibility and compliance of each feature value combination under the on-site pressing conditions. The final output verification fingerprint vector is [1.8 rad, 0.5 mm, 5.0 Pa·s, 3]. This provides stable control of the compensation strategy generation process in subsequent graph neural network model inference, significantly improving the accuracy and robustness of the predicted layer bias. S3.5: The four-dimensional vector of the stage coupling state fingerprint that has passed the verification in S3.4 is used as a valid output and provided to the subsequent graph neural network model for compensating the dynamic generation of the instruction set.

[0015] like Figure 3 As shown, step S4 involves constructing a lightweight graph neural network model, where nodes represent pressing stages, edge weights are determined based on the Euclidean distance of the stage coupling state fingerprints between adjacent stages and the rate of change of layer offset residuals, and the training objective is to minimize the weighted deviation between the predicted layer offset and the measured layer offset in the next stage, while constraining the compensation command output to meet the physical limiting and step response characteristics of the press actuator. Specifically, this includes: S4.1: Perform node initialization processing on the pressing stage sequence, mapping each pressing stage to an independent node in the graph neural network, generating a graph node set to provide basic input for building the model topology; The four-dimensional vector dataset of stage coupling state fingerprints that passed the verification of the previous stage S3.5 output is used as input conditions. The sequence number of each pressing stage and the corresponding fingerprint feature items are read for the initial configuration of node construction parameters. A stage number mapping method (parameters: stage number, node identifier generation rule) is adopted to uniquely map each pressing stage number to a set of independent node identifiers in the graph neural network, so as to ensure the uniqueness and indexability of stage data in the graph structure. Furthermore, by using a node attribute binding algorithm (parameters: node identifier set, stage coupling state fingerprint), each stage node is bound to its standardized four-dimensional feature vector as a node feature vector, and a set of node attributes with stage feature loads is obtained. Furthermore, a node set densification processing method (parameters: node attribute set, idle node marking strategy) is adopted to complete and remove redundancy from the node set, generate a node set with consecutive ordinal numbers and eliminate duplicate mappings, thus ensuring the integrity of the graph topology. Furthermore, by using a node sequence verification algorithm (parameters: node set, logical order of stage numbers), the logical correctness of the arrangement of stage numbers in the node set is verified, and a node topology legality identifier is generated. Through node set initialization processing, the stage coupling state fingerprint data of the previous step is transformed into a node set that can be used for graph neural network topology construction, so as to realize the accurate mapping and feature initialization of each compression stage in the graph structure. For example, in a 20-layer multi-stage lamination process, the stage number ranges from 1 to 8, and the fingerprint feature values ​​of the stage coupling state read are the phase difference. = rad, slip cumulative S= mm, impedance R= Pa·s and stage number k= The sequence number mapping method is used to map stage 3 to node N3, binding the feature vector [0.12, 0.45, 1.8, 3] to the attribute set of node N3. During node densification, the node numbers are arranged into a continuous set of [1, 2, 3, 4, 5, 6, 7, 8], and duplicate node N2 copies generated during the device testing stage are removed. The node sequence verification algorithm is used to check the correctness of the node set order, confirming that the pressing stage sequence is strictly progressive without jumps. After initialization, node N3 has a unique identifier and a complete feature vector in the graph structure. Subsequently, S4.2 can directly call this feature to calculate Euclidean distance and generate edge weights, thereby ensuring high consistency between the stage coupling state fingerprint and the graph structure nodes, and achieving an accurate mapping effect from the pressing stage to the model structure. S4.2: Based on the stage coupling state fingerprint data, Euclidean distance is calculated on the stage coupling state fingerprints of adjacent pressing stages to obtain the feature difference measure between stages; at the same time, the layer offset residual change rate data is obtained; a linear weighted fusion operation is performed on the feature difference measure and the layer offset residual change rate to generate dynamic edge weight parameters to quantify the correlation strength of state transitions between stages. The input conditions are the set of four-dimensional vectors of stage coupling state fingerprints effectively output by step S3 and the sequence of the rate of change of layer partial residuals corresponding to adjacent stages. The Euclidean distance calculation method (parameter: input is a four-dimensional vector of stage coupling state fingerprint of adjacent compression stages) is used to realize the quantitative measurement of the feature space difference between adjacent stages; Furthermore, through the distance calculation formula

[0016] This method calculates the distance between fingerprint vectors of adjacent stages in a multi-dimensional feature space and obtains a measure of feature difference between stages. Let i be the i-th dimension feature value of the fingerprint at the current stage. The i-th dimension feature value of the fingerprint in adjacent stages; Furthermore, by using a differential calculation method (parameter: input is a sequence of measured layer partial residuals from two consecutive stages), the rate of change of layer partial residuals is extracted, and the rate of change sequence data is generated. Furthermore, through a linear weighted fusion algorithm (parameter: weight coefficients) , (Obtained by fitting historical training data), achieving a combined calculation of feature difference measurement and layer-biased residual change rate, the fusion formula is as follows:

[0017] in The result is the Euclidean distance calculation. The rate of change of layer offset residuals; Furthermore, by using a normalization method (parameter: edge weight sample set), the range of dynamic edge weight parameters is unified, different numerical scales are converted to the [0,1] interval, and a dynamic edge weight matrix for graph neural network models is generated. Through the above chain operation, the feature differences and dynamic response of layer bias error between stage-coupled state fingerprints are transformed into edge weight data with unified dimensions, so as to realize the accurate quantification of the correlation strength of state transitions between stages. For example, on a high-multilayer PCB lamination machine, the fingerprint vectors of adjacent stages are respectively and Substituting into the Euclidean distance formula, we obtain... The rate of change of the layer-specific residual is obtained by the difference between the residuals of the previous stage and the current stage, and the residual values ​​are respectively... μm and μm, rate of change μm / stage. The fusion weighting coefficient is set to... , Substituting into the fusion formula yields = After normalization, the value is mapped to the [0,1] interval to obtain... This serves as the dynamic weight for the corresponding edge. This weight effectively enhances the ability to capture stage state transition patterns during subsequent graph neural network training, significantly reducing the error of the model's predicted layer bias and meeting the process's accuracy requirements for control within ±10μm. S4.3: Based on the graph node set and dynamic edge weight parameters, a loss function is defined for calculation. This loss function performs a weighted absolute deviation calculation on the predicted layer bias and the measured layer bias in the next stage, generating a minimized objective function to guide the direction of model parameter optimization. S4.4: Apply physical limiting constraints to the compensation command output, mapping the pressure gradient slope correction, temperature plateau offset, and pressure holding time increment to the operating range boundary of the press actuator; at the same time, perform step response characteristic constraint processing to ensure that the compensation command output meets the dynamic response time requirements of the press actuator, generating a constraint-compliant compensation command output space; The input conditions are the preliminary compensation instruction set generated by the graph neural network inference module, which includes three structured parameters: pressure gradient slope correction, temperature plateau offset, and pressure holding time increment. The physical limit threshold matrix and step response time constant of the compressor actuator are obtained simultaneously as constraint benchmarks. A limit constraint algorithm (parameters: maximum and minimum allowable values ​​for each channel) is used to implement boundary limit processing for the pressure gradient slope correction amount; further, through saturation truncation function operation, the command parameters exceeding the physical limit are truncated to the limit range, and a correction parameter vector that meets the safety threshold is obtained; Furthermore, a temperature control command limiting mapping algorithm (parameters: maximum platform offset of temperature control channel, step response time constant) is adopted to achieve saturation truncation of temperature platform offset and smoothing of time response, generating a temperature compensation parameter sequence that conforms to thermal inertia characteristics; Furthermore, an incremental timing embedding algorithm for holding pressure duration is adopted (parameters: upper limit of holding pressure time, equipment response delay) to achieve dynamic correction of the incremental holding pressure duration and generate holding pressure compensation parameters that have been limited and time-synchronized verified to ensure that the stability requirements of the mechanical structure are met. The step response characteristic constraint calculation method is adopted (parameter: response time constant). The target step amplitude (A) is used to achieve dynamic response matching for three types of parameters. The step matching deviation of each parameter is calculated using the following formula. :

[0018] in, This is the current step response output value. For the target response value, The number of sampling points is used to adjust parameters to meet the step characteristics. The constraint verification function transforms the result of the previous step, which involves amplitude limiting and step response matching, into a constraint-compliant compensation instruction output space, thereby enabling the executability of the compensation instructions within safe and stable boundaries. For example, in a high-multilayer PCB lamination equipment implementation scenario, the initial compensation instruction set includes a pressure gradient slope correction of 0.15 MPa / s, a temperature plateau offset of 6°C, and a holding time increment of 12 seconds. The press limiting thresholds are pressure ±0.12 MPa / s, temperature ±5°C, and holding time ±10 seconds. The limiting constraint process truncates the pressure correction to 0.12 MPa / s, the temperature offset to 5°C, and the holding time increment to 10 seconds. In the step response characteristic constraint, the equipment response time constant... =2.5s, the target step amplitude A is taken as the target value of the temperature plateau offset of 5℃, the number of sampling points n=10, the collected response data y=[5.2,5.1,5.0,4.9,5.0,5.0,5.0,5.1,5.0,5.0], substitute into the formula to calculate the matching deviation: The result was E = 0.012 ℃², and after adjustment, the temperature command maintained the response trajectory corresponding to a 5℃ step value. This constraint-compliant compensation command output space significantly improved the thermo-mechanical control stability of the pressing process during subsequent press execution. S4.5: Perform backpropagation iterative training on the graph neural network model using the training dataset with stage labels. Based on minimizing the output space of the compensation instructions that comply with the constraints, update the model parameters until the loss function converges, and obtain the optimized lightweight graph neural network model parameter set. After the lightweight graph neural network model is built and the loss function and constraints are defined, the input is a training dataset with stage labels, including the four-dimensional feature vector of stage coupling state fingerprint, the weight parameters of adjacent stage edges, and the corresponding measured layer bias data. The batch gradient descent method (parameters: batch size m=64, learning rate η=0.001) is used to realize the calculation of the first iteration update of the model parameters; Furthermore, through forward propagation calculation (parameters: number of input layer nodes = 4, number of hidden layer nodes = 32), the state transition simulation under the action of node feature aggregation and edge weights is realized, and the bias value of the prediction layer in the next stage and the initial value of the compensation instruction in the current stage are obtained. Furthermore, a loss function calculation method is adopted (parameter: weighted absolute deviation weight). Constraints on violation penalty coefficients This enables the weighted difference assessment between predicted layer bias and measured layer bias, and generates an overall loss value with a constraint penalty term; Furthermore, the backpropagation algorithm (parameters: ReLU activation function, Adam optimizer) is employed to calculate the gradients of the weight matrices and bias parameters of each layer based on the gradient chain rule, and the update rule is used. Perform iterative updates on the parameters, where For the current model parameter set, The learning rate is 0.001-0.005. The gradient vector of the loss function; Furthermore, compliance testing methods are implemented by constraining the parameters (pressure gradient slope limit). Temperature plateau offset limit Incremental limit of pressure holding time This allows for the verification of the physical limiting and step response characteristics of the updated compensation instruction space, the removal of non-compliant parameters, and the generation of a corrected compliant model output space. Furthermore, training is performed iteratively until the loss function value converges (the number of iterations is no less than 100, and the threshold is met). This enables the stable acquisition of the optimized lightweight graph neural network model parameter set; Through the above chain algorithm and constraint verification processing method, the model initialization result of the previous step is transformed into high-precision graph neural network parameters that converge within the constraints, thereby achieving a stable improvement in the dynamic generation capability of compensation instructions in the multi-stage pressing process. For example, in the dynamic compensation strategy training scenario, the input training data contains 1000 stage-labeled samples. Each sample contains a four-dimensional stage coupling state fingerprint (temperature-pressure coupling phase difference ranging from 0.2 to 1.2 rad, interlayer slip accumulation ranging from 0.05 to 0.30 mm, resin flow resistance coefficient ranging from 1.5 to 6.0 Pa·s, stage number 1–4) and the corresponding measured layer bias. An Adam optimizer with a batch size of 64 and a learning rate of 0.001 is used, with ReLU activation function. After 500 iterations, the loss function... From the initial Reduced to the convergence threshold The following conditions were met, and all output compensation commands satisfied the constraints of pressure gradient slope ≤ 5 N / mm, temperature plateau offset ≤ 2℃, and holding time increment ≤ 15 s. With increased data perturbation range at different stages (interlayer slip accumulation increased to 0.35 mm), the model training convergence time was extended to 700 iterations, and the difference between the predicted and measured layer offsets was reduced to ±4 μm, significantly improving the stability and accuracy of multi-stage compression compensation control.

[0019] Step S5: After each pressing stage, a stage coupling state fingerprint is generated based on the data collected in the current stage. This fingerprint is then input into the lightweight graph neural network model to obtain a compensation instruction set adapted to the physical constraints and disturbance characteristics of the current stage. This compensation instruction set includes pressure gradient slope correction, temperature plateau offset, and pressure holding time increment. Specifically, it includes: S5.1: The core coupling feature data output by the real-time acquisition system of the pressing process is called and processed. Based on the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient as input conditions, the data interface reading operation is executed to obtain the core coupling feature dataset after the end of the current stage, so as to provide the initial input basis for the generation of stage coupling state fingerprint. S5.2: Based on the core coupling feature dataset and the current pressing stage number as input conditions, a four-dimensional feature vector construction process is performed through the feature vector combination algorithm to generate a stage coupling state fingerprint containing the temperature-pressure coupling phase difference, interlayer slip accumulation, resin flow resistance coefficient and the current pressing stage number, so as to form a standardized input representation of the lightweight graph neural network model. Based on the core coupling feature dataset and the current pressing stage number as input conditions, a feature vector combination algorithm (parameter: feature dimension = 4, splicing order depends on physical property sensitivity sorting) is adopted to realize the structured splicing function of multi-dimensional coupling features; Furthermore, through a feature standardization processing algorithm (parameters: mean μ is taken from the historical sample set, standard deviation σ is estimated by a sliding window), the uniformity of each feature dimension is achieved, and a standardized feature value sequence with eliminated dimension differences is obtained; Furthermore, a data type encoding method is adopted (parameters: temperature-pressure phase difference, slip accumulation, and impedance coefficient are all encoded as floating-point type, and stage number is encoded as integer long character format) to achieve a unified precision representation of feature values ​​and generate an encoding vector with consistent physical semantics; Furthermore, through vector splicing operations (parameter: splicing end order is fixed with temperature-pressure phase difference value at the first position and stage number at the last position), a four-dimensional combined vector is constructed, and stage coupling state fingerprint data containing temperature-pressure coupling phase difference value, interlayer slip accumulation, resin flow resistance coefficient value and current pressing stage number is generated. By using a range boundary verification algorithm (parameter: the boundary interval is taken from the upper and lower limits of the process feasible domain), the result of the previous step is transformed into a verified four-dimensional feature vector, achieving the technical effect of standardized input representation of lightweight graph neural network model; For example, under the lamination condition of a high-multilayer (24-layer) board, the core coupling feature dataset includes the temperature-pressure coupling phase difference Δφ = 0.015 seconds, the interlayer slip accumulation S = 0.28 mm, the resin flow resistance coefficient R = 1.12 N·s / mm, and the current lamination stage number k = 3. A zero-mean normalization algorithm is applied to Δφ, S, and R, as shown in the following formula:

[0020] in, The feature value of the current sample. This is the historical mean of this feature. This represents the historical standard deviation of this feature. In this embodiment, the historical mean of Δφ is μ = 0.010 seconds, and σ = 0.004 seconds. The standardized Δφ feature value is then calculated. =1.25; similarly, S and R are calculated to obtain standardized values ​​of 0.95 and 1.08, respectively. The stage number k is encoded as 3 using integer. In terms of concatenation order, the standardized Δφ is placed first, followed by S and R, and finally the stage number k, forming a vector [1.25, 0.95, 1.08, 3]. The range boundary check is performed on this vector. Δφ∈[-3,3], S∈[-2,2], R∈[-5,5], and k∈[1,6] all satisfy the process feasible region, and the effective stage coupling state fingerprint is obtained and input to the graph neural network. In this embodiment, the generated four-dimensional fingerprint vector significantly improves the model's response matching ability to stage disturbances in subsequent inference, and achieves a significant improvement in the stability and accuracy of the compensation control parameter output; S5.3: Using the stage coupling state fingerprint as input condition, perform data format conversion processing through the model interface protocol to map the four-dimensional feature vector to the node input layer of the lightweight graph neural network model, so as to complete the input preparation for model inference and activate the inter-stage state transition feature transmission. Based on the four-dimensional feature vector of the stage coupling state fingerprint output by S5.2 as input condition, the unified data encapsulation function of the feature vector is implemented by the model interface protocol parsing module (parameters: interface type Ethernet / IP, data encapsulation format JSON). Furthermore, through a data standardization mapping algorithm (parameter: the mapping rule is to normalize the node input domain [-1,1], and use the feature data after zero mean standardization), the numerical domain transformation of the four-dimensional feature vector is realized, and a floating-point array structure matching the input layer of the lightweight graph neural network node is obtained; Furthermore, by utilizing the data structure rearrangement module (parameters: memory alignment is 32 bytes, field order is [temperature-pressure coupling phase difference, interlayer slip accumulation, resin flow resistance coefficient, stage number]), the arrangement of physical feature vectors in memory is optimized, and efficient contiguous memory blocks are generated to improve model input efficiency; Furthermore, a feature tensor quantization processing algorithm (parameters: tensor dimension shape=(1,4), tensor type float32) is used to convert the rearranged and optimized feature vectors into tensor form, so that matrix operations can be performed in the input layer of the graph neural network. Furthermore, by mapping the inter-stage state transition feature activation function (parameters: activation type ReLU, threshold parameter 0.0), the nonlinear feature activation of the input feature vector in the node input layer is realized, and an initial node state matrix adapted to the stage sequence is generated. Through the above data format conversion and feature activation processing, the S5.2 results are transformed into tensor data that has been standardized and structured by the node input layer, realizing the input preparation for model inference and creating conditions for the efficient transmission of state transition features between stages. For example, in a production scenario comprising four pressing stages, the temperature-pressure coupling phase difference is... The cumulative amount of interlayer slip is mm, resin flow resistance coefficient is N·s / mm², the current pressing stage number is After being encapsulated by the model interface protocol parsing module, the data is formatted as JSON: {"phaseDiff":0.45,"slipAcc":1.25,"resistCoef":0.85,"stageNum":2}. Through a normalization mapping algorithm, the feature normalization interval for columns 1 to 3 is [-1,1], while the stage number retains its original value. After the field order is rearranged, the vector becomes [0.45,1.25,0.85,2]. This vector is then processed by feature tensor quantization to form a float32 tensor input with shape=(1,4). The input layer is activated by ReLU while retaining its original values, generating an initial node state matrix for graph neural network inference. This matrix significantly improves the numerical stability of state transition calculations in the subsequent node feature aggregation in S5.4, ensuring that the stage characteristics during the compensation instruction generation process are fully preserved. S5.4: Based on the inference engine of the lightweight graph neural network model as input conditions, perform forward propagation calculation processing, and generate a compensation instruction set including pressure gradient slope correction, temperature plateau offset and pressure holding time increment through node feature aggregation and edge weight optimization, so as to output structured control parameters that adapt to the physical constraints and disturbance characteristics of the current stage. S5.5: Using the compensation instruction set as input conditions, perform constraint verification processing based on the physical limiting and step response characteristics of the press actuator. Verify the feasibility of the compensation parameters through limiting threshold comparison and step response simulation to ensure that the output compensation instruction set meets the equipment operation safety requirements and improves the real-time stability of the control strategy.

[0021] Step S6: Automatically map the compensation instruction set to the control word table of the programmable logic controller of the pressing equipment, and perform compensation operations for pressure gradient slope correction, temperature plateau offset, and holding time increment. Specifically, this includes: S6.1: Extract structured parameters from the compensation instruction set. Based on the triplet data format of pressure gradient slope correction, temperature plateau offset, and pressure holding time increment, perform field segmentation and numerical type conversion to generate a standardized parameter vector containing correction values, unit identifiers, and effective time windows. Based on the triplet data of the compensation instruction set after S5.5 constraint verification as input conditions, a field segmentation algorithm is used (parameters: the triplet separator is a comma ",", and the field identifier sequence number is fixed as 1, 2, 3) to extract independent fields of pressure gradient slope correction, temperature plateau offset and pressure holding time increment. Furthermore, a numerical type conversion algorithm is used (parameters: target type is IEEE 754 standard double-precision floating-point number, conversion precision is...). This process standardizes the types of correction values ​​for each field and yields a continuously numerical correction value dataset. A unit identifier appending processing method is adopted (parameters: pressure gradient unit MPa / s, temperature plateau unit ℃, pressure holding time unit s) to realize the binding operation of each numerical field with its corresponding physical quantity unit, and generate a parameter record set with unit identifier; Furthermore, by using an effective time window calculation method (parameters: the window start point is the current PLC cycle count, and the window size is set according to the dynamic compensation duration output by the graph neural network), the action time boundary of the triplet parameter is determined, and time window data containing start and end timestamps is obtained; By encapsulating the parameters into vectors, the aforementioned continuous numerical correction values, unit identifiers, and effective time windows are combined into a standardized parameter vector according to a fixed field order, thereby achieving format compatibility with the PLC control word table mapping interface. Through the above chained processing method, the compensation instruction set that has been constrained and verified is transformed into a standardized parameter vector that can be directly used for PLC address mapping, thereby realizing the execution preparation of the pressing process compensation parameters at the data structure level. For example, after a multilayer board lamination stage, the triplet data format of the compensation instruction set is "0.025,1.2,15", where the pressure gradient slope correction is 0.025 MPa / s, the temperature plateau offset is 1.2℃, and the holding time increment is 15s. The field segmentation algorithm separates this string into three independent fields by commas, which are then converted into double-precision floating-point numbers 0.025, 1.2, and 15, respectively. The unit identifier appending method binds the unit identifiers of MPa / s, ℃, and s to the three fields respectively; the effective time window calculation method starts with the PLC cycle count of 100, with a window size of 20 cycles, resulting in start and end timestamps (100, 120). Finally, after vectorization and encapsulation, a standardized parameter vector [0.025|MPa / s|(100,120);1.2|℃|(100,120);15|s|(100,120)] is obtained. After being mapped by the PLC control word table, this vector realizes the correction of the pressure control channel, the offset of the temperature platform, and the dynamic extension of the holding time, effectively improving the alignment stability of the pressing process. S6.2: Based on the standardized parameter vector, the control word table address mapping is executed. The predefined register address index table of the programmable logic controller of the pressing equipment is used to linearly scale the pressure gradient slope correction value to determine the target register address and write value of the D / A converter of the pressure control channel. S6.3: Based on the D / A converter target register address and write value of the pressure control channel, perform physical constraint boundary verification, and use the preset limiting threshold of the press actuator to saturate and truncate the temperature platform offset to generate a temperature control command sequence that conforms to the step response characteristics. S6.4: Based on the temperature control instruction sequence that conforms to the step response characteristics, the timing embedding operation of the pressure holding time increment is performed. The clock reference of the pressure holding stage is dynamically offset and corrected through the timer module of the programmable logic controller to form a composite compensation control word that includes pressure gradient slope correction, temperature plateau offset and pressure holding time increment. Based on a temperature control command sequence that conforms to step response characteristics as input, a timing embedding method is employed (parameters: pressure holding time increment ΔT, temperature control command trigger timestamp). The PLC system clock accuracy ε) enables the timing control chain to accurately superimpose the pressure holding time increment onto the current pressure holding stage; Furthermore, the offset calculation algorithm of the timer module (parameters: ΔT, start time of the current pressure holding stage) is used. This enables dynamic offset correction of the clock reference during the pressure holding phase and yields a new clock reference value. ; Furthermore, the original holding voltage timing is updated using a dynamic offset correction formula, as shown in the following formula:

[0022] in, This is the start time of the initial pressure holding phase. To increase the pressure holding time, This is the revised start time of the pressure holding phase; Furthermore, the pressure holding end time is corrected using the formula (parameter: original end time). The holding time increment ΔT is used to adjust the overall holding time, and the formula is as follows:

[0023] in, This is the end time of the initial pressure holding phase. This is the corrected end time; Furthermore, through the PLC control word generation algorithm, the pressure gradient slope correction, temperature plateau offset, and corrected pressure holding time parameters are integrated into a composite compensation control word and mapped to the corresponding logical address space to realize the synchronous drive of the PLC to the actuator. By using a control word sequence packaging process, the aforementioned composite control word is transformed into executable, low-latency control data, thereby achieving a coordinated compensation effect for the three parameters of pressure, temperature, and time. For example, in an industrial press with a rated control accuracy of ±0.1s, the pressure holding time increment ΔT is set to 1.5s, and the start time of the current pressure holding stage is... The original end time is 35.0s. The temperature control command trigger timestamp is 95.0s. The time is 34.8s, and the PLC system clock accuracy ε is 0.01s. Calculate the corrected start time according to the formula. =35.0 + 1.5 = 36.5s, the corrected end time =95.0 + 1.5 = 96.5s. The timer module will... The timeout was updated to 36.5s, and the end time was updated to 96.5s, forming a composite compensation control word with ternary parameters (pressure gradient slope correction: +0.02MPa / s, temperature plateau offset: +1.0℃, holding time increment: +1.5s). This control word was mapped to the pressure channel address 0x1A, temperature channel address 0x1B, and timer address 0x1C via PLC. In the actual pressing process, it extended the holding time. Simultaneously, with synchronous adjustment of pressure and temperature, the absolute value of the layer deviation residual was significantly reduced and stabilized. Verification showed that this method can significantly improve the stability of ±10μm-level alignment accuracy in multi-stage pressing. S6.5: Based on the composite compensation control word, the execution status is verified in real time. The hardware interrupt mechanism of the programmable logic controller is called to perform frequency domain characteristic analysis on the servo motor current ripple signal fed back by the pressing equipment, so as to generate a compensation operation completion confirmation flag and a residual disturbance intensity assessment report.

[0024] Step S7: Within a preset time window before the end of each pressing stage, the embedded optical measurement module is invoked to perform sub-pixel positioning of the board edge markers, calculate the actual layer offset residual, and determine whether the absolute value of the layer offset residual exceeds a preset residual threshold or whether the direction undergoes a sudden change. Specifically, this includes: S7.1: Within a preset time window before the end of each pressing stage, send a synchronous trigger command to the deployed embedded optical measurement module to start the high-resolution image acquisition process, obtain the original image frame sequence of the plate edge markers, perform data caching operations, and add precise timestamp markers to generate a raw image dataset with time-series labels. S7.2: Perform image denoising, edge enhancement, feature point extraction, subpixel-level coordinate calculation and positioning accuracy verification sequentially on the original image dataset with time-series labels to generate the actual coordinate positions of the board edge markers; S7.3: Based on the actual coordinate position of the board edge marker and the theoretical coordinate position in the PCB design data, perform coordinate alignment transformation, layer offset calculation, unit conversion to micrometer level, residual vectorization and noise filtering in sequence to calculate the actual layer offset residual. S7.4: For the actual layer offset residual, perform the following operations in sequence: residual absolute value extraction, comparison with preset residual threshold, direction change rate calculation, direction change determination, and comprehensive judgment of trigger conditions, so as to output the layer offset residual status identifier. For the actual layer partial residual vector data calculated in step S7.3, an absolute value extraction algorithm (parameters: double-precision floating-point operation mode, component-by-component processing method) is used to generate the residual modulus sequence; Furthermore, a threshold comparison algorithm is used (parameter: preset residual threshold). The comparison operation type is "greater than", which realizes the component-by-component comparison between the residual modulus sequence and the residual limit, and obtains the Boolean logic judgment result of the over-limit; Furthermore, the direction change rate calculation method (parameters: adjacent stage residual vector, Euclidean normalization) is adopted to calculate the change amplitude of the current residual direction vector relative to the previous stage residual direction vector, and obtain the scalar value of the direction change rate. Furthermore, a direction change determination algorithm is adopted (parameter: direction change rate threshold). This allows for the comparison of the calculated scalar value of the rate of change of direction with the mutation threshold, and the resulting Boolean logic judgment of the directional mutation. Furthermore, by using a trigger condition comprehensive judgment method (parameters: over-limit Boolean judgment result, directional change judgment result, and logic type "OR"), the logical aggregation of the two judgment results is achieved, and a layer-partial residual state identifier Boolean quantity is generated. Through the above algorithm chain, the results of absolute value extraction, threshold comparison, direction change rate calculation and direction change judgment are transformed into a single layer deviation residual status identifier, realizing the effect of automated detection and early warning technology for abnormal residual at the end of the pressing stage. For example, in a high-multilayer PCB lamination process, the residual vector output by S7.3 at the end of the lamination stage is: The residual vector of the previous stage is Absolute value extraction yields a sequence of modulo values. , and residual limit The second component exceeded the limit after comparison. The rate of change of direction was calculated using the Euclidean normalized dot product method. Greater than the threshold for determining directional abrupt changes When the trigger condition is determined to be true by a comprehensive judgment (logical "OR"), the output layer partial residual status is marked as "abnormal". This status is further encoded and used to trigger the historical fingerprint backtracking process in subsequent steps S7.5 and S8.1, thereby significantly improving the response speed and adaptability of the real-time compensation strategy; S7.5: Based on the layer offset residual state identifier, perform result encoding, confidence assessment, state marker generation, data encapsulation and output transmission operations to generate the layer offset residual state judgment result.

[0025] Step S8: If the layer offset residual is determined to meet the triggering condition, then the stage coupling state fingerprint combination of the previous two pressing stages is traced back, historical similar fingerprint clusters are retrieved, the corresponding compensation strategy fine-tuning rules are retrieved from the knowledge graph, and an enhanced compensation instruction with confidence labels is generated. Specifically, this includes: S8.1: Data cleaning is performed on the stage coupling state fingerprint sequence and interlayer bias residual data during the historical pressing process to remove outlier noise points; based on the cleaned stage coupling state fingerprint sequence, feature normalization is performed to eliminate the dimensional differences of temperature-pressure coupling phase difference, interlayer slip accumulation, and resin flow resistance coefficient; hierarchical clustering algorithm is used to perform clustering on the normalized stage coupling state fingerprint vector to generate a set of historically similar fingerprint clusters; based on the linear regression relationship between interlayer bias residual and compensation command within the set of historically similar fingerprint clusters, compensation strategy fine-tuning rules are extracted; the set of historically similar fingerprint clusters and their associated compensation strategy fine-tuning rules are stored as a knowledge graph to support real-time retrieval operations; S8.2: Obtain the stage coupling state fingerprints of the two previous stages of the current pressing stage; perform vector concatenation on the two stage coupling state fingerprints to form a stage coupling state fingerprint combination vector; perform time series alignment operation on the stage coupling state fingerprint combination vector to match the temporal structure characteristics of historical similar fingerprint clusters; calculate the Euclidean distance between the stage coupling state fingerprint combination vector and the center point of each historical similar fingerprint cluster in the knowledge graph; determine the most similar historical similar fingerprint cluster as the target cluster based on the minimum Euclidean distance principle; S8.3: Obtain a set of related compensation strategy fine-tuning rules from the knowledge graph based on the target historical similar fingerprint clusters; perform confidence weighting on the set of compensation strategy fine-tuning rules based on the distribution density of data points within the historical similar fingerprint clusters; perform rule fusion operation to generate a unified compensation strategy fine-tuning instruction; calculate the confidence index of the compensation strategy fine-tuning instruction; and output the compensation strategy fine-tuning rules with confidence labels. S8.4: Perform vector superposition processing on the compensation strategy fine-tuning rules with confidence labels and the basic compensation instructions of the current stage; perform physical constraint verification on the superposition result to ensure that the physical limit requirements of the press actuator are met; optimize the compensation instruction parameters based on the physical constraint verification results to minimize the predicted value of the layer deviation residual; generate an enhanced compensation instruction that includes the pressure gradient slope correction, temperature plateau offset, and pressure holding time increment; attach a confidence label to the enhanced compensation instruction to indicate its reliability level. S8.5: Adapt the enhanced compensation command execution mechanism and map it to the control word table of the programmable logic controller of the pressing equipment; output compensation operation commands based on the control word table, including pressure gradient slope correction, temperature plateau offset, and holding time increment; monitor the layer deviation residual change trend during the compensation operation execution process; calculate the layer deviation residual convergence rate index after the compensation operation; generate a phase closed-loop verification report to evaluate the effectiveness of the enhanced compensation command.

[0026] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time prediction and compensation control of lamination layer deviation in multilayer boards, characterized in that, Includes the following steps: S1: Deploy a multi-point distributed sensor array in the high-multilayer board lamination equipment to synchronously collect local temperature gradient, pressure transmission delay time, medium flow rate and real-time layer offset optical measurement values ​​at each lamination stage, forming a spatiotemporal sampling sequence. S2: Perform causal sensitivity analysis on the spatiotemporal sampling sequence, remove redundant parameters whose correlation with the layer bias response is lower than a preset threshold, and retain the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient as core coupling features. S3: Based on the core coupling features and the current pressing stage number, construct a stage coupling state fingerprint vector; S4: Construct a graph neural network model, define a loss function to minimize the weighted deviation between the predicted layer bias and the measured layer bias in the next stage, generate a minimization objective function, and constrain the compensation instruction to generate a compensation instruction output space. Based on the minimization objective function and the compensation instruction output space, update the graph neural network model parameters to obtain the optimized graph neural network model. S5: After each pressing stage is completed, a stage coupling state fingerprint is generated based on the data collected in the current stage. The stage coupling state fingerprint is then input into the optimized graph neural network model to obtain a set of compensation instructions that adapts to the physical constraints and disturbance characteristics of the current stage. S6: Map the compensation instruction set to the control word table of the programmable logic controller of the pressing equipment, and execute the compensation operation.

2. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 1, characterized in that, Following step S6, the following is also included: S7: Within a preset time window before the end of each pressing stage, call the embedded optical measurement module to perform sub-pixel positioning of the board edge markers, calculate the actual layer offset residual, and determine whether the absolute value of the actual layer offset residual exceeds the preset residual threshold or whether the direction changes abruptly. S8: If the layer offset residual is determined to meet the triggering condition, the stage coupling state fingerprint combination of the previous two pressing stages is traced back, historical similar fingerprint clusters are retrieved, the corresponding compensation strategy fine-tuning rules are retrieved from the knowledge graph, and an enhanced compensation instruction with confidence label is generated.

3. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 1, characterized in that, The multi-point distributed sensor array deployment includes sensor location planning. Based on the size parameters of the press plate and the distribution of the thermal expansion coefficient, the sensor layout is optimized to the area with the maximum temperature gradient change rate and stress concentration area through spatial discretization analysis, finite element mesh generation, thermo-mechanical coupling field calculation and characteristic sensitivity screening.

4. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 1, characterized in that, The causal sensitivity analysis adopts the Granger causality test algorithm. The sampling interval of the input temperature, pressure and flow rate sequences is preferably less than or equal to 1 ms, the test lag order is 3-10, the significance level is 0.01-0.1, and the feature screening threshold correlation is not less than 0.

3.

5. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 1, characterized in that, Step S3 specifically includes: The core coupling features output in step S2 are read to obtain the temperature-pressure coupling phase difference, interlayer slip accumulation, and resin flow resistance coefficient, and the current pressing stage number is obtained simultaneously. Based on the core coupling characteristics, a zero-mean normalization algorithm is applied to the temperature-pressure coupling phase difference, interlayer slip accumulation, and resin flow resistance coefficient to generate a normalized feature vector. The standardized feature vector is concatenated with the current pressing stage number to construct a four-dimensional vector of the initial stage coupling state fingerprint. Perform a range boundary verification operation on the initial stage coupled state fingerprint four-dimensional vector and output the fingerprint vector that passes the verification.

6. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 5, characterized in that, The stage coupling state fingerprint four-dimensional vector includes the standardized temperature-pressure coupling phase difference, interlayer slip accumulation, resin flow resistance coefficient, and pressing stage number. Each feature has been verified by physical boundary checks and consistency verification.

7. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 1, characterized in that, Step S4 specifically includes: Based on the stage coupling state fingerprint vector output in step S3, node initialization processing is performed on the pressing stage sequence, mapping each pressing stage to an independent node in the graph neural network, generating a graph node set. Based on the stage coupling state fingerprint vector, Euclidean distance is calculated for the stage coupling state fingerprints of adjacent pressing stages to obtain the feature difference measure between stages. At the same time, the layer offset residual change rate is obtained. A linear weighted fusion operation is performed on the feature difference measure and the layer offset residual change rate to generate dynamic edge weight parameters. Based on the graph node set and the dynamic edge weight parameters, a loss function is defined for calculation and processing. The loss function performs a weighted absolute deviation calculation on the predicted layer bias and the measured layer bias in the next stage to generate a minimized objective function. Physical limiting constraints are applied to the compensation instruction set, which is then mapped to the operating range boundary of the press actuator. Simultaneously, step response characteristic constraints are performed to generate a constraint-compliant compensation instruction output space. The graph neural network model is trained using a training dataset with stage labels through backpropagation. Based on the output space of the minimized objective function and the constraint-compliant compensation instructions, the parameters of the graph neural network model are updated until the loss function converges, thereby obtaining the optimized parameter set of the graph neural network model.

8. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 7, characterized in that, The graph neural network model adopts four-dimensional features in the input layer, 16-64 nodes in the hidden layer, ReLU activation function, and the edge weights are linearly weighted and fused with the Euclidean distance of the stage coupling state fingerprint and the rate of change of the layer partial residual, and normalized to the [0,1] interval. The training adopts the Adam optimizer with a learning rate of 0.001-0.005 and the number of iterations is not less than 100.

9. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 1, characterized in that, Step S5 specifically includes: The core coupling feature data output by the real-time acquisition system of the pressing process is called and processed. Based on the temperature-pressure coupling phase difference, interlayer slip accumulation and resin flow resistance coefficient as input conditions, the data interface reading operation is executed to obtain the core coupling feature dataset after the current stage ends. Based on the core coupling feature dataset and the current pressing stage number as input conditions, a four-dimensional feature vector construction process is performed through a feature vector combination algorithm to generate a stage coupling state fingerprint. Using the stage coupling state fingerprint as input condition, data format conversion processing is performed through the model interface protocol to map the four-dimensional feature vector to the node input layer of the optimized graph neural network model; Using the inference engine based on the graph neural network model as input, forward propagation computation is performed. Compensation instruction sets are generated through node feature aggregation and edge weight optimization, and structured control parameters that adapt to the physical constraints and disturbance characteristics of the current stage are output.

10. The method for real-time prediction and compensation control of lamination layer deviation in multilayer boards according to claim 9, characterized in that, The compensation instruction set includes pressure gradient slope correction, temperature plateau offset, and pressure holding time increment.

Images