Twelve-layer automobile PCB fusion area glue shortage defect prediction and process compensation method

Abstract

The invention provides a twelve-layer automobile PCB fusion area glue shortage defect prediction and process compensation method, which comprises the steps of multi-source process parameter and defect detection data acquisition, normalization exception processing, domain knowledge constrained causal relationship dynamic modeling, and realization of real-time adaptive updating of a causal network in combination with a sliding window mechanism. According to the method, anti-fact intervention reasoning is carried out based on a structural equation model, a process compensation optimization strategy is generated, and closed-loop adjustment of process parameters is realized through an MES system, so that the accuracy of glue shortage defect prediction and the pertinence of compensation intervention are effectively improved, rapid optimization of process abnormal response is realized, and the stability of PCB manufacturing quality and the defect suppression capability are enhanced.

Description

Technical Field

[0001] This invention relates to the field of automotive electronics manufacturing process control and defect prediction technology, and in particular to a method for predicting and compensating for missing adhesive defects in the fusion zone of a twelve-layer automotive PCB. Background Technology

[0002] Currently, the automotive electronics manufacturing industry places stringent demands on the quality control of multilayer printed circuit boards (PCBs) requiring high reliability and consistency. For the manufacturing process of twelve-layer automotive PCBs, insufficient adhesive in the fusion zone is one of the core challenges affecting the quality and electronic performance stability of the finished product. Existing manufacturers primarily rely on historical data-driven correlation modeling methods to control this type of defect, such as defect prediction models based on machine learning or statistical analysis. These models focus on mining the correlation between multi-source process parameters (such as lamination pressure, temperature gradient, resin rheological properties, interlayer alignment error, vacuum fluctuations, etc.) and defect detection results. For example, by constructing prediction frameworks such as regression models, decision trees, and neural networks, early warnings of the probability of insufficient adhesive in the fusion zone and rough determination of defect locations can be achieved, and suggested process adjustment strategies can be output based on the model. However, current technology still has the following prominent problems and limitations: First, most existing defect prediction models are based on statistical correlation, which makes it difficult to reveal the fundamental causal mechanism of missing glue in the fusion zone. They are also weak in determining the main cause and analyzing the path in complex multivariate interaction processes. Therefore, compensation and process optimization are prone to the phenomenon of "treating the head when the head hurts and the foot when the foot hurts", making it difficult to accurately locate the real cause of the defect. Second, in the compensation strategy generation stage, traditional methods often use parameter empirical shifting or simple sensitivity ranking based on predicted output, failing to identify key controllable parameters and their optimal intervention paths from the causal mechanism level. This results in adjustment suggestions lacking pertinence and process interpretability, with large fluctuations in actual process adjustment effectiveness, and the compensation closed loop failing to form positive reinforcement. Third, in a dynamic manufacturing environment, the causal relationship between process parameters is dynamically evolving due to factors such as equipment aging, material batch differences, and environmental fluctuations. However, existing models have a fixed structure and lack the ability to adapt and update in real time with the manufacturing status. Long-term use of these models can easily lead to prediction drift and intervention failure. Fourth, the industry's current technical approaches have not widely adopted causal reasoning algorithms and counterfactual analysis mechanisms. Process intervention decisions are mostly limited to data correlation interpretation, and cannot quantitatively predict the changes in defect probability caused by "adjusting a certain parameter", nor can they dynamically verify and correct the intervention effect, thus limiting their practical value and theoretical guidance significance. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a method for predicting and compensating for missing adhesive in the fusion zone of a twelve-layer automotive PCB.

[0004] The technical solution of this invention is implemented as follows: a method for predicting and compensating for insufficient adhesive in the fusion zone of a twelve-layer automotive PCB board, comprising: S1: Collect multi-source process parameters and corresponding glue deficiency detection results in the fusion zone during the manufacturing process of a twelve-layer automotive PCB board. The process parameters include pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation parameters, and construct a pairwise observation dataset. S2: Normalize and remove outliers from the observed dataset, and construct the input matrix for the time-series causal discovery algorithm based on domain knowledge constraints. The processing includes dynamic window truncation of the time delay correlation between process parameters. S3: Based on the PCMCI+ algorithm and combined with process mechanism constraints, an initial process-defect dynamic cause-effect graph model is constructed, and the output is a cause-effect network topology structure containing node process parameters, directed causal edges and edge weights. S4: The dynamic causal graph model is updated online using a sliding time window mechanism. By calculating the KL divergence difference between the new window data and the historical causal edges, the weights of the causal edges are dynamically adjusted and the network topology is reconstructed. S5: Construct a counterfactual reasoning module based on structural equation model. When the predicted probability of insufficient glue in the fusion zone exceeds the preset threshold, perform do-intervention operation on the parameters of each node in the causal graph to simulate the marginal causal effect of different process parameter adjustments on the defect probability. S6: Calculate the compensation priority index based on the marginal causal effect, generate a multi-level compensation strategy by taking into account the process adjustment cost constraints, and output compensation suggestions that include the adjustment amount of key process parameters and the expected reduction in defect probability. S7: Feedback the compensation suggestion to the MES system to adjust the process parameters, and collect the process parameters and defect detection results of the compensated batch to update the node association strength of the dynamic cause-effect graph model; S8: Based on the change in defect probability after compensation, calculate the prediction error of the causal graph model. When the error exceeds the preset tolerance value, trigger the causal network reconstruction mechanism to form a closed-loop optimization link of prediction-intervention-verification.

[0005] The method for predicting and compensating for insufficient adhesive in the fusion zone of a twelve-layer automotive PCB board provided by this invention has the following beneficial effects: (1) This invention significantly improves the scientificity and interpretability of process optimization by integrating dynamic causal discovery and interventionist reasoning mechanisms. It introduces a time-series causal discovery algorithm based on PCMCI+ and applies structural constraints in conjunction with domain knowledge to effectively identify the direct causal relationships and interlayer transmission paths between core variables such as pressing pressure, temperature gradient, and resin rheological properties, overcoming pseudo-correlation interference. Furthermore, a sliding time window is used to continuously update the causal graph structure and edge weights, enabling the model to adapt to dynamic changes caused by production line operating conditions drift and equipment aging, realizing the online evolution of the causal network. This design transforms defect cause analysis from "black box correlation" to "white box tracing," providing a reliable causal basis for subsequent precise intervention and greatly enhancing the usability and credibility of the prediction results in actual engineering decision-making. (2) This invention innovatively constructs an interventionizable decision-making module based on counterfactual reasoning, realizing a leap from "passive early warning" to "active optimization." When a high risk of glue deficiency is predicted in a certain batch, the counterfactual reasoning mechanism is automatically activated to simulate the change in defect probability caused by "adjusting a specific process parameter." The marginal causal effect of each intervention measure is quantified by using structural equation modeling or neurocausal modeling. This process not only outputs the probability of defect occurrence but also provides a ranking of the expected improvement degree of different control schemes, thereby supporting the generation of differentiated and refined compensation suggestions. On this basis, the compensation strategy prioritizes key parameters with strong causal influence and low implementation cost for recommendation, taking into account both process feasibility and economy. At the same time, the system uses the actual compensated quality feedback data to reversely correct the connection strength in the causal graph, forming a self-evolving closed loop of "prediction-intervention-verification-learning," continuously improving the model's generalization ability and robustness in future scenarios. Compared with traditional fixed rule or experience-driven parameter tuning methods, this invention significantly improves the efficiency and response sensitivity of compensation actions, effectively overcoming the quality fluctuation problem caused by ineffective intervention or over-adjustment. (3) This invention constructs an intelligent process optimization system with adaptive, interpretable, and strong closed-loop characteristics, fully meeting the stringent requirements of automotive-grade PCB manufacturing for high reliability and high consistency. By deeply integrating dynamic causal modeling with the Manufacturing Execution System (MES), end-to-end linkage from data perception to decision execution is achieved, breaking down information silos and control delays, and making process adjustments more timely and accurate. Compared with traditional offline modeling and static model deployment modes, this method significantly improves real-time response capabilities and long-term stability while ensuring computational efficiency, and can cope with the challenges of mixed-line production of multiple batches and models of products without frequent manual parameter tuning. In addition, the causal graph itself has a natural visualization advantage, which makes it easier for engineers to understand the defect propagation mechanism and participate in strategy review, enhancing the trust foundation of human-machine collaboration. This architecture is not only applicable to the quality control of the fusion zone of twelve-layer automotive PCBs, but can also be extended to micro-defect prediction and process compensation scenarios of other high-density interconnect structures, possessing good versatility and engineering implementation value. Attached Figure Description

[0006] Figure 1 This is a flowchart of the method for predicting and compensating for insufficient adhesive in the fusion zone of a twelve-layer automotive PCB board according to the present invention. Figure 2 This is a sub-flowchart of the method for predicting and compensating for insufficient adhesive in the fusion zone of a twelve-layer automotive PCB board according to the present invention. Figure 3 This is another sub-flowchart of the present invention's method for predicting and compensating for insufficient adhesive in the fusion zone of a twelve-layer automotive PCB board. 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 predicting and compensating for insufficient adhesive in the fusion zone of a twelve-layer automotive PCB board, specifically including: S1: Collect multi-source process parameters and corresponding glue deficiency detection results in the fusion zone during the manufacturing process of a twelve-layer automotive PCB board. The process parameters include pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation parameters, and construct a pairwise observation dataset. S2: Normalize and remove outliers from the observed dataset, and construct the input matrix for the time-series causal discovery algorithm based on domain knowledge constraints. The processing includes dynamic window truncation of the time delay correlation between process parameters. S3: Based on the PCMCI+ algorithm and combined with process mechanism constraints, an initial process-defect dynamic cause-effect graph model is constructed, and the output is a cause-effect network topology structure containing node process parameters, directed causal edges and edge weights. S4: The dynamic causal graph model is updated online using a sliding time window mechanism. By calculating the KL divergence difference between the new window data and the historical causal edges, the weights of the causal edges are dynamically adjusted and the network topology is reconstructed. S5: Construct a counterfactual reasoning module based on structural equation model. When the predicted probability of insufficient glue in the fusion zone exceeds the preset threshold, perform do-intervention operation on the parameters of each node in the causal graph to simulate the marginal causal effect of different process parameter adjustments on the defect probability. S6: Calculate the compensation priority index based on the marginal causal effect, generate a multi-level compensation strategy by taking into account the process adjustment cost constraints, and output compensation suggestions that include the adjustment amount of key process parameters and the expected reduction in defect probability. S7: Feedback the compensation suggestion to the MES system to adjust the process parameters, and collect the process parameters and defect detection results of the compensated batch to update the node association strength of the dynamic cause-effect graph model; S8: Based on the change in defect probability after compensation, calculate the prediction error of the causal graph model. When the error exceeds the preset tolerance value, trigger the causal network reconstruction mechanism to form a closed-loop optimization link of prediction-intervention-verification.

[0010] Step S1: Collect multi-source process parameters and corresponding glue deficiency detection results in the fusion zone during the manufacturing process of the twelve-layer automotive PCB board. The process parameters include pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation, and vacuum fluctuation parameters, constructing a pairwise observation dataset. Specifically, this includes: S1.1: Acquire pressing pressure time-series data based on the sensor network of the pressing equipment. The pressing pressure time-series data includes pressure peak, holding time, and pressure fluctuation variance, so as to obtain the mechanical action characteristic parameters in the pressing process. The input object is the pressing pressure time-series data stream output by the sensor network of the pressing equipment, which includes the pressure peak, holding time and pressure fluctuation raw signals from multiple key detection nodes of the equipment; A high-precision strain gauge pressure sensor array (sampling frequency ≥ 1 kHz, resolution ≤ 0.01 MPa) is used to collect the original pressure signals at different locations during the pressing process, and generate a basic time-series data matrix of pressure changes over time. Furthermore, the original pressure signal was noise-suppressed by digital filtering (parameters: fourth-order Butterworth low-pass filter, cutoff frequency 50Hz), eliminating high-frequency components introduced by electromagnetic interference and mechanical vibration, and obtaining the purified pressure time-series curve. Furthermore, by using a peak detection algorithm (parameter: dynamic threshold = mean + 2σ), the maximum pressure during the pressing process is identified on the pressure time-series curve after purification, and the corresponding time index is recorded as the pressure peak parameter; Furthermore, by using an interval integration algorithm, based on the pressure holding phase time range set by the pressure control program, the average pressure and duration within the interval are calculated to obtain pressure holding time index and stable pressure level characteristic data. Furthermore, the intensity of pressure fluctuation throughout the entire pressing cycle is evaluated using the variance calculation formula, and the pressure fluctuation variance is output as a quantitative indicator reflecting the stability of the pressing process.

[0011] in, Let be the pressure value at time i. This represents the average pressure over the entire cycle. The total number of sampling points is used to obtain the pressure fluctuation variance, a characteristic parameter of mechanical action, through this formula. Through the above signal acquisition, noise reduction, peak value calculation, interval integration and variance evaluation processing methods, the original pressure signal from the previous step is transformed into multi-dimensional mechanical action characteristic data including pressure peak value, holding time and pressure fluctuation variance, so as to achieve accurate quantification of the mechanical state of the pressing process. For example, during the lamination process of a twelve-layer automotive PCB board, the lamination equipment is equipped with eight distributed strain gauge pressure sensors with a sampling frequency of 2000Hz, a sensor range of 0-10MPa, and an accuracy of 0.005MPa. The original signal is denoised using a fourth-order Butterworth low-pass filter (cutoff frequency 30Hz) to obtain a purified pressure time-series curve. Dynamic threshold peak detection (threshold = mean + 1.8σ) identifies the pressure peak during the lamination stage as 8.25MPa, with a peak occurrence time of 35.6 seconds. Based on the pressure holding settings of the lamination control program (35 to 85 seconds), the holding time is calculated to be 50.0 seconds using interval integration, and the average holding pressure is 8.10MPa. The full-cycle pressure fluctuation variance is calculated using the pressure fluctuation variance formula to be 0.012MPa², which is significantly lower than the process stability warning line (0.05MPa²), indicating that the mechanical state of this batch of lamination is stable. The extracted mechanical feature data can provide highly reliable input for subsequent causal inference algorithms. S1.2: Temperature gradient data is collected based on a multi-point infrared temperature measurement module. The temperature gradient data includes heating rate, temperature difference in constant temperature section, and temperature difference fluctuation between layers, in order to extract the characteristics of uneven heat conduction during hot pressing. S1.3: Acquire resin rheological property data based on an online rheometer. The resin rheological property data includes viscosity change curves, gel time, and flow window width to reflect the evolution of the physical state of the binder material during the fusion process. S1.4: Based on a high-precision optical positioning system, interlayer alignment deviation data is collected. The interlayer alignment deviation data includes X / Y direction offset, angular deflection, and cumulative alignment error to characterize the stacking accuracy of multilayer structures. S1.5: Collect vacuum fluctuation data based on vacuum sensor, including vacuum pumping rate, vacuum extreme value maintenance, and periodic fluctuation amplitude, to quantify the stability of the degassing process; A multi-point acquisition method based on a vacuum sensor array (sampling frequency: 10Hz, sensor range: 0~100kPa) is used to acquire the original time-series signal of vacuum degree during hot-pressing degassing process; Furthermore, noise suppression and trend extraction are performed on the original vacuum degree time series signal through first-order difference operation and moving average filtering (window length: 5), and the vacuum pumping rate parameter is obtained. The calculation formula is:

[0012] in For vacuum degree, and Sampling time; Furthermore, the extreme points of the vacuum maintenance stage are extracted by the peak search algorithm (threshold: 0.1 kPa), and the extreme value index of vacuum maintenance is generated to characterize the stability of the vacuum state. Furthermore, based on Fast Fourier Transform (FFT, sampling points: 1024), spectral analysis was performed on the vacuum degree time series data to extract the dominant frequency component and its amplitude of periodic fluctuations, and the periodic fluctuation amplitude parameter was calculated. The calculation formula is:

[0013] in It is a frequency domain vacuum degree signal; Through the above feature extraction and parameter calculation processing methods, the results of each time series operation are transformed into quantitative indicators such as vacuum pumping rate, vacuum degree extreme value and periodic fluctuation amplitude, so as to effectively quantify the stability of the degassing process. For example, in a lamination test of a 12-layer automotive-grade PCB, the vacuum sensor sampling frequency was set to 10Hz, with a range of 0~100kPa. The original vacuum data collected during the degassing process ranged from 0~85kPa. A moving average filter (window length 5) was applied to the original signal to obtain a smooth curve. In the initial stage of degassing, from second 0 to second 30, the vacuum level decreased from 85kPa to 10kPa. According to the formula... The calculated vacuum pumping rate was 2.5 kPa / s. During the constant vacuum phase, a peak search with a threshold of 0.1 kPa revealed a maximum vacuum level of 9.8 kPa with a fluctuation range of 0.05 kPa. FFT analysis showed a dominant frequency fluctuation period of 60 seconds and a calculated periodic fluctuation amplitude of 0.025 kPa. These degassing process characteristics indicate a significantly improved vacuum environment stability, providing highly reliable process input data for subsequent causal models. S1.6: Obtain the detection results of missing glue in the fusion zone based on the AOI detection system. The detection results of missing glue in the fusion zone include the proportion of missing glue area, the coordinates of the defect location, and the defect morphology classification label, so as to form a defect observation sample that corresponds one-to-one with the process parameters. The input conditions for acquiring the detection results of missing glue in the fusion zone of a twelve-layer automotive PCB board based on the AOI inspection system include: the lamination process completion command signal corresponding to the production batch number, the color image sequence output by the high-resolution industrial camera of the inspection station, the coordinate range of the inspection area provided by the control system, and the parameter settings of the light source system. A multi-angle, multi-spectral imaging method (parameters: simultaneous acquisition of RGB and near-infrared channels, resolution ≥5μm) is used to achieve full-coverage, high-precision image acquisition of the fusion zone surface; Furthermore, by using an adaptive illumination equalization algorithm (parameters: local histogram equalization, dynamic adjustment of light intensity compensation coefficient), the illumination conditions of the detected images are normalized, and a set of original defect detection images with enhanced contrast is obtained. Furthermore, a feature segmentation algorithm based on a convolutional neural network (parameters: U-Net structure, loss function optimized by Dice coefficients) is used to achieve accurate segmentation of the glue-deficient region in the image and generate a binary mask matrix of the glue-deficient region. Furthermore, a connected component analysis method (parameters: minimum pixel threshold set to 50, combined with morphological opening and closing operations) is employed to identify independent adhesive-deficient regions in the binary mask and calculate the area value of each adhesive-deficient region. ,in This represents the area of ​​the glue-deficient region. The total area of ​​the detection area is used to determine the percentage of area lacking adhesive. Furthermore, by using centroid calculation and pixel coordinate mapping methods (parameter: double affine transformation matrix from workstation coordinate system to PCB design coordinate system), the precise coordinates of the defect location are extracted, and the defect location coordinate set is output. Furthermore, a classification algorithm based on contour morphology feature extraction (parameters: Hu invariant rectangular matrix features + support vector machine classifier) ​​is used to classify the morphology of glue-deficient defects and generate defect morphology classification labels, including interface fracture type, resin loss type and porous void type. The above algorithm transforms the detected image data from the previous step into structured detection results, including the percentage of glue-deficient area, defect location coordinates, and defect morphology classification labels. This enables the construction of defect observation samples that correspond one-to-one with process parameters, providing high-confidence defect label inputs for subsequent causal analysis and model training. For example, in a production batch of automotive-grade twelve-layer PCB boards, the industrial camera sampling resolution of the AOI inspection system was set to 4096×2160 pixels, using RGB+850nm near-infrared dual-channel acquisition, with the light source color temperature controlled at 5500K and an exposure time of 8ms. The image underwent local histogram equalization, with the light intensity compensation coefficient dynamically adjusted between 0.95 and 1.05 to obtain a contrast-enhanced inspection image. A U-Net segmentation network (learning rate set to 0.001, batch size of 16) was used to segment the image into glue-deficient regions, obtaining a binary mask matrix for these regions. With a minimum pixel threshold of 50 for connected component analysis, seven independent glue-deficient regions were identified, ranging in area from 0.15mm² to 1.02mm², and their total area accounted for 0.025 to 0.148 of the total inspection area. The centroid coordinates were mapped to the PCB design coordinate system through a double affine transformation, achieving a positioning accuracy better than 10μm. The defect coordinate range was concentrated in the center-right position of the fusion zone. The Hu moment classifier was used to identify four defects as resin loss type and three as porous and void type. The final defect detection results were then paired with process parameters such as pressing pressure and temperature gradient in subsequent causal analysis to achieve high-precision training of the defect prediction model, significantly improving the predictive effectiveness of the model and the targeting of intervention strategies in this batch of products. S1.7: The pressing pressure time series data, temperature gradient data, resin rheological property data, interlayer alignment deviation data, vacuum fluctuation data and the detection results of missing glue in the fusion zone are timestamped and stored in a structured manner to construct a pairwise observation dataset containing process parameters and defect labels.

[0014] Step S2: Normalize and remove outliers from the observed dataset, and construct the input matrix for a time-series causal discovery algorithm based on domain knowledge constraints. This processing includes dynamic windowing of the time delay correlation between process parameters. Specifically, it includes: S2.1: The bonding pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation parameters collected during the manufacturing process of the twelve-layer automotive PCB are subjected to minimum-maximum normalization processing to eliminate the interference of different dimensions on the time-series causal discovery algorithm and obtain a standardized process parameter matrix. S2.2: Based on the sliding window Z-score method, outlier detection and removal operations are performed on the data of each dimension in the standardized process parameter matrix. 3σ is set as the outlier threshold to remove abnormal data points caused by sensor drift or equipment failure, and a cleaned time series dataset is obtained. S2.3: Based on the process mechanism knowledge provided by experts in the automotive electronics manufacturing field, a priori causal constraint graph between process parameters is constructed, and the existence of direct causal relationship, time delay dependency relationship or prohibited connection relationship between each parameter is marked as the input constraint condition of the time-series causal discovery algorithm. S2.4: Perform time delay correlation analysis based on mutual information and partial autocorrelation coefficient on the cleaned time series dataset to identify the optimal time delay window between key process parameters and obtain the set of time delay offsets between each parameter for subsequent dynamic window truncation; S2.5: Based on the set of time delay offsets, the cleaned standardized time series data is reconstructed using a dynamic time window truncation method to generate a multi-dimensional time series input matrix with causal inference capability, where each dimension corresponds to the observation value of a parameter under different time lags, in order to support the modeling requirements of the subsequent PCMCI+ causal discovery algorithm. Based on the time delay offset set, a dynamic time window truncation method is used to reconstruct the cleaned and standardized time series data, aiming to generate a multi-dimensional time series input matrix that meets the requirements of PCMCI+ causal modeling. A dynamic time window partitioning algorithm (parameters: time delay offset set Δτ, window length Lw, step size Sw) is adopted to realize the function of multi-time lag observation and extraction of standardized time series data after cleaning in various process parameter dimensions. Furthermore, by using the sliding window resampling method (parameters: window start point t0, time delay set Δτ), synchronous data segments of each process parameter under different time lags are extracted, and a multi-dimensional time slice set is obtained; Furthermore, a sequence stacking integration algorithm (parameters: time slice set, parameter dimension N) is adopted to stack different lag observations of the same process parameter in the column direction to form a parameter lag sequence matrix; Furthermore, by using a matrix dimension rearrangement method (parameters: lag sequence matrix, merging rule R), the lag sequence matrices of all process parameters are merged into a multi-dimensional time series input matrix with a unified structure, where each dimension corresponds to the observation data of a process parameter under a specific lag. Through the matrix reconstruction process described above, the time delay offset analysis result of the previous step is transformed into time series input data with causal reasoning ability, so as to realize the effective modeling of the subsequent PCMCI+causal discovery algorithm in dynamic time delay scenarios. For example, in the fusion process of a twelve-layer automotive PCB manufacturing process, the time delay offset set Δτ is set to {0, 1, 3, 5} seconds, the window length Lw is set to 10 seconds, the step size Sw is set to 2 seconds, and the process parameter dimension N=5, corresponding to five types of parameters: bonding pressure, temperature gradient, resin rheological properties, interlayer alignment deviation, and vacuum fluctuation. Based on this configuration, the dynamic time window partitioning algorithm sequentially performs window truncation on the cleaned and standardized time series data to obtain a set of time slices for each type of process parameter under different lag conditions. Taking the bonding pressure parameter as an example, the data segment length corresponding to its 3rd second lag window is 10 seconds, and it moves within the entire sampling period with a step size of 2 seconds. The generated multiple segments are stacked column-wise to form a lag sequence matrix. Subsequently, the lag sequence matrices of the five types of parameters are rearranged in dimensions, and the observation columns of different parameters with the same lag value are arranged according to a unified index position, resulting in the final input matrix structure as (number of samples × lag dimension × parameter dimension). Assuming a sample size of 100, the final matrix size is 100×4×5, where 4 represents the lag number and 5 represents the number of parameters. After inputting this matrix into the PCMCI+ causal discovery algorithm, the causal direction and strength of different parameter lag values ​​can be identified within a dynamic window, significantly improving the accuracy of causal inference and the targeting of process interventions.

[0015] like Figure 2 As shown, step S3 involves constructing an initial process-defect dynamic causal graph model based on the PCMCI+ algorithm and process mechanism constraints, outputting a causal network topology structure that includes node process parameters, directed causal edges, and edge weights. Specifically, this includes: S3.1: Based on the multi-source process parameter dataset after normalization and outlier removal in S2, it is input into the PCMCI+ causal discovery algorithm framework to identify the temporal causal relationship of variables such as pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation, so as to construct a preliminary causal network structure. Based on the multi-source process parameter cleaning time series dataset output by step S2, after minimum-maximum normalization and sliding window Z-score outlier removal, five types of variables, namely pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation, are selected as input objects for the PCMCI+causal discovery algorithm. The PCMCI+ algorithm (parameters: maximum delay L is determined by the delay offset set output by S2.4, the conditional independence test method is the partial correlation test based on Fisher Z transform, and the significance level α is taken as the empirical value of the process) is adopted to realize the search for potential causal paths and the identification of delay dependency structures among multidimensional time series variables. Furthermore, by configuring the prior causal constraint matrix constructed by S2.3 in the PCMCI+ framework, the search space of the algorithm is constrained, and a set of candidate causal edges after removing prohibited connections is obtained; Furthermore, for each candidate causal edge, the conditional independence statistic and corresponding p-value distribution within a given maximum time delay range are calculated, the set of causal relationships that pass the significance test is screened, and the directed edge orientation results are generated. Furthermore, the aforementioned significant causal edges and their corresponding time delay information are integrated into a preliminary causal network topology, and a set of nodes is generated, where each node corresponds to a single process parameter variable, the edges represent the causal direction, and the edges are accompanied by time delay labels. By using the causal structure output of the PCMCI+ algorithm, the clean multi-source process parameter data from the previous step are transformed into a quantifiable preliminary causal network model, realizing a structured expression of the temporal causal relationship between process variables, and providing a basic structure for the application of the physical mechanism correction rules in S3.2. For example, for a batch of lamination dataset of a twelve-layer automotive PCB board, the lamination pressure variable uses a pressure peak sequence with a sample frequency of 1Hz, the temperature gradient variable uses an interlayer temperature difference fluctuation sequence, the resin rheological properties use discrete values ​​of viscosity change curves collected by an online rheometer, the interlayer alignment deviation uses an X / Y offset sequence output by high-precision optical positioning, and the vacuum fluctuation uses a periodic fluctuation amplitude sequence collected by a vacuum pressure sensor. These variables are used to construct a five-dimensional time series input matrix. The maximum time delay L is set to... seconds, significance level α is In the PCMCI+ algorithm, the partial correlation test is called to calculate the partial correlation between 0 and 1. Conditional independence statistics under a second delay are used, and physically impossible causal connections, such as paths where vacuum fluctuations directly lead to temperature gradients, are eliminated using a priori constraint matrix. For the remaining candidate edges, the p-value curve for each edge is calculated, and edges with a p-value lower than 1 / 2 are selected. The significant range of the output, such as pressing pressure → probability of missing glue (delay) (seconds), temperature gradient → resin rheological properties (delay) The causal relationships (such as seconds) are used to form a preliminary causal network topology with 5 nodes and 12 causal edges, providing directional input for subsequent process mechanism correction; S3.2: Introduce the physical mechanism knowledge of automotive PCB lamination process as a priori constraint for the construction of cause-effect graph, make reasonable corrections to the preliminary cause-effect structure output by PCMCI+ algorithm, eliminate false cause-effect paths that do not conform to process logic, so as to improve the domain applicability and engineering credibility of cause-effect graph model. The input condition is the preliminary causal network structure output from step S3.1. This structure is automatically generated by the PCMCI+ algorithm based on a multi-source process parameter dataset, and includes process variable nodes such as lamination pressure, temperature gradient, resin rheological properties, interlayer alignment deviation, and vacuum fluctuation, along with their preliminary directed connections. It is necessary to introduce the physical mechanism knowledge of automotive PCB lamination process as a priori constraint for this structure, and systematically eliminate and correct false causal paths that do not conform to the process logic. A process mechanism constraint mapping method based on expert knowledge (parameters: process link classification, material physical properties, and equipment operating state variable set) is adopted to determine the process rationality of each directed edge in the preliminary causal network. Furthermore, by using the process-variable dependency matrix calculation method (parameters: node variable physical correlation matrix, causal edge statistical significance), the consistency of causal interpretation for each edge is evaluated, and the process logic matching index is obtained. Furthermore, a false causal detection algorithm (parameters: conditional independence test threshold, process prohibition connection table) is adopted to mark statistically significant but process-prohibited paths in the causal network and generate a list of false causal paths; Furthermore, by using a causal path correction algorithm (parameters: edge weight adjustment coefficient α, constraint satisfaction β), causal edges that conform to the domain logic are retained, and false edges are subjected to weight zeroing or structure deletion operations, resulting in a corrected causal network topology. By integrating process mechanism constraints and correcting causal structure, the output of the PCMCI+ algorithm in the previous step is transformed into a dynamic causal network structure that conforms to the real transmission relationship of automotive PCB lamination process, thereby improving the domain applicability and engineering credibility of the causal model. For example, in the lamination process of a twelve-layer automotive PCB, the initial causal network structure contains directed edges pointing from the temperature gradient to the resin rheological properties. These edges have high statistical significance but do not reflect the mediating role of lamination pressure. Based on prior process constraint mapping, it is assumed that there must be an intervening coupling term between lamination pressure and the temperature gradient, meaning changes in resin rheological properties must be transmitted through lamination pressure. A process dependency matrix is ​​constructed, where the physical correlation between the temperature gradient and resin rheological properties is 0.4, lower than the set effective transmission threshold of 0.6, triggering the false path detection algorithm. Using a conditional independence test, under controlled lamination pressure, the correlation coefficient between the temperature gradient and resin rheological properties drops to 0.05, indicating that this edge is a false causal relationship. The causal path correction formula is then executed.

[0016] in, The original edge weight is 0.35. The adjustment factor is 0. The constant term is 0, which corrects the subsequent weights. This achieves the complete removal of the false path. The corrected causal network exhibits higher consistency in subsequent dynamic window updates and counterfactual reasoning, significantly improving the explanatory power of defect prediction and the pertinence of compensation strategies; S3.3: Based on the modified causal structure, Granger causality test and partial correlation analysis are used to identify the direct causal effects and indirect transmission paths between various process parameters, and to calculate the initial weight of each causal edge in order to quantify the causal influence of different process variables on the occurrence of glue shortage defects in the fusion zone. Based on the preliminary causal structure modified by S3.2, the input objects are the standardized and cleaned time sequence process parameter matrix, such as pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation, as well as the causal connection constraint information between nodes constructed in the early stage. The Granger causality test method (parameters: significance level α=0.05, maximum time delay order p is set according to the time delay offset set output by S2.4) is used to determine the significance of the direct causal effect of each process parameter node relative to other nodes. Furthermore, by using partial correlation analysis (parameter: the partial correlation order is selected as the minimum order to eliminate non-directly related factors), the independence relationship between the identified significant causal pairs is stripped away, and the separation results of indirect transmission paths are obtained; Furthermore, by jointly normalizing the Granger F statistic and partial correlation coefficient of the direct causal effect node pairs, a basic effect matrix for weight calculation is generated. Furthermore, a weighted fusion function is used to combine the relative magnitude of the F-statistic with the absolute value of the partial correlation coefficient as the side weights. The calculation formula is:

[0017] in, The F-statistic is the result of Granger causality test. This represents the partial correlation coefficient for the corresponding node pair; Furthermore, the initial weights of each directed causal edge are calculated in batches according to the above formula, and a causal connection data table containing direct causal effects, indirect transmission paths and weight values ​​is generated. By combining Granger causality test and partial correlation analysis, the modified causal structure of the previous step is transformed into a quantified directed edge weight parameter, thus achieving a preliminary quantitative characterization of the causal influence of process parameters on the occurrence of glue shortage defects in the fusion zone. For example, in a twelve-layer automotive PCB lamination process, with a significance level α of 0.05 and a maximum time delay order p of 3, the Granger F-statistic for the effect of lamination pressure on resin rheological properties is 8.2, with a partial correlation coefficient of 0.74; the F-statistic for the effect of temperature gradient on defect probability is 5.6, with a partial correlation coefficient of 0.62. Substituting these parameters into the weighting formula, the causal weight of lamination pressure → resin rheological properties is... ≈0.66, the causal edge weight of temperature gradient → defect probability is ≈0.51. In the initial causal graph generated by weighted fusion, the edge weight of the pressing pressure node is significantly higher than that of the temperature gradient node, indicating that its direct causal influence on the missing glue defect is stronger in this batch. In subsequent counterfactual reasoning, prioritizing intervention on the pressing pressure parameter can significantly improve the controllability of defect risk reduction and the effectiveness of process optimization; S3.4: Integrate the identified causal relationships and weight information into a directed weighted causal graph, where nodes represent process parameters, directed edges represent causal directions, edge weights represent causal strength, and the initial dynamic causal graph model file is output as the input basis for subsequent sliding window updates and counterfactual reasoning. S3.5: Perform interventionability verification on the completed dynamic causal graph model. Verify whether the model supports intervention reasoning for key process nodes by simulating do-intervention operations, and ensure that the model has the causal interventionability characteristics required for the generation of subsequent compensation strategies. Based on the completed dynamic causal graph model file, a causal interventionability verification method (parameters: causal graph node set, directed causal edge set, initial edge weight) is used to evaluate the effectiveness of intervention reasoning for key process nodes in the model. Furthermore, by loading a dynamic causal graph model through a structural equation model (parameters: node variable function mapping, parent node set, initial coefficient matrix), each process parameter node is mapped to a structural variable in the equation to form a causal function system with input intervention channels; Furthermore, a do-dry budget injection mechanism (parameters: target node identifier, intervention value amplitude Δp) is adopted to fix the input value of the specified process parameter node to the preset intervention value, disconnect the original parent node input signal of the node, and realize direct external control simulation of the node; Furthermore, based on the structural equation model, post-intervention inference calculations are performed, and the forward propagation method is used to resolve the variables of the affected downstream nodes to obtain the defect probability distribution values ​​of each output node under the intervention state. The defect probability values ​​are calculated using the following expression:

[0018] in, This represents the set of parent nodes of node X. For structural equation functions, The total number of samples, This is a defect event. Set values ​​for intervention; Furthermore, a difference operation is performed on the defect probability output by do-intervention at different nodes and the original predicted defect probability to generate the marginal causal effect vector of each node, so as to quantify the degree of influence of the node parameter change on the defect probability. Furthermore, by verifying the significance of the marginal causal effect of each key process parameter node (parameters: significance level α, sample variance σ²), it is determined whether the causal inference pathway in the model has a stable intervention response capability. If the significance meets the preset conditions, the model is confirmed to meet the interventionability requirements. By using the causal interventionability verification algorithm, the dynamic causal graph structure constructed in the previous step is transformed into a causal reasoning model with external intervention capability and predictable response, thereby realizing the causal interventionability characteristics required for the generation of compensation strategies. For example, to verify the interferability of a pressing pressure node, the original predicted defect probability is set to 0.35. In the structural equation model, the intervention amplitude Δp = +0.12 MPa is set. A do-intervention operation is performed to disconnect the original parent node input and fix the pressure value at 6.40 MPa. The model recalculates the downstream temperature gradient and resin rheological properties, and outputs an intervention defect probability of 0.21. According to the marginal causality formula, the difference between the original probability and the intervention probability is calculated... At the same significance level of α=0.05, a significance test was conducted using the sample variance σ²=0.004. A p-value less than 0.05 indicates the intervention is effective. This validation result demonstrates that the pressing pressure node has a significant intervention effect in the current model and can be included as a high-priority parameter in the compensation strategy library to improve the targeting and feasibility of defect prediction and process optimization.

[0019] like Figure 3 As shown, step S4 involves using a sliding time window mechanism to update the dynamic causal graph model online. This is achieved by calculating the KL divergence difference between the new window data and historical causal edges, dynamically adjusting the weights of the causal edges, and reconstructing the network topology. Specifically, this includes: S4.1: Based on the initial process-defect dynamic cause-effect graph model constructed in the preceding step S3, the directed causal edges between each process parameter node and their corresponding causal strength weights are extracted as the initial input parameters of the sliding window update mechanism to form the basic topology structure for dynamic updates. S4.2: Set the length and step size parameters of the sliding time window. Based on the multi-source process parameters and defect detection data collected online, normalize the data in the current window and extract the corresponding causal edge distribution features to obtain the current causal graph probability distribution for KL divergence calculation. S4.3: Perform KL divergence calculation between the current window causal graph distribution and the historical causal graph distribution to quantify the degree of difference between the old and new causal structures, and output the KL divergence value of each causal edge as the key basis for determining whether the causal graph needs to be updated. The probability distributions of the current window causal graph and the historical causal graph are normalized, and a difference quantization algorithm based on Kullback-Leibler divergence (parameter: node edge weight vector after distribution alignment) is used to measure the degree of difference between the new and old causal structures. Furthermore, by using the probability mass function resampling method (parameter: node edge normalization weight), the mapping and comparison of the current window causal graph distribution and the historical distribution in a unified statistical space are realized, and the probability vector difference matrix corresponding to each causal edge is obtained. Furthermore, based on the KL divergence calculation formula, the difference is calculated for each directed causal edge, as follows:

[0020] in, This represents the probability distribution of causal edges within the current window. For historical window causal edge probability distribution, and Let represent the probability values ​​of the i-th directed causal edge in the two distributions, respectively; Furthermore, a numerical integration approximation method is adopted (parameter: integration step size is set to 0.001) to achieve high-precision calculation of KL divergence values ​​and generate a divergence vector containing the difference of each causal edge. By standardizing the difference vector, the KL divergence result from the previous step is transformed into an edge difference index that can be directly used in the subsequent threshold judgment mechanism, thereby achieving accurate quantification of the causal graph update requirements. For example, in the manufacturing process of a twelve-layer automotive PCB, the sliding window length is set to 300 seconds, the step size is 60 seconds, the edge weight vector between the current window pressing pressure node and the temperature gradient node is [0.25, 0.75], and the historical corresponding edge weight vector is [0.4, 0.6]. The normalized distributions after resampling are as follows: and The KL divergence formula above is used to calculate... The calculated result is 0.091. The difference value, after standardization, yields an edge difference index of 0.182, corresponding to a high structural change trend. This index can be identified as a candidate causal edge that needs to be updated in the next threshold judgment process, effectively supporting the real-time adjustment of the dynamic causal network. S4.4: Based on the KL divergence threshold judgment mechanism, the KL divergence value of each causal edge is compared. If it exceeds the preset threshold, it is determined that the causal relationship has changed significantly, triggering the dynamic adjustment mechanism of the weight of the causal edge, and marking it as a candidate edge for network reconstruction. The KL divergence calculation results are based on the probability distribution of the current window causal graph and the probability distribution of the historical causal graph. The input objects are the distribution parameter set of each directed causal edge obtained by the sliding time window mechanism and the preset threshold reference table. An edge comparison algorithm (parameters: causal edge identifier ID, current edge KL divergence value, historical edge statistical mean) is used to compare the latest KL divergence value of each causal edge with the corresponding preset threshold one by one. Furthermore, the statistical significance of the KL divergence value is verified by using a method for determining the significance of differences (parameters: preset threshold δ, significance factor α), and a list of marked edges with significant changes is obtained. Furthermore, through a candidate set generation algorithm (parameters: list of significantly changed edges, historical weight records), causal edges that are determined to have changed significantly are marked as candidate edges for network reconstruction, and a candidate edge index matrix is ​​generated; Furthermore, by using an abnormal pattern filtering method (parameters: candidate edge index matrix, edge attribute feature vector), abnormal changes in candidate edges that may be caused by sensor failure or transient interference are eliminated, and the final set of candidate edges for network reconstruction is generated. By using a significant change determination algorithm, the KL divergence calculation results from the previous step are transformed into a trigger flag for dynamic causal graph reconstruction, enabling rapid response to changes in the process and real-time adaptability of the causal graph structure. For example, in the lamination process of a twelve-layer automotive PCB, assuming a sliding time window length of 15 minutes and a step size of 5 minutes, the KL divergence value of a certain causal edge "temperature gradient → probability of missing adhesive" within the current window is... The historical average is The preset threshold δ is The edge comparison algorithm is used to perform the operation on that edge. The difference operation yields Combined with the significance factor α, let it be... The condition is that the edge has changed significantly within the current window. The candidate set generation algorithm marks this edge as a candidate edge and records it in the index matrix. The anomaly pattern filtering method checks that the change pattern of this edge in the three adjacent windows all show a continuous upward trend, eliminates false judgments caused by single-point anomalies, and retains this edge as a final network reconstruction candidate edge. In the subsequent weighted moving average update process, the weight of this candidate edge will be adjusted to combine historical weights. Compared with the current intensity The calculated weighted average value significantly improves the real-time model sensitivity to the impact of temperature gradient changes on the probability of missing glue. S4.5: The weighted moving average algorithm is used to update the weights of candidate causal edges. The weights are combined with the historical weights and the causal strength reflected by the current window data to calculate the new causal edge weights, so as to reflect the real-time dynamic change trend of the process. S4.6: Based on the updated causal edge weights and candidate edge reconstruction markers, perform local or global reconstruction operations on the causal graph topology, delete weakened or invalid causal connections, and enhance or add significantly enhanced causal paths to generate a dynamic causal graph model that adapts to the current process state. S4.7: Output the updated dynamic causal graph model to the subsequent counterfactual reasoning module S5 as the structural basis for its process parameter intervention simulation, so as to improve the causal explanatoryness and intervention accuracy of defect prediction and compensation strategy generation.

[0021] Step S5: Construct a counterfactual reasoning module based on structural equation modeling. When the predicted probability of insufficient adhesive in the fusion zone exceeds a preset threshold, perform do-intervention operations on the parameters of each node in the causal graph to simulate the marginal causal effect of different process parameter adjustments on the defect probability. Specifically, this includes: S5.1: Based on the node process parameters, directed causal edges and their weights output by the dynamic causal graph model, construct a structural equation model (SEM) framework, where each process parameter is used as a structural variable in the model, and its input path is determined by the parent node in the causal graph to form an interpretable causal function relationship between variables. Based on the set of node process parameters, directed causal edges and corresponding causal strength weights of the dynamic causal graph model output by S4.7, a structural equation model (SEM) construction method (parameters: node set, edge set, edge weight matrix, causal direction constraint) is adopted to map each process parameter to a structural variable in the model. Furthermore, by extracting the set of parent nodes of each node in the causal graph and using it as the set of input paths for the structural variable, the causal relationship between variables in the model is kept consistent with the dynamic causal graph, and an interpretable process parameter transmission link is formed. Furthermore, a functional definition method (parameters: linear influence term, nonlinear interaction term, noise term) is adopted to realize the causal function expression framework of the node structure variables, and the weighted causal contribution of the parent node input to the child node is explicitly reflected in the formula; Furthermore, by associating and binding the weight matrix with the variable mapping table, the parameterized embedding of causal edge weights in the structural equation function is realized, enabling SEM to fully inherit the intensity information of the dynamic causal graph in subsequent inference. Furthermore, by using a graph topology traversal algorithm, the computation order of the structural equation model is determined, ensuring that the parent node is computed before the child node, thereby avoiding dependency conflicts in causal computation. By using structural equation modeling, the dynamic causal graph results from the previous step are transformed into an interventionist mathematical model, realizing an interpretable causal function expression of process parameters and defect probabilities, and providing a precisely quantifiable causal calculation interface for subsequent counterfactual reasoning steps. For example, in the scenario of predicting missing adhesive in the fusion zone of a twelve-layer automotive PCB, based on the latest updated dynamic causal graph, the node set includes bonding pressure P, temperature gradient T, resin viscosity V, interlayer deviation M, vacuum fluctuation Q, and defect probability D. The edge set contains directed relationships such as P→V, T→V, V→D, M→D, and Q→D. In the edge weight matrix, the weight of P→V is 0.62, the weight of T→V is 0.48, the weight of V→D is 0.73, the weight of M→D is 0.41, and the weight of Q→D is 0.35. When constructing using SEM, the structural equation of V is set as:

[0022] in This represents the rheological measurement noise term. The structure equation for D is set as follows:

[0023] in This represents the defect detection error term. When the model calculation order is P, T→V→D, it ensures that the parent node completes its calculation before the child node. In practical applications, for a batch of data, input P=3.5MPa, T=45℃, M=0.08mm, Based on the real-time measured V = 12.5 Pa·s, the output D = 0.173 can be calculated using the above structural equations. When performing do-intervention, adjusting P to 3.8 MPa allows the model to directly recalculate V and D, outputting D = 0.152, thus significantly reducing the defect probability. This embodiment demonstrates that the SEM framework fully inherits the structure and weight information of the causal graph, supports causal response calculations with simultaneous multi-parameter or independent single-parameter interventions, and has the ability to be directly used as input for generating process compensation strategies. S5.2: For each process parameter node in the structural equation model, the coefficients of the structural equation are fitted using the least squares method based on historical observation data to obtain the quantitative causal function expression of each node variable, so as to establish an accurate process parameter-defect probability mapping model. S5.3: When the probability of missing glue in the fusion zone of the current batch output by the fusion defect prediction module exceeds the preset threshold, a do-intervention operation is performed based on the structural equation model to apply parameter perturbation of preset magnitude to the process parameters of each node in the causal graph in sequence, so as to simulate the causal response change of defect occurrence under different process adjustments. The input conditions for the structural equation model are the set of node process parameters, directed causal edges and their weights provided by the dynamic causal graph model, and the current batch defect probability index output by the fusion defect prediction module as the triggering basis. Threshold comparison method (parameter: predicted defect probability value) Preset threshold This enables the defect risk trigger determination function; Furthermore, by using the target node traversal method (parameter: set of causal graph node indices), the process parameter nodes are selected sequentially, and a preset disturbance amplitude Δx is assigned to each selected node to form an intervention operation input sequence; Furthermore, by using the do-intervention variable injection method (parameters: SEM object, intervention node ID, perturbation value Δx), the causal exogenous variable replacement of the target node is achieved, the original parent node output path is truncated and replaced with a fixed perturbation value, forming an independent intervention state model; Furthermore, a parameter iterative loading method (parameter: intervention input sequence) is adopted to realize the one-to-one intervention simulation of process parameters of all nodes in the dynamic causal graph, ensuring the observability of the causal response of each process parameter adjustment path; By using the intervention process control algorithm, the intervention simulation results of the previous step are transformed into an adjusted structural equation model input that can be used for forward propagation calculation in the next sub-step, thus realizing the prerequisite preparation for defect probability causal response analysis. For example, in a certain batch manufacturing process, the defect probability output by the fusion defect prediction module is 0.35, and the setting is... A value of 0.30 triggers an intervention decision. The dynamic cause-effect diagram includes five nodes: pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation, and vacuum fluctuation. A disturbance amplitude Δx is set for each node: pressing pressure + 0.2 MPa, temperature gradient... Resin rheological properties: viscosity +50 Pa·s; interlayer alignment deviation. Vacuum fluctuation range The do-dry perturbation injection method was used to perform exogenous substitution on the pressing pressure node, truncating the corresponding input path in the SEM and setting it to a perturbation value of +0.2 MPa to complete the model modification. The same method was applied to the remaining four nodes. After processing the full intervention sequence, five independent intervention state-structured equation model files were generated and output to S5.4 for forward propagation calculation. In this scenario, it was verified that each intervention model maintained the causal independence of the variable paths, meeting the accuracy requirements for subsequent defect probability change calculations. S5.4: Perform forward propagation calculations on the structural equation model after each do-intervention, output the probability distribution of missing glue in the fusion zone after intervention, and record the marginal causal effect value of each intervention parameter on the defect probability to quantify the causal influence of each process parameter on defect risk. S5.5: Based on the marginal causal effect value and the adjustment cost of the intervention parameters, calculate the compensation priority index of each process parameter, and output the counterfactual reasoning result including the intervention parameter name, the expected reduction in defect probability, and the strength of the intervention suggestion, for the compensation strategy generation module to call.

[0024] Step S6: Calculate the compensation priority index based on the marginal causal effect, generate a multi-level compensation strategy by considering the cost constraints of process adjustment, and output compensation recommendations including the adjustment amount of key process parameters and the expected reduction in defect probability. Specifically, this includes: S6.1: Normalize the marginal causal effects corresponding to each process parameter output by the counterfactual reasoning module to eliminate the influence of differences in the dimensions of different parameters on the calculation of compensation priority and obtain a standardized marginal causal effect vector. S6.2: Based on the standardized marginal causal effect vector and the preset process adjustment cost factor, construct a compensation priority index calculation model. The process adjustment cost factor includes equipment response delay, energy consumption increment and process window tolerance. Output the compensation priority index of each process parameter. S6.3: Sort each process parameter according to the compensation priority index, and generate a multi-level compensation strategy candidate set in combination with the feasibility constraints of process parameter adjustment. Each level strategy contains one or more high-priority process parameters and their suggested adjustment amounts. S6.4: For each level of the multi-level compensation strategy candidate set, simulate the predicted reduction of the probability of missing glue in the fusion zone based on the structural equation model to form an evaluation matrix of the expected improvement effect of the compensation strategy. For each level of the multi-level compensation strategy candidate set, a structural equation model simulation method (parameters: nodal variable function coefficient set, causal edge weight set, time lag matrix) is used to realize the defect probability response simulation based on process causal constraints. Furthermore, by sequentially inputting the set of process parameter adjustment amounts corresponding to different strategies into the structural equation model and performing forward inference operations, the predicted value matrix of the probability of missing glue in the fusion zone under each strategy is obtained. Furthermore, a differential calculation method (parameters: pre-adjustment predicted defect probability matrix, post-adjustment predicted defect probability matrix) is used to calculate the reduction in defect probability for each strategy and generate an effect vector containing probability change values. Furthermore, through effect vector normalization, the reduction in defect probability of each strategy is correlated with the corresponding process parameter adjustment amount, forming a data table of expected improvement effects of multi-level compensation strategies; Furthermore, a matrix construction method (parameters: strategy number set, adjustment vector, effect vector) is used to generate a two-dimensional compensation strategy effect evaluation matrix, where rows represent different strategy numbers and columns represent the correspondence between the adjustment amount of key process parameters and the expected reduction in defect probability. By using structural equation modeling simulation and matrix construction, the multi-level compensation strategy candidate set generated in the previous step is transformed into a quantitative evaluation matrix containing the predicted reduction magnitude, thereby realizing the quantitative evaluation of the effect of the process compensation strategy before execution. For example, in the lamination process of a twelve-layer automotive PCB, the three highest priority strategies are selected, with the first strategy adjusting the lamination pressure. The second group adjusts the temperature difference in the hot-pressing constant-temperature section. The third group increases the vacuum holding time. By inputting the adjustment vectors corresponding to the three strategies into the structural equation model, and performing forward inference calculations, the adjusted defect probability prediction values ​​are obtained as follows: , , Compared with the defect probability value before adjustment Differential calculations were performed to obtain the vectors of the magnitude of the decrease in defect probability, respectively. , , The reduction vector was correlated with the adjustment amounts of each strategy to construct a three-row, two-column effect evaluation matrix. The first column of the matrix recorded the process parameter adjustment amount, and the second column recorded the expected reduction in defect probability. In simulation verification, the second group of strategies showed a significant improvement in the overall effect between the probability reduction and adjustment cost, and was marked as the preferred strategy. It can significantly reduce the defect probability without increasing the additional pressure on the equipment, achieving the expected technical effect of process optimization. S6.5: Based on the priority index of the compensation strategy and the evaluation matrix of expected improvement effect, the optimal compensation strategy that meets the requirements of process stability and cost control is selected, and the compensation suggestions including the adjustment amount of key process parameters and the expected reduction in defect probability are output.

[0025] Step S7: The compensation suggestion is fed back to the MES system to adjust the process parameters, and the process parameters and defect detection results of the compensated batch are collected to update the node association strength of the dynamic cause-effect graph model. Specifically, this includes: S7.1: Based on the compensation suggestions output by S6, update the parameters of the process parameter configuration module in the MES system to achieve real-time adjustment of key process variables, including pressing pressure, temperature gradient, resin rheological properties, interlayer alignment deviation and vacuum fluctuation parameters. S7.2: Deploy a compensation execution status monitoring module in the MES system to collect execution feedback after parameter updates in real time, and obtain the actual process parameter set after compensation execution for subsequent causal model updates; S7.3: Deploy an automated optical inspection (AOI) system at the fusion zone inspection station in the PCB manufacturing process to perform high-precision identification of missing adhesive in the fusion zone of batch products after compensation is performed, and output a defect detection result set, including defect type, location and probability; After the batch of products undergoing compensation enters the fusion zone inspection station, the image acquisition module of the automated optical inspection (AOI) system (parameters: resolution ≥8K, acquisition frame rate ≥60fps) is invoked to achieve high-precision raw image acquisition function covering the entire fusion zone. Furthermore, by using a multispectral imaging algorithm (parameters: visible light band 400~700nm, near-infrared band 800~1000nm), the reflectivity of different materials is enhanced and the separated multispectral image datasets are obtained. Based on the acquired multispectral image data, an adaptive threshold segmentation method (parameters: local window size 51×51 pixels, threshold dynamic adjustment coefficient 0.8) is used to achieve initial screening and segmentation of the target area of ​​the glue-deficient region and generate a defect binarization mask matrix. Furthermore, a morphological filtering algorithm (parameter: structuring element radius of 5 pixels) is used to remove noise and smooth the edges of the binarized mask, resulting in an optimized set of defect contours. For the optimized defect contour set, the boundary fitting and geometric feature extraction method (parameter: minimum bounding rectangle fitting accuracy ≤ 0.1mm) is used to realize the quantitative calculation of geometric parameters such as defect area, aspect ratio and boundary curvature, and to obtain the preliminary classification basis of defect type; Furthermore, a convolutional neural network (CNN) defect classification model (parameters: network depth of 50 layers, training set covering 8 glue-deficiency morphologies) is used to determine the morphological category of glue-deficiency defects and output defect type labels; Based on type label and contour coordinate data, a sub-pixel positioning algorithm (parameter: positioning accuracy ≤ 0.05mm) is used to achieve high-precision coordinate calculation of defect location, and then registered with the workstation reference coordinate system in the AOI system to obtain a standardized coordinate set of defect location; Furthermore, based on the defective pixel area and the pixel density collected at the workstation, an area ratio calculation method is used to calculate the proportion of glue-deficient area. The quantitative assessment is expressed by the following formula:

[0026] in, This represents the actual area corresponding to the pixel with missing adhesive. This represents the total area of ​​the detection region. A comprehensive assessment of defect probability is achieved by using a mapping table between the percentage of areas with insufficient adhesive and defect types, combined with the defect location coordinates. This generates a defect detection result set containing defect type, defect location coordinates, and defect probability. By combining multispectral imaging, geometric feature quantization, classification and recognition and probability assessment, the batch detection images of the previous compensation step are transformed into structured defect detection results, so as to realize the quantifiable verification of the compensation effect and the preparation of input data for subsequent causal model updates. For example, in the inspection following a compensation process of a twelve-layer automotive PCB board fusion process, the AOI image acquisition module was set to a resolution of 8192×8192 pixels, with a frame rate of 60fps, covering the entire fusion area of ​​the single board. A multispectral imaging system acquired visible and near-infrared images, and after adaptive threshold segmentation (local window 51×51 pixels, coefficient 0.8), a preliminary defect mask was obtained. Morphological filtering was performed using a structuring element with a radius of 5 pixels to remove isolated noise. Defect contour boundary fitting achieved area quantization with an accuracy within 0.1mm. The CNN classification model determined the defect morphology to be edge-deficient glue, and sub-pixel localization obtained the defect coordinates (x=102.35mm, y=57.80mm). The area percentage calculation formula... It is 5.8mm². The area is 120mm², and the area ratio is as follows: Based on the defect probability assessment rules, the defect probability assessment is classified as high-risk. The defect detection result set accurately reflects the missing glue defect status of the compensated batch, which is used for subsequent updating of the node association strength of the dynamic cause-effect graph model. S7.4: The actual process parameter set collected in S7.2 and the defect detection result set output in S7.3 are timestamped to construct a pairwise observation dataset after compensation execution, which is used for dynamic updating of the node association strength of the causal graph model. S7.5: Based on the pairwise observation dataset after compensation execution, the edge weights of each directed causal edge in the dynamic causal graph model are re-estimated using the causal effect estimation method to correct the causal strength between process parameters and fusion defects in the model, thereby improving the accuracy and robustness of model prediction.

[0027] Step S8: Based on the change in defect probability after compensation, calculate the prediction error of the causal graph model. When the error exceeds a preset tolerance value, trigger the causal network reconstruction mechanism to form a closed-loop optimization link of prediction-intervention-verification. Specifically, this includes: S8.1: Perform feature alignment and timestamp matching on the process parameters collected after compensation and the detection results of missing glue in the fusion zone to generate a causal consistency verification dataset that can be used for model verification. S8.2: Based on the causal consistency verification dataset, perform forward inference calculation using the dynamic causal graph model, and output the predicted probability value of glue shortage defect in the fusion zone of the current batch as the model prediction output result; S8.3: Compare the predicted probability of missing glue in the fusion zone with the actual test results sample by sample, and calculate the model prediction error index. The error index includes the mean absolute error and KL divergence to quantify the deviation of the causal model prediction performance. S8.4: Compare the model prediction error with a preset error tolerance threshold. If the error index exceeds the threshold for more than a preset number of batches consecutively, generate a causal network reconstruction trigger signal to start the model adaptive update mechanism. S8.5: In response to the causal network reconstruction trigger signal, based on the latest verification data and historical causal graph structure, the causal discovery process based on the PCMCI+ algorithm is re-executed, and domain knowledge constraints are integrated to generate an updated dynamic causal graph model in order to restore the model prediction accuracy and the effectiveness of process intervention.

[0028] 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.

[0029] 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 predicting and compensating for a lack of glue defect in a fusion area of a twelve-layer automotive PCB board, characterized in that, The method comprises the following steps: S1: Collecting multi-source process parameters in the manufacturing process of a twelve-layer automobile PCB board and corresponding defect detection results of the fusion area, and constructing a paired observation data set; S2: Normalizing and removing outliers from the paired observation data set, and constructing an input matrix of a time series causal discovery algorithm based on domain knowledge constraints; S3: Based on the PCMCI+ algorithm combined with process mechanism constraints, an initial process-defect dynamic causal graph model is constructed, and a causal network topology structure is output; S4: The initial process-defect dynamic causal graph model is updated online by using a sliding time window mechanism, the KL divergence difference between new window data and historical causal edges is calculated, the weight of the causal edge is dynamically adjusted, and the network topology structure is reconstructed; S5: An anti-reality reasoning module based on a structural equation model is constructed, when the predicted probability of the fusion area defect exceeds a preset threshold, a do-intervention operation is performed on each node parameter in the causal graph, and the marginal causal effect of different process parameter adjustments on the defect probability is simulated; S6: According to the marginal causal effect, a compensation priority index is calculated, a multi-level compensation strategy is generated by comprehensively considering the process adjustment cost constraint, and a compensation suggestion containing the key process parameter adjustment amount and the expected defect probability reduction amplitude is output.

2. The method according to claim 1, wherein the twelve-layer automotive PCB board fusion area glue deficiency defect prediction and process compensation method is characterized in that, The step S6 further comprises: S7: The compensation suggestion is fed back to the MES system to execute process parameter adjustment, and the process parameters and defect detection results of the compensated batch are collected to update the node correlation strength of the dynamic causal graph model; S8: Based on the defect probability change value after compensation execution, the prediction error of the causal graph model is calculated, and when the error exceeds a preset tolerance value, the causal network reconstruction mechanism is triggered, forming a closed-loop optimization link of prediction-intervention-verification.

3. The method of claim 1, wherein the method is used for a twelve-layer automotive PCB board fusion area glue defect prediction and process compensation. The step S1 specifically comprises: Based on the sensor network of the pressing equipment, the pressing pressure time series data is obtained; Based on the multi-point infrared temperature measurement module, temperature gradient data is collected; Based on the online rheometer, resin rheological property data is obtained; Based on the high-precision optical positioning system, interlayer alignment deviation data is collected; Based on the vacuum degree sensor, vacuum degree fluctuation data is collected; Based on the AOI detection system, the defect detection result of the fusion area is obtained; The pressing pressure time series data, the temperature gradient data, the resin rheological property data, the interlayer alignment deviation data, the vacuum degree fluctuation data and the defect detection result of the fusion area are time stamped and structured stored, and a paired observation data set containing process parameters and defect labels is constructed.

4. The method of claim 3, wherein the method is used for a twelve-layer automotive PCB board, and the method comprises the following steps: The pressing pressure time series data includes pressure peak value, pressure holding time and pressure fluctuation variance; the temperature gradient data includes temperature rising rate, temperature difference in constant temperature section and interlayer temperature difference fluctuation; the resin rheological property data includes viscosity change curve, gel time and flow window width; the interlayer alignment deviation data includes X / Y direction offset, angle deflection and cumulative alignment error; the vacuum degree fluctuation data includes vacuum pumping rate, vacuum degree extreme value and periodic fluctuation amplitude; and the defect detection result of the fusion area includes defect area proportion, defect position coordinates and defect morphology classification label.

5. The method for predicting and compensating the lack of adhesive defects in the fusion area of a twelve-layer automotive PCB board according to claim 1, characterized in that, The step S2 specifically comprises: The minimum-maximum normalization processing is performed on the collected twelve-layer automobile PCB manufacturing process parameters, including the pressing pressure, temperature gradient, resin rheological characteristics, interlayer alignment deviation and vacuum fluctuation parameters, to obtain a standardized process parameter matrix; Based on the sliding window Z-score method, abnormal value detection and elimination operations are performed on each dimension data in the standardized process parameter matrix to obtain a cleaned time series data set; According to the process mechanism knowledge provided by the experts in the field of automobile electronic manufacturing, a prior causal constraint graph is constructed among the process parameters, and whether there is a direct causal relationship, time delay dependence relationship or prohibited connection relationship among the parameters is labeled; Based on the mutual information and partial autocorrelation coefficient, time delay correlation analysis is performed on the cleaned time series data set to identify the optimal time delay window between the key process parameters, and a time delay offset set among the parameters is obtained. Based on the time delay offset set, a dynamic time window interception method is used to reconstruct the cleaned time series data set to generate a multi-dimensional time series input matrix.

6. The method for predicting and compensating the lack of adhesive defects in the fusion area of a twelve-layer automotive PCB board according to claim 1, characterized in that, The step S3 specifically includes: Based on the multi-source process parameter data set after normalization processing and abnormal value elimination, the pressing pressure, temperature gradient, resin rheological characteristics, interlayer alignment deviation and vacuum fluctuation variables are input into the PCMCI+ causal discovery algorithm framework for time series causal relationship identification, and a preliminary causal network structure is constructed; The physical mechanism knowledge of automobile PCB pressing process is introduced as a prior constraint condition for constructing the causal graph, and the preliminary causal network structure is reasonably modified; Based on the modified causal structure, Granger causality test and partial correlation analysis method are used to identify the direct causal effect and indirect conduction path among the process parameters, and the initial weight of each causal edge is calculated; The identified causal relationship and weight information are integrated into a directed and weighted causal graph, and an initial dynamic causal graph model is output; The dynamic causal graph model is verified for intervenability, and whether the model supports intervention reasoning on the key process nodes is verified by simulating do-intervention operation.

7. The twelve-layer automotive PCB fusion area glue defect prediction and process compensation method according to claim 6, characterized in that, In the directed and weighted causal graph, the nodes represent the process parameters, the directed edges represent the causal direction, and the edge weight represents the causal strength.

8. The method for predicting and compensating the lack of adhesive defects in the fusion area of a twelve-layer automotive PCB board according to claim 1, characterized in that, The step S4 specifically includes: Based on the constructed initial process-defect dynamic causal graph model, the directed causal edges between the process parameter nodes and the corresponding causal strength weights are extracted to form a dynamically updated basic topology structure; The length and step parameters of the sliding time window are set, the data in the current window is normalized based on the online collected multi-source process parameters and defect detection data, and the corresponding causal edge distribution features are extracted to obtain the current causal graph probability distribution; The KL divergence is calculated between the current window causal graph distribution and the historical causal graph distribution, and the KL divergence values of each causal edge are output; Based on the KL divergence threshold judgment mechanism, the KL divergence values of each causal edge are compared, if the threshold is exceeded, it is determined that the causal relationship has changed significantly, the dynamic adjustment mechanism of the causal edge weight is triggered, and it is marked as a candidate causal edge for network reconstruction; The weight of the candidate causal edge is updated by using a weighted moving average algorithm, and a new causal edge weight is calculated by combining the historical weight and the causal strength reflected by the current window data; Based on the updated causal edge weight and the candidate edge reconstruction mark, a reconstruction operation of the causal graph topology is performed to generate a dynamic causal graph model adapted to the current process state; The dynamic causal graph model adapted to the current process state is output to the counterfactual reasoning module.

9. The method for predicting and compensating the lack of adhesive defects in the fusion area of a twelve-layer automotive PCB board according to claim 1, characterized in that, The step S5 specifically includes: Based on the node process parameters, directed causal edges and edge weights output by the dynamic causal graph model, a structural equation model framework is constructed, wherein each process parameter is a structural variable in the model, and the input path is determined by the parent node in the causal graph; For each process parameter node in the structural equation model, the least squares method is used to fit the structural equation coefficients based on the historical observation data, to obtain a quantitative causal function expression of each node variable, and to establish an accurate process parameter-defect probability mapping model; When the current batch fusion zone glue deficiency probability output by the fusion defect prediction module exceeds a preset threshold, a do-intervention operation is performed based on the structural equation model to sequentially apply a preset amplitude of parameter disturbance to each node process parameter in the causal graph; Forward propagation calculation is performed on the structural equation model after each do-intervention, the probability distribution of the intervention fusion zone glue deficiency is output, and the marginal causal effect value of each intervention parameter on the defect probability is recorded; Based on the marginal causal effect value and the adjustment cost of the intervention parameter, a compensation priority index of each process parameter is calculated, and a counterfactual reasoning result is output.

10. The method of claim 1, wherein the twelve-layer automotive PCB board fusion area glue deficiency defect prediction and process compensation method is characterized in that, The process adjustment cost constraint includes a process adjustment cost factor, and the process adjustment cost factor includes equipment response delay, energy consumption increment and process window tolerance.

Images