PCB manufacturability intelligent analysis and early warning method and system based on artificial intelligence

Abstract

The invention provides a PCB manufacturability intelligent analysis and early warning method and system based on artificial intelligence, and the method comprises the steps: collecting and marking the multi-source time sequence process parameter data in the PCB design and manufacturing process under working conditions, building a dynamic causal graph model with direction and time lag marks through a sliding window and standardization processing by applying a causal discovery algorithm, and carrying out the calculation of the dynamic causal graph model. Dynamic expression of causal relationships among process variables is realized; when manufacturing abnormity is detected, abnormity attribution is carried out by combining a Bayesian back propagation algorithm, high-contribution-degree root dependent variables are screened, the effectiveness of root causes is verified through virtual intervention simulation and statistical test, and finally verification results and a causal mode are stored in a knowledge base to support subsequent rapid matching and reasoning. According to the method, the accuracy, efficiency and interpretability of PCB manufacturing abnormity attribution are improved, and process optimization and preventive intervention are facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial process analysis, and particularly relates to a PCB manufacturability intelligent analysis and early warning method and system based on artificial intelligence. BACKGROUND

[0002] The mainstream technology in the field of current PCB manufacturability intelligent analysis and early warning system is mostly centered on artificial intelligence and machine learning, and widely adopts big data analysis, deep learning or statistical correlation modeling to predict quality risks and abnormalities in the production process. These schemes usually collect and train a large amount of production parameters, historical defective product data and equipment sensor logs to build a risk prediction model, realize online early warning in the manufacturing link or quality trend analysis; In terms of representative technical solutions, most of the existing intelligent PCB risk prediction systems adopt machine learning algorithms such as regression analysis, clustering, decision tree, random forest or deep neural network. These technologies can automatically discover complex correlations between data for fault early warning, process abnormality determination and process optimization suggestions. Typical applications include defect rate prediction in printed board manufacturing execution system, copper thickness uniformity analysis, drilling deviation detection, impedance compliance evaluation and equipment parameter abnormality reminder. The above-mentioned schemes have certain self-learning and data-driven capabilities and have become the technical foundation of PCB intelligent manufacturing; However, the existing technology also has significant limitations. The main problem is that the current AI model for manufacturing risk explanation still highly depends on statistical correlation, lacks explicit causal reasoning mechanism, and makes it difficult to trace the root cause of the explainability after the model outputs risk early warning. On the one hand, correlation-driven models are susceptible to data collinearity, hidden variables and occasional noise, which may produce pseudo-correlation or false diagnosis conclusions, reducing the credibility of risk attribution; on the other hand, the current model mostly adopts static and fixed structure of causal relationship setting, which cannot flexibly adapt to the complex dynamic environment under the frequent iteration, process replacement and equipment change in PCB manufacturing process. The causal path is difficult to evolve with the data flow in real time, making it difficult to cover new scenarios or variant working conditions for root cause positioning; The existing risk prediction system can only give abnormal detection or parameter abnormality score, and does not support risk mechanism verification based on intervention simulation (such as do-operation), lacks causal effectiveness verification mechanism for candidate root causes. System users are difficult to obtain operational explanations such as "why does the risk occur" and "which parameters can be changed to radically solve the problem", which affects the implementation of decision-making closed loop and the acceptance of engineering practice. SUMMARY

[0003] The present application provides a PCB manufacturability intelligent analysis and early warning method and system based on artificial intelligence to solve the above-mentioned technical problems.

[0004] The technical scheme of the present application is implemented as follows: the PCB manufacturability intelligent analysis and early warning method based on artificial intelligence comprises: S1: acquiring multi-source time sequence process data in the PCB design stage and manufacturing process, the multi-source time sequence process data comprising circuit layer stacking parameters, drilling density sequence, copper thickness distribution time sequence, impedance control parameters, and device state logs from different process nodes, and attaching process condition labels to each type of data to identify corresponding production environment conditions; S2: performing sliding window segmentation and standardization processing on the multi-source time sequence process data, identifying potential causal connection relationships between variables in each time window based on local condition independence testing and delayed mutual information analysis, generating an initial causal edge set with direction and time lag labels as the basis for constructing a dynamic causal graph model; S3: constructing an evolvable dynamic causal graph model based on the initial causal edge set, the model being represented in the form of a weighted directed graph, wherein a node represents a process parameter variable, and an edge weight reflects causal influence strength, and the existence consistency of each causal edge under different conditions is evaluated by introducing a sliding historical window comparison mechanism to filter pseudo-causal connections caused by noise or accidental correlation, and output a robust updated dynamic causal structure; S4: when an AI analysis module detects a manufacturing risk abnormal event, injecting the observation value of an abnormal variable as an evidence node into the dynamic causal graph model, starting a confidence update algorithm based on Bayesian back propagation, reversely transferring abnormal support along the causal edges, calculating the contribution degree of each upstream variable to the abnormal event, and generating a hierarchical attribution score vector; S5: screening high-contribution candidate root cause variables based on the attribution score vector, calling a virtual intervention simulator to perform a do-operation intervention simulation on the candidate root cause variables, i.e., forcibly setting their values to a reference state within the normal process range, observing and recording the change trend of the posterior probability of downstream risk variables, and forming an intervention response curve set; S6: judging the effectiveness of the candidate root cause according to the intervention response curve set, if a certain candidate variable causes a significant decrease in downstream risk probability after intervention, it is confirmed as the root cause, and the causal path pattern is stored together with the corresponding process condition label in the causal knowledge memory bank, constituting the result output of this time attribution tracing; S7: in the subsequent risk analysis process, when a similar process condition label or a similar abnormal pattern is detected, a matching historical causal path template is retrieved from the causal knowledge memory bank, and the retrieval result is used as prior guide information to initialize the structure search space of the current dynamic causal graph model, improving the convergence speed and stability of attribution reasoning; S8: generating an explainability report based on the attribution tracing result, the report containing root cause variable name, causal impact path, intervention verification result and applicable working condition range, fed back to the design end or manufacturing control system through a man-machine interaction interface, triggering preventive process adjustment or design rule revision action.

[0005] The application also provides an artificial intelligence-based PCB manufacturability intelligent analysis and early warning system, which uses the artificial intelligence-based PCB manufacturability intelligent analysis and early warning method to intelligently analyze and early warn the PCB manufacturability.

[0006] The artificial intelligence-based PCB manufacturability intelligent analysis and early warning method has the following beneficial effects: (1) The application introduces a time series causal discovery engine and a dynamic causal structure learning mechanism to realize the ability to autonomously identify and continuously update potential causal relationships from multi-dimensional process parameter time series, combined with a causal stability evaluation strategy under a sliding window, the system can automatically identify and eliminate pseudo-causal connections caused by noise interference or accidental correlation, significantly improving the accuracy and robustness of the causal graph construction, making the entire reasoning architecture have stronger environmental adaptability and long-term operation reliability.

[0007] (2) The application innovatively constructs a message passing-based reverse belief propagation mechanism and integrates a virtual intervention simulator to form a "reasoning-verification" double-loop attribution system: when a risk alert is triggered, the system takes the abnormal node as the observation evidence, propagates the contribution score along the dynamic causal graph in reverse, realizes the hierarchical attribution ordering of upstream variables, and accurately locks the key influencing factors; further, through do-operation simulation, the system simulates virtual intervention on high-contribution nodes to quantitatively evaluate their regulation effect on downstream risk states, thereby completing the effectiveness test of causal hypothesis, greatly enhancing the credibility and explainability of the diagnosis conclusion; this mechanism avoids the misjudgment risk caused by the dependence of traditional statistical methods on correlation ordering, truly realizes the transition from "phenomenon correlation" to "causal driving", and significantly improves the scientificity and decision support ability of root cause positioning. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a flowchart of the artificial intelligence-based PCB manufacturability intelligent analysis and early warning method of the application; Figure 2 is a sub-flowchart of the artificial intelligence-based PCB manufacturability intelligent analysis and early warning method of the application; Figure 3 is another sub-flowchart of the artificial intelligence-based PCB manufacturability intelligent analysis and early warning method of the application. DETAILED DESCRIPTION

[0009] Embodiments of the present application are described below in detail with reference to the drawings, in which like or similar elements or components each having the same or similar function are designated by the same reference numerals throughout the several views. The embodiments described below are exemplary only, and are not to be taken in a limiting sense, but the entire scope of the application shall include any novel one, or combination of any novel ones, of the features described herein.

[0010] The disclosure given below provides a number of different embodiments or examples for implementing different aspects of the application. In order to simplify the disclosure of the application, the components and arrangements of the specific examples are described below. Of course, these are merely examples and are not intended to limit the application. Furthermore, the application can refer to the same reference numerals and / or letters in different examples. Such repetitions are for the sake of a simplified and clear disclosure, and do not in themselves indicate a relationship between the various embodiments and / or arrangements discussed.

[0011] As shown in the drawings, Figure 1 The present application provides an artificial intelligence-based PCB manufacturability intelligent analysis and early warning method, specifically comprising: S1: Obtain multi-source time sequence process data in the PCB design stage and manufacturing process, the multi-source time sequence process data including circuit layer stacking parameters, drilling density sequence, copper thickness distribution time sequence, impedance control parameters, and equipment state logs from different process nodes, and attaching a process condition label to each type of data to identify the corresponding production environment conditions; S2: Perform sliding window segmentation and standardization processing on the multi-source time sequence process data, identify potential causal connection relationships between variables in each time window based on local conditional independence testing and delayed mutual information analysis, generate an initial causal edge set with direction and time lag labels, and use it as the basis for constructing a dynamic causal graph model; S3: Based on the initial causal edge set, an evolving dynamic causal graph model is constructed, which is represented in the form of a weighted directed graph, where the nodes represent process parameter variables and the edge weights reflect the causal influence strength. By introducing a sliding history window comparison mechanism, the consistency of each causal edge under different process conditions is evaluated, and false causal connections caused by noise or accidental correlation are filtered out, outputting a robust updated dynamic causal structure; S4: When the AI analysis module detects a manufacturing risk abnormal event, the observation value of the abnormal variable is injected as an evidence node into the dynamic causal graph model, a confidence update algorithm based on Bayesian back propagation is started, the abnormal support is transferred along the causal edges in reverse direction, the contribution degree of each upstream variable to the abnormal event is calculated, and a hierarchical attribution score vector is generated; S5: Based on the attribution score vector, high-contribution candidate root cause variables are screened out, a virtual intervention simulator is called to perform a do-operation intervention simulation on the candidate root cause variables, i.e., to forcibly set their values to a reference state within the normal process range, to observe and record the change trend of the posterior probability of the downstream risk variables, and to form an intervention response curve set; S6: According to the intervention response curve set, the effectiveness of the candidate root cause is judged, if a certain candidate variable causes a significant decrease in downstream risk probability after intervention, it is confirmed as the root cause, and the causal path pattern is stored together with the corresponding process condition label in the causal knowledge memory bank, which constitutes the result output of this attribution tracing; S7: In the subsequent risk analysis process, when similar process conditions or similar abnormal patterns are detected, the matching historical causal path template is retrieved from the causal knowledge memory bank, and the retrieval result is used as prior guidance information to initialize the structure search space of the current dynamic causal graph model, improving the convergence speed and stability of the attribution reasoning; S8: Based on the attribution tracing result, an explainability report is generated, which includes the root cause variable name, the causal impact path, the intervention verification result and the applicable process condition range, which is fed back to the design end or the manufacturing control system through the man-machine interaction interface, triggering preventive process adjustment or design rule correction action.

[0012] The step S1: Obtain multi-source time-series process data in the PCB design stage and manufacturing process, the multi-source time-series process data includes circuit layer stacking parameters, drilling density sequence, copper thickness distribution time sequence, impedance control parameters and equipment state logs from different process nodes, and a process label is attached to each type of data to identify its corresponding production environment condition. Specifically, it includes: S1.1: Obtain circuit layer stacking parameter data from the PCB design system, the circuit layer stacking parameters include the thickness, dielectric constant, copper foil type and stacking order of each signal layer and dielectric layer, automatically extract based on the design file parsing interface, generate a standardized layer stacking structure description table as a basic input factor representing the PCB physical configuration; Based on the original engineering file data derived from the PCB design system, a design file parsing interface (parameters: support Gerber, ODB++, IPC-2581 format, parsing depth accurate to layer element attribute) is used to realize automatic identification and data extraction of the structure parameters of each signal layer and dielectric layer; Further, through the layer stacking structure analysis algorithm (parameters: dielectric constant extraction rule, thickness calibration error ≤0.001mm), the thickness of each layer of dielectric and the type of copper foil are classified and arranged, and multi-dimensional physical configuration attribute results are obtained; Further, a parameter consistency verification method (parameter: matching threshold of 99% between original design file annotation value and material database measured value) is adopted to realize consistency verification of the extracted dielectric constant and thickness data in cross-layer combination, and to generate traceable verification records; Further, a layer stacking sequence serialization algorithm (parameter: sequential labeling and adjacent layer dielectric property difference calculation) is adopted to realize strict coding of the stacking order of different copper foil types and dielectric materials, and to generate a structure sequence key-value pair set; Through a standardized mapping processing method (parameter: global node namespace rule is adopted), the results of the previous step are converted into a standardized layer stacking structure description table, realizing the physical configuration data representation which can be directly used as the input node of the dynamic causal model; For example, in a certain 12-layer high-speed signal PCB design file, the design file parsing interface configuration supports IPC-2581 format, the parsing accuracy is 0.0005mm, the signal layer thickness is 0.035mm, the dielectric layer thickness is 0.10mm, the dielectric constant range is between 4.2 and 4.5, and the copper foil type is standard electrolytic copper and deposited copper alternately stacked. The maximum thickness deviation of the dielectric extracted by the layer stacking structure analysis algorithm is 0.0008mm, which meets the preset accuracy requirement. The parameter consistency verification method compares the dielectric constant annotated in the design file with the measured value in the material database, and the matching degree reaches 99.7%, generating a verification record ID LAYERCHK202405. The layer stacking sequence serialization algorithm outputs the code [Cu-E1|Die-4.23|Cu-D1|Die-4.22|...|Cu-E2], and the maximum sequence difference is 0.12, which meets the difference calculation accuracy. In the layer stacking structure description table generated by the standardized mapping processing, each node has a unique identifier (such as LayerID001 corresponding to signal layer 1), and contains thickness, dielectric constant, material type and sequence label attribute fields. Finally, this table is used as the input of the initial node set of the dynamic causal graph, realizing the data link connectivity from the design end to the causal modeling end; S1.2: Collecting drilling density sequence data, the drilling density sequence is organized in the form of time sequence of the number of through holes, blind holes and buried holes per unit area changing with position coordinates, using the CAM process decomposition module for coordinate mapping and region division, calculating the local drilling density sliding mean, and outputting the drilling load distribution factor with spatial positioning attribute; S1.3: Collecting continuous monitoring of copper thickness distribution time series data in the manufacturing process, the copper thickness distribution time series is obtained by periodic sampling of multiple sampling points by the online X-ray thickness gauge of the electroplating line, and the original measurement value is subjected to outlier rejection and trend smoothing processing, and the copper thickness fluctuation spectrum graph factor reflecting the process stability characteristics is generated based on the moving average filtering algorithm; S1.4: Obtain the impedance control parameter sequence, which includes the trace width, trace spacing, reference plane distance, and actual measured differential impedance value, extract the effective dielectric constant time curve based on the transmission line model inversion, mark the deviation interval combined with the target impedance tolerance range, and generate the impedance compliance criterion factor; In the multi-source time sequence process data collection of the PCB manufacturability analysis system, the impedance control parameters are used as input data sources, which include the trace width, trace spacing, reference plane distance, and online measured differential impedance value; A high-precision impedance test module (parameters: sampling frequency 1 kHz, impedance resolution 0.1 Ω) is used to realize continuous sampling of differential impedance values and obtain the original impedance time sequence signal; Further, the transmission line model inversion algorithm (parameters: based on the edge-coupled microstrip line model, input trace geometric parameters and impedance value) is used to calculate the effective dielectric constant and obtain the effective dielectric constant time curve. The formula is

[0013] Wherein is the effective dielectric constant, is the speed of light constant, is the signal transmission speed, which is determined by measuring the signal flight time and transmission path length value; Further, the effective dielectric constant time curve is smoothed by sliding window mean filtering (window length of 5 sampling points) to suppress transient peak noise and generate smoothed curve data; Further, by comparing with the preset target differential impedance tolerance range (tolerance range based on design specifications, center value such as 50 Ω, allowed deviation ±5%), the deviation interval is marked, and the deviation marking sequence is obtained; Further, through the logic rule matching processing method, the deviation marking sequence is associated with the working condition label to generate the impedance compliance criterion factor, which includes variable identification, deviation start and end time, corresponding working condition label, and deviation degree quantization index; Through the above algorithm chain, the original impedance test results are converted into structured criterion data with working condition context, realizing the expected technical effect of impedance compliance analysis; For example, in the impedance monitoring process of a certain high-speed signal PCB, the geometric data of trace width 0.15 mm, trace spacing 0.20 mm, and reference plane distance 0.10 mm are collected, and the impedance test module records the differential impedance time sequence value range of 48.7 Ω to 51.3 Ω. The effective dielectric constant time curve is calculated by transmission line model inversion, and the average value is about After 5-point window mean filtering, the curve is significantly smoothed, and the instantaneous peak is reduced to no more than Compared with the target impedance of 50Ω, ±5% tolerance, three deviation intervals are marked out, occurring at 120ms to 130ms, 310ms to 315ms and 500ms to 505ms, with a deviation amplitude of 1.3Ω to 2.5Ω. Combined with the working condition label record, these three deviation intervals all correspond to the high-speed drilling and post-compression stage of the same batch production. The finally generated impedance compliance criterion is used for subsequent sliding window condition independence test, which significantly improves the capture accuracy of manufacturing risks related to impedance abnormalities in the causal structure learning stage; S1.5: Integrate device state log data from different process nodes, including drill speed, compression temperature and pressure curve, light plotter exposure energy record and AOI detection alarm code, use a unified log protocol for timestamp alignment and semantic annotation, add the working condition label corresponding to the current production batch, and output the multi-source heterogeneous process data stream factor with environment context identification.

[0014] The step S2: sliding window segmentation and standardization processing of the multi-source time series process data, based on local conditional independence test and delay mutual information analysis, identify the potential causal connection relationship between variables in each time window, generate an initial causal edge set with direction and time delay label, as the basis for constructing dynamic causal graph model. Specifically includes: S2.1: Perform sliding window segmentation processing on the multi-source time series process data stream from the PCB design and manufacturing links, based on the preset time window length and step length parameters, divide the continuous time series data into overlapping time segment sequences, generate a segmented data block set with local time series integrity to support dynamic capture of local causal relationship; The multi-source time series process data from the PCB design and manufacturing links is segmented by the sliding window segmentation method (parameters: time window length Δt, step length δt, window overlap rate r), which divides the continuously collected multi-dimensional process parameter time series into data segment sequences containing local time context according to fixed intervals; Further, through the window index mapping mechanism (parameters: window number, start timestamp, end timestamp), the exact positioning of the original data stream position corresponding to each time window is realized, and the segmented data block set with index annotation is obtained; Further, through the working condition label inheritance strategy (parameters: working condition label type, label mapping table), the segmented data block inherits the working condition identification of the corresponding time interval, and generates a structured data block record containing time window parameters and working condition labels; Further, by the cross-parameter synchronization slicing algorithm (parameters: variable number M, synchronization tolerance ε), the values of the multi-source process parameter variables in the same time window are aligned column by column, a locally time-sequentially complete multi-variable data slice matrix is constructed, and a segmented data block set factor is generated; Through the above sliding window segmentation processing mode, the continuous time series data stream of the previous step is converted into a segmented data block set factor with local time sequence integrity and structured label, realizing the expected technical effect of supporting dynamic capture of local causal relationship; For example, in a PCB batch production data processing scenario, for four types of variables including circuit layer stacking parameters, drilling density sequence, copper thickness distribution time sequence and impedance control parameters, the time window length Δt is set to seconds, the step size δt is seconds, the window overlap rate r is , the sliding window segmentation method is applied to cut the multi-variable sequence with a length of seconds into window segments. Through the window index mapping mechanism, the starting and ending time stamps of each window are recorded, and the working condition label of the interval (such as the pressing temperature range T= ℃, equipment model combination E1-E3) is associated. The cross-parameter synchronization slicing algorithm is used to align the data columns of the four types of variables in each time window, generating local time sequence data matrices, each with a dimension of . The segmented data block set factor significantly improves the stability of the test statistic and the suppression effect on noise in the subsequent zero-mean standardization and conditional independence test steps, ensuring the local sensitivity and global robustness of dynamic causal structure learning; S2.2: Zero-mean standardization and range normalization processing is performed on the segmented data block set factor, based on the statistical characteristics (mean and standard deviation) within the sliding window of each process parameter variable, to eliminate the interference caused by different dimensions and amplitude ranges, and output the standardized local time sequence segment data matrix factor, to improve the numerical stability of the subsequent correlation and independence test; S2.3: Based on the standardized local time sequence segment data matrix factor, a local conditional independence test algorithm (such as the partial correlation test in the PC-mid algorithm) is executed to detect whether there is a significant direct dependency relationship between any two process parameter variables under the premise of controlling the influence of other covariates, and generate a preliminary undirected dependency edge set factor for excluding pseudo-correlation caused by indirect association; S2.4: For each pair of variable combination in the set of undirected dependency edges, apply the delayed mutual information analysis method to calculate the mutual information value under different time lags, determine the direction of causal action and time lag parameter based on the time delay corresponding to the maximum mutual information, generate a set of directed causal candidate edges with direction labels and time lag labels, and reflect the possible driving-response relationship between variables; S2.5: Perform significance threshold filtering and minimum description length criterion checking on the set of directed causal candidate edges, eliminate weak causal connections with mutual information values below the noise level or not conforming to the principle of simplicity, and output the final initial causal edge set with direction and time lag labels as the core input basis for constructing an evolvable dynamic causal graph model.

[0015] As shown in Figure 2 S3: Based on the initial causal edge set, an evolvable dynamic causal graph model is constructed in the form of a weighted directed graph, where nodes represent process parameter variables and edge weights reflect causal impact strength. The consistency of each causal edge under different working conditions is evaluated by introducing a sliding history window comparison mechanism to filter false causal connections caused by noise or accidental correlation, and a robust updated dynamic causal structure is output. Specifically, it includes: S3.1: Based on the initial causal edge set with direction and time lag labels output by S2, define the node set of the dynamic causal graph model, where each node corresponds to a process parameter variable in the multi-source time series process data. The process parameter variables include line layering parameters, drilling density sequences, copper thickness distribution time series, impedance control parameters, and device state log items. A unique identifier is generated using a variable namespace mapping mechanism to form a standardized node registry; Based on the initial causal edge set with direction and time lag labels output by S2, a node mapping initialization method (parameters: node namespace prefix, variable type code set) is used to convert each process parameter variable into a node entity of the dynamic causal graph model; Further, through a multi-source variable analysis algorithm (parameters: line layering parameter set, drilling density sequence set, copper thickness distribution time series set, impedance control parameter set, device state log item set), semantic decoupling and type classification of various input data sources are realized, and a structured variable metadata table is obtained; Further, a unique identifier generation algorithm (parameters: hash length 16 bits, collision detection threshold 0.01) is used to generate a non-repetitive node ID for each process parameter in the variable metadata table, and a node attribute mapping matrix is generated to ensure the consistency and traceability of node references during the evolution of causal structure; Further, by the node registration rule engine (rule set: variable type constraint, data source identification, working condition label binding), the system registration of node attributes and their metadata is realized, forming a standardized node registration table factor; By using the node visualization coding method (parameters: color mapping table, shape coding rule), the node registration table factor is converted into a visual symbol set, providing unified identification support for subsequent causal graph structure display and interaction; Through the node mapping initialization algorithm and the unique identifier generation processing method, the initial causal edge set in the previous step is converted into a node registration table that is symbolized, standardized, and has a unique constraint, realizing the expected technical effect of node set definition and attribute solidification in the dynamic causal graph model construction stage; For example, in a multi-source data processing scenario involving five types of process parameters, the input line layering parameters include signal layer thickness 0.12mm, dielectric layer thickness 0.08mm, dielectric constant 4.2, copper foil type Hoz, drill hole density sequence with 120 through holes, 45 blind holes, and 30 buried holes per unit area, copper thickness distribution time sequence with a mean value of 35μm at 10 sampling points on the electroplating line, target impedance of the impedance control parameter is 90 ohms with an allowable deviation of ±5 ohms, and the device state log items include drill speed 3500rpm, pressing temperature 150℃, pressure 2.0MPa, and exposure energy 200mJ. The multi-source variable analysis algorithm classifies each type of parameter into the corresponding variable type set; the unique identifier generation algorithm calculates the hash value for each variable metadata, such as the line layering parameter node ID , and the drill hole density sequence node ID , effectively avoiding duplication; the node registration rule engine binds working condition labels for each node, such as the corresponding batch "Batch202309" and production line number "Line05", forms a node registration table and attaches visual coding, with the line layering parameter node represented by a blue square and the drill hole density sequence node represented by a green circle. The final output standardized node registration table covers all five types of variables, meeting the node set definition requirements of the dynamic causal graph model construction, and providing good uniqueness and readability support for subsequent causal structure maintenance and evolution; S3.2: Perform impact strength quantification processing on each causal edge in the initial causal edge set, calculate the edge weight value based on the Granger causality strength coefficient and the standardized regression weight fusion algorithm, the Granger causality strength coefficient is determined by the multivariate vector autoregressive model residual ratio, and the standardized regression weight is corrected by the cross-time window parameter stability test, output the weighted directed edge set to represent the causal impact strength factor; The input condition for performing the influence intensity quantification process on the initial causal edge set is that each causal edge is composed of a source node variable and a target node variable, and is attached with a direction and a time lag label, and that the source and target variables both have standardized time series data, satisfying the numerical stability requirement for constructing a multivariate vector autoregressive model; A multivariate vector autoregressive (VAR) model is used to establish the joint dynamic relationship of the source node and the target node and their related control variables, a least squares method is used to fit the coefficient matrix of each node at different lag stages, and the Granger causality strength coefficient of the source node to the target node is calculated. Based on the residual variance ratio defined as:

[0016] Wherein, is the residual standard deviation of the restricted model without the source node lag term, is the residual standard deviation of the complete model including the source node lag term; Further, the stable influence intensity of the source node to the target node at different time windows is calculated by the standardized regression weight, and the following formula is used for standardization:

[0017] Wherein, is the standardized regression weight after stabilization, is the original value of the source node coefficient obtained by regression fitting, is the standard deviation of the target node. The standardization process eliminates the dimensional difference, so that the weights between different variables are comparable; Further, a cross-time window parameter stability test is performed, and the variance of the sequence obtained for the same causal edge in the continuous sliding window is calculated. When the variance is lower than the preset stability threshold, the weight value is retained; when the threshold is higher, the weight value is corrected by exponential smoothing to suppress the abnormal strength caused by accidental fluctuations; Further, the Granger causality strength coefficient is weighted and fused with the standardized regression weight after stabilization The following fusion formula is used:

[0018] Wherein, is the maximum weight of the causal edge, and is the fusion weight coefficient, which is determined according to the performance optimization of the verification set; Through the above fusion calculation, the weight matrix of the weighted directed edge set is obtained, and the weight value represents the comprehensive influence intensity of each causal connection, which is used as the input for subsequent robustness evaluation and filtering of the dynamic causal graph model; Exemplarily, in the causal edge set of a certain PCB production batch, the source node is selected as the copper thickness distribution parameter X1, and the target node is selected as the impedance control parameter Z2. Using the VAR model, the order is set to 2, and the fitting is obtained under the restricted model = 0.15, and under the complete model = 0.10, the Granger strength coefficient = 1.5 is calculated. Based on the standardization regression weight formula, the original coefficient = 0.08, the target node standard deviation = 0.02, and = 4.0 is obtained. The variance of the cross-time window test is 0.05, which is lower than the threshold value 0.1, so the weight is directly retained. The fusion coefficient = 0.6, = 0.4, and the fusion formula is calculated =(0.6×1.5+0.4×4.0) / (1.0)=2.5, which is used as the causal influence strength value of the copper thickness distribution to the impedance control. The value remains at a high level after subsequent robustness filtering, and the edge is finally retained in the dynamic causal graph and used for risk attribution analysis; S3.3: A sliding historical window comparison mechanism is constructed, which maintains the existence records of causal edges under N similar working conditions in the past according to batch rolling, and based on the working condition label matching strategy, the appearance frequency of the current causal edge in different windows is statistically analyzed. When the consistency of a certain causal edge is lower than the preset threshold value, it is determined that it is caused by accidental correlation or noise interference, and a pseudo-causal screening flag is generated; Based on the weighted directed edge set and the corresponding working condition label record, a working condition label matching algorithm (parameters: the label matching threshold is set to 0.85, and the similar working condition category is defined according to the process constraint) is used to realize the historical working condition screening function, and batches with high similarity to the current working condition are selected from the historical data set as the maintenance object of the sliding historical window; Further, through the batch rolling maintenance mechanism (parameters: the window length N is 5, and the window step is 1 batch), the existence records of causal edges under N matching working condition conditions in the past are accumulated, and a time stamp and a working condition category identifier are added to each record to obtain a causal edge time sequence existence matrix data with working condition labels; Further, through the appearance frequency statistical algorithm (parameters: the frequency calculation method is the edge existence count / window total batch number), the appearance frequency of the current weighted directed causal edge in different historical windows is solved, and an edge consistency score vector is obtained, reflecting the stability of the causal connection in the cross-working condition dimension; Further, a consistency threshold judgment method is adopted (parameter: preset signal threshold is 0.6, and a stable edge is judged below the threshold), a low consistency edge recognition function is realized, a binary pseudo-causal flag signal is generated when the frequency value of a certain causal edge in a historical window is lower than the set threshold, and the edge is identified as being caused by accidental correlation or noise interference; Through the sliding historical window comparison mechanism and the consistency statistical analysis, the weighted causal edge set after the influence intensity quantization in the previous step is converted into a causal connection list with a stability label, and the expected technical effect of the robustness evaluation of the causal structure is realized. For example, for the weighted causal edge between the line layer stacking parameter and the copper thickness distribution time sequence in the PCB manufacturing process causal graph model, the working condition label similarity threshold is set to 0.85, the sliding window is maintained under the similar working condition for 5 batches, and the existence state of the edge in each batch is recorded. The statistical result shows that the appearance frequency of the edge is 0.4, which is lower than the preset consistency threshold 0.6, and a pseudo-causal screening flag is generated. In this example, the window length N = 5 batches, the step is 1 batch, and the frequency calculation formula is

[0019] Wherein is the number of batches in which the causal edge e appears, and N is the total number of batches in the sliding window. In this scenario, = 2, N = 5, frequency = 0.4, and the edge is determined as a low consistency edge and marked as a pseudo-causal flag. This edge is removed in S3.4 according to the flag. This mechanism significantly improves the robustness of the dynamic causal structure in actual operation, and reduces the false connection caused by accidental noise; S3.4: Perform pseudo-causal filtering operation on the weighted directed edge set, remove the low consistency causal edge according to the pseudo-causal screening flag generated in S3.3, and retain the high frequency stable appearing causal connection. The weight of the remaining edge is subjected to variance weighting and then normalized to improve the robustness of the causal structure, and the cleaned and updated dynamic causal graph structure is output. S3.5: Inject the cleaned and updated dynamic causal graph structure into the evolvable graph storage engine, which supports incremental structure update and version snapshot management, and realizes the joint storage of node attributes, edge weights and working condition context metadata based on graph database technology, so that the dynamic causal graph model has the characteristics of continuous evolution with new data inflow, and constitutes the final output result of this dynamic causal modeling. On the basis of the cleaned and updated dynamic causal graph structure as the input condition, an incremental graph structure updating method (parameters: new causal edge set, update batch number and working condition label index) is adopted to realize the structure merging and optimization of the registered nodes and edges in the existing storage engine. Further, through the version snapshot management algorithm (parameters: version number generation rule, update trigger condition), a structural snapshot file of the dynamic causal graph model after each incremental update is generated, and the difference information between versions is recorded to support subsequent backtracking and comparison functions. Further, a joint storage method based on a graph database (parameters: node attribute mode, edge weight mode, and working condition context metadata mode) is used to realize unified storage of node sets, weighted directed edges, and associated working condition labels, and to accelerate the retrieval performance of nodes and edges through an indexing mechanism. Further, an attribute-edge weight consistency verification algorithm (parameters: threshold comparison, weight drift rate calculation formula) is used to verify the consistency of the injected nodes and edges, ensuring the compatibility of the new database structure and the original dynamic causal graph in terms of semantics and numerical values. The formula is as follows:

[0020] wherein, is the original edge weight, is the updated edge weight, and when the calculation result exceeds the preset drift rate threshold, the weight correction process is triggered. Through the graph structure persistence processing method, the cleaned dynamic causal graph result of the previous step is converted into a storage unit that can evolve with new data flow, realizing the continuous availability and traceability of the dynamic causal model structure. For example, under the condition that the PCB manufacturing working condition label is "thick plate high-density interconnection", a cleaned dynamic causal graph structure containing 120 nodes and 275 weighted directed edges is injected into the Neo4j graph database storage engine, the incremental update batch number is set to 45, and the new edge set is 25. The version snapshot management algorithm is triggered by the condition that the time since the last update is not more than 5 minutes and the number of new edges is ≥20, generating version number "V20240512_45" and recording the difference matrix of new edges and weight changes compared with version "V20240512_44". In the joint storage process, the node attribute mode includes variable ID, physical quantity category, and statistical range, the edge weight mode records Granger causality strength and stability score, and the working condition context metadata mode includes production line code and temperature and humidity environment parameters. In the attribute-edge weight consistency verification result, the drift rate of 3 edges exceeds the threshold of 0.15, triggering the automatic regression correction process of the weight. The final output of the persistent dynamic causal graph realizes automatic addition of new nodes and edges under the condition of subsequent data flow, and maintains the full comparison capability with historical versions, significantly improving the controllability and operation efficiency of causal structure evolution.

[0021] As Figure 3As shown, the step S4: when the AI analysis module detects a manufacturing risk abnormal event, the observation value of the abnormal variable is injected as an evidence node into the dynamic causal graph model, a confidence update algorithm based on Bayesian back propagation is started, abnormal support is reversely transmitted along the causal edge, the contribution of each upstream variable to the abnormal event is calculated, and a hierarchical attribution score vector is generated. Specifically, it includes: S4.1: Based on the manufacturing risk abnormal event identification output by the AI analysis module, the abnormal variable name associated with it and its observation value sequence in the current time window are obtained, the abnormal variable is a node variable in the dynamic causal graph, the over-limit type, occurrence time and spatial positioning information of the variable are extracted by using the abnormal detection result analysis interface, and the abnormal evidence factor with context annotation is generated as the input basis for subsequent evidence injection operation; Based on the manufacturing risk abnormal event identification output by the AI analysis module in real time, the abnormal event analysis method (parameters: event ID, time window range, working condition label set) is used to realize the unique identification and correlation search of the abnormal variable; Further, through the variable index analysis algorithm (parameters: dynamic causal graph node registration table, abnormal event field code), the abnormal event identification is mapped to the target node variable in the dynamic causal graph, and the original observation value sequence of the variable in the current time window is obtained; Further, the over-limit type extraction algorithm (parameters: variable observation value sequence, control specification limit table) is called by using the abnormal detection result analysis interface to realize the classification of the abnormal variable deviating from the normal process range, and the over-limit type data including the over-limit category, deviation amplitude and unit are obtained; Further, the time positioning analysis method (parameters: observation value sequence timestamp set, abnormal event start and end marker) is used to accurately obtain the first time and duration of the abnormal occurrence, and generate the time positioning index; Further, the spatial mapping calculation method (parameters: process parameter spatial position information, PCB design coordinate system conversion matrix) is used to realize the spatial position calibration of the abnormal variable, and output the spatial positioning coordinate vector; Through the context information integration algorithm (parameters: over-limit type data, time positioning index, spatial positioning coordinate vector, working condition label set), the above results are fused to generate the abnormal evidence factor with context annotation, and the technical effect of recording the abnormal state and its occurrence environment is realized; Exemplarily, in a high-speed drilling process monitoring scene, the AI analysis module outputs an abnormal event ID EVT-20240315001, and the current working condition label contains "batch number BCH-5501", "drilling machine model DRL-X200", and "material type FR4". The event ID is matched with the dynamic causal diagram node registry by the analysis algorithm, the node variable name is locked as "blind hole drilling density", the observation value sequence in the past 180 seconds is extracted, the unit is hole / mm2, and the maximum value reaches 3.25. The over-limit type extraction algorithm is called, combined with the control specification limit value 2.80 holes / mm2, it is judged as "density over-limit" type, and the deviation amplitude is 0.45 holes / mm2. The time positioning analysis method calculates the abnormal starting time as 2024-03-15 10:32:14 according to the monitoring timestamp, and the duration is 58 seconds. The space mapping calculation method uses the CAD coordinate system conversion matrix to locate the abnormal area to the third area coordinate (120 mm, 45 mm) of the PCB board surface. The context integration algorithm fuses the above data to generate abnormal evidence factors, which contain abnormal variable name, observation value sequence, over-limit type "density over-limit", time "start point 2024-03-15 10:32:14 / duration 58s", spatial position "(120 mm, 45 mm)" and working condition label, as the input basis of the S4.2 evidence injection operation, to ensure that the causal reasoning in the subsequent backward confidence propagation process has a complete time and space and working condition context; S4.2: The abnormal evidence factor with context annotation is injected into the corresponding variable node of the constructed dynamic causal diagram model as an observation evidence node, and the evidence fixing operation is performed, that is, the value of the node is locked as its actual observed abnormal state, the conditional probability constraint is formed, the joint probability distribution of the whole graph network is updated, and the local belief state factor affected by the evidence is generated, providing an initial driving condition for the backward confidence propagation; S4.3: On the basis of the local belief state factor affected by the evidence, the confidence updating algorithm based on Bayesian backward propagation is started, the upstream parent nodes are traversed in reverse along the directed edges in the dynamic causal diagram, the conditional probability table on each path is recursively derived, the posterior probability increment of each upstream variable to the current abnormal event is calculated, and the backward propagation support factor containing the path weight and time delay attenuation factor is generated, reflecting the influence intensity difference of different preconditions; In the local belief state factor affected by the evidence, the confidence updating algorithm based on Bayesian backward propagation (parameters include dynamic causal diagram structure, initial joint probability distribution, and conditional probability table set) is used to realize the reverse probability calculation of the abnormal event influence path; Further, through the reverse traversal mechanism of the directed edge (the parameter is the path set from the evidence node to the upstream parent node), the posterior probability value is recursively derived along each path according to the parent node conditional probability table. Further, the conditional probability update formula is adopted: wherein is the updated posterior probability, is the prior probability of the parent node under the evidence state, is the conditional probability value of the node, is the normalization coefficient; Further, the multiplication accumulation mechanism is adopted in combination with the path weight parameter and the time lag decay factor wherein is the path support degree, is the causal edge weight, is the time lag decay coefficient, is the time lag step number of the edge, and the weighted adjustment of the path support degree is realized to reflect the real physical impact strength; Further, the reverse propagation support degree factor of each upstream variable is generated by the path support degree accumulation and the node posterior probability increment fusion, and the influence difference of different upstream factors on the abnormal event is quantified as a comparable numerical index; Through the Bayesian reverse propagation algorithm, the local belief state factor of the previous step is converted into the reverse propagation support degree data containing the path weight and the time lag decay, and the accurate description of the upstream variable impact strength is realized; For example, for an abnormal signal node A in a dynamic causal graph model, the input conditional probability table is The value is , the prior probability of the parent node B is , the path weight is , the time lag decay coefficient is , and the time lag step is . After recursive derivation, the posterior probability of B is calculated as = , the path support degree is = , and the reverse propagation support degree is formed by the fusion of the two In this scenario, similar indexes are calculated for different upstream variables, and the system can significantly improve the attribution calculation accuracy and enhance the true reflection of the time lag effect; S4.4: The reverse propagation support degree factor is normalized and weighted fusion processed, the historical stability score and the working condition consistency index of each causal edge are combined, the contribution degree distribution strategy is optimized, the interference signal caused by unstable or low confidence causal connection is suppressed, the comprehensive attribution contribution degree of each upstream process parameter variable is calculated, and the hierarchical attribution score vector factor is generated as the basic input for candidate root cause screening, which is sorted according to the contribution strength; S4.5: Based on the hierarchical attribution score vector, set a contribution threshold, filter out a high-contribution upstream variable set higher than the threshold, mark its variable name, process stage, and maximum impact path, and generate a high-contribution candidate root cause list for the next stage of virtual intervention simulator to call to perform do-operation verification, complete the logical connection from statistical attribution to causal verification.

[0022] The step S5: based on the attribution score vector, filter out high-contribution candidate root cause variables, call the virtual intervention simulator to perform do-operation intervention simulation on the candidate root cause variables, i.e. set their values to the reference state within the normal process range, observe and record the change trend of the posterior probability of downstream risk variables, and form an intervention response curve set. Specifically, it includes: S5.1: Based on the hierarchical attribution score vector, set an attribution contribution threshold, filter out upstream process parameter variables that exceed the threshold, and generate a high-contribution candidate root cause variable set as the operation object of virtual intervention simulation to reduce the root cause verification space and improve reasoning efficiency; Based on the hierarchical attribution score vector output by step S4.5 as input conditions, a threshold setting method (parameter: contribution threshold γ, determined according to the ROC curve analysis of the historical attribution verification set to determine the optimal value interval) is used to realize the preliminary screening of each upstream process parameter variable in the score vector; Further, by using a sorting index construction algorithm (parameters: score vector dimension n, ascending flag false), the index mapping of each variable according to the attribution contribution from high to low is realized, and the association table of the contribution sequence and the corresponding variable identification is obtained; Further, an interval clipping method (parameters: threshold γ, variable contribution table) is used to realize the deletion of variable items with a contribution lower than γ, and to retain a high-contribution variable set higher than the threshold, and to generate process stage labels and maximum impact path reference data of the set elements; Further, based on the process condition label matching method (parameters: variable set, dynamic causal graph process context), the process condition consistency filtering of the high-contribution variable set is realized, the items that do not match the current process condition label and the historical process condition label of the variable are removed, and the process condition adapted candidate root cause set is formed; Through the set normalization processing method (parameters: set size m, variable contribution weight), the candidate set of the previous step is converted into an operation object list that can be directly accepted by the virtual intervention simulator, realizing the accurate reference and operation mapping of the candidate root cause in the subsequent intervention verification; Exemplarily, in a six-layer PCB manufacturability intelligent analysis process, the input hierarchical attribution score vector contains 15 upstream process parameter variables, with a score range of 0.12 to 0.87. Using the threshold setting method, the contribution threshold γ is determined to be 0.65 according to the ROC curve analysis of the historical validation set; the sorting index construction algorithm sorts the variables in descending order, and the first variable "drilling density region B" is 0.87, the variable "copper thickness fluctuation 3 region" is 0.82, and the variable "impedance deviation layer 4" is 0.78. The interval clipping method deletes 9 variables with a contribution less than 0.65, and retains a set of 6 high-contribution variables. The working condition label matching method detects that the current production environment is "laminated material FR4, board thickness 1.6mm, mass production batch 500", and eliminates 2 variables that do not match this working condition, finally forming 4 candidate root cause sets, including "drilling density region B", "copper thickness fluctuation 3 region", "impedance deviation layer 4" and "pressing temperature fluctuation machine 7". Using the set normalization processing method, the weight vector of the set size m=4 is standardized to In order to accurately select nodes and perform do-operations by the virtual intervention simulator. The execution effect shows that the set significantly improves the efficiency and accuracy of root cause verification in subsequent intervention simulation steps, ensuring the credibility of the subsequent intervention response curve generation; S5.2: Construct a configurable virtual intervention simulator based on the structure and parameterized edge weight of the dynamic causal graph model, which supports do-operation intervention on any node; use the Bayesian network inference engine to realize the conditional probability recalculation mechanism, when a node is forced to be a reference state, the system automatically updates the probability distribution of its child nodes, and outputs the risk variable posterior probability sequence after intervention; After receiving the high-contribution candidate root cause variable set from the output of S5.1, use the node set and weighted directed edge parameters of the dynamic causal graph model as system initialization input conditions to construct a configurable virtual intervention simulator structure framework; Using a graph traversal-based node selection algorithm (parameter settings: intervenable node set = high-contribution candidate set, graph structure reference version = latest robust dynamic causal graph version), the positioning and attribute reading of any target node are realized, ensuring that the simulator can accurately access the target variable and load its conditional probability table parameters; Further, through the Bayesian network inference engine initialization algorithm (parameter settings: inference mode = conditional probability recalculation, evidence locking strategy = do-operation), the model structure is imported into the engine core, and the conditional probability matrix between each node and its parent node is preloaded, realizing the computability of the global joint probability distribution; Further, through the conditional probability recalculation formula, the intervention operation is realized, when a node j is forced to be a normal process reference state At this time, the engine resets the conditional probability value of the node under the premise that the parent node set remains unchanged

[0023] and performs recursive recalculation on the conditional probability table of all child nodes k:

[0024] wherein, is the parent node set of node k, is the conditional probability of node k before intervention, is the conditional probability of node k after intervention; Further, the recalculation result is adjusted by applying a decay factor adjustment through the edge weight and time lag parameter in the graph structure, and an exponential time lag decay model (parameter τ derived from the causal edge label) is used to realize the fidelity of the time sequence characteristics of the probability update of the downstream nodes; Through the above conditional probability recalculation chain, the posterior probability sequence of the risk variable after intervention is output, which carries the time index and retains the context of the intervention parameter, and realizes the forward probability propagation driven in S5.3; For example, in the PCB manufacturing process analysis scene, the high-contribution candidate root variable is selected as the impedance control parameter node, and the reference state of the normal process range is set as the impedance value 50Ω, and its parent nodes are the trace width and the dielectric layer distance. The node selection algorithm is used to locate the node index number = ID_23, and the conditional probability table is loaded, and the joint probability distribution of the impedance value and the parent node combination in the table is a 32×32 matrix. Through the do-operation of the Bayesian network inference engine, the impedance value is locked to 50Ω, and the observation conditions of the parent nodes are kept unchanged. The engine recursively updates the copper thickness distribution, AOI detection alarm code and other node probabilities, recalculates the conditional probability table of each node according to the formula chain, and adjusts the exponential decay factor of the time lag parameter τ = 2 batches, generates an array of 10 AOI alarm probability sequences after intervention, with a time unit of production batch, and the sequence shows that the alarm probability is significantly reduced after the 3rd batch. The posterior probability sequence of the risk variable after intervention is output as a structured array, which is transmitted to S5.3 to perform single-variable intervention risk evolution trajectory calculation, and realizes the input preparation of intervention effectiveness verification; S5.3: For each variable in the set of high-contribution candidate root variables, call the virtual intervention simulator to perform do-operation intervention simulation, lock its value in the causal graph to the reference state within the normal process range, keep other observation conditions unchanged, start the forward probability propagation algorithm, calculate the posterior probability change process of the downstream risk variable under the intervention condition layer by layer, and generate the risk evolution trajectory under single-variable intervention; For each variable in the high-contribution-degree candidate root cause variable set, call the virtual intervention simulator to perform a node state locking operation (parameter: reference state value taken from the historical mean of the normal process range) to implement the value setting function of the node in the causal graph; Further, by the conditional probability recalculation method (parameter: using the Bayesian network inference engine, updating order is expanding from the intervened node to its child nodes layer by layer), the posterior probability distribution of each child node is recalculated under the constraint of keeping other observation conditions unchanged, and the local probability update result under the intervention condition is obtained; Further, the forward probability propagation algorithm (parameter: layer sequence traversal according to the topological sorting result of the causal graph, edge weight combined with time delay decay factor adjustment) is used to propagate the influence to the more downstream risk variable nodes based on the updated local probability, calculate the risk variable posterior probability value after each step of propagation, and form a layer-by-layer probability evolution data set; Further, through the recursive traversal and caching mechanism (parameter: cache each layer propagation result to avoid repeated calculation), after completing the forward propagation to the risk variable node, the posterior probability values at consecutive time points are arranged in chronological order, and finally the single-variable risk evolution trajectory is generated during the intervention period; By combining node locking and conditional probability recalculation, the action process of the candidate root cause variable selected in the previous step is converted into risk probability time series data, and the intervention effect is described based on single variable; For example, in a certain PCB production environment, the drilling density sequence parameter is selected as the high-contribution-degree candidate root cause variable, and the normalized reference state value is set to 10 through-holes, 3 blind holes, and 1 buried hole per unit area. Call the virtual intervention simulator to lock the value of this variable node to the reference state, and use the conditional probability recalculation mechanism of the Bayesian network to update the posterior probability distribution of the lamination quality node and the impedance control node directly related to the drilling density. The topological sorting of the causal graph is drilling density → lamination quality → impedance control → risk variable, and the forward probability propagation algorithm is used to update each node in turn. During the propagation process, the edge weight is multiplied by the time delay adjustment coefficient to reflect the delay effect of the physical process. Finally, the posterior probability sequence of the risk variable under the intervention condition is obtained , which is arranged in chronological order to form a complete risk evolution trajectory. In subsequent analysis, this trajectory serves as an important input for evaluating the effectiveness of the intervention, and it is also verified that the driving effect of high drilling density on downstream manufacturing risk is significantly reduced, significantly improving the reliability of causal attribution; S5.4: Time alignment and normalization are performed on the posterior probability sequence of the risk variable generated by each intervention simulation, the risk decay rate feature is extracted based on sliding window difference analysis, and the preset significance criterion (such as relative decline amplitude ≥ 30% and p < 0.05) is combined to judge whether the intervention causes significant inhibition of downstream risk, forming an intervention effectiveness marker with statistical significance; S5.5: The risk evolution trajectory after intervention corresponding to each candidate root cause variable is combined with its intervention effectiveness marker to generate an intervention response curve set containing variable name, intervention starting time, posterior probability dynamic response curve, and effectiveness determination result, which is used as input basis for root cause confirmation and causal path storage in the next stage.

[0025] The step S6: judging the effectiveness of the candidate root cause according to the intervention response curve set, if a certain candidate variable causes a significant decrease in downstream risk probability after intervention, it is confirmed as the root cause, and the causal path pattern is stored together with the corresponding working condition label in the causal knowledge memory library, which constitutes the result output of this time of attribution tracing. Specifically, it includes: S6.1: Based on the intervention response curve set, the posterior probability sequence of the risk variable corresponding to each candidate root cause variable after performing do-operation is extracted, and the instantaneous change rate of risk probability before and after intervention is calculated using sliding window difference algorithm to generate a dynamic sensitivity index sequence; S6.2: Statistical significance test is performed on the dynamic sensitivity index sequence, Wilcoxon signed rank test method is used to analyze the distribution difference before and after intervention, if the p value is lower than the preset threshold and the effect size is greater than the set standard, it is determined that the candidate variable has significant causal influence ability, forming a significance criterion; S6.3: Based on the significance criterion, combined with the original contribution weight in the attribution score vector, a composite confidence score function is constructed to sort all candidate variables that pass the significance test, and the variable with the highest confidence is selected as the root cause output of this time of abnormal event; S6.4: The confirmed root cause variable, the causal influence path it triggers, the intervention response curve set and the corresponding working condition label are integrated into a structured causal pattern template, a unique identification code is generated using hash coding technology, and it is written into the causal knowledge memory library as a standardized knowledge unit; S6.5: The historical templates in the causal knowledge memory library are subjected to deduplication and clustering maintenance operations, based on the working condition label and path similarity matrix, highly overlapping causal pattern clusters are merged, the storage structure is optimized and the subsequent retrieval matching efficiency is improved, forming a sustainable evolving domain causal knowledge graph; Based on the structured causal pattern template set in the causal knowledge memory base, the hash index mapping mechanism (parameters: template unique identification code, working condition label, path structure code) is used to realize the generation of the preliminary template retrieval list; Further, through the path similarity matrix calculation algorithm (parameters: node set difference, edge weight difference, time lag vector difference), multi-dimensional path structure similarity quantization evaluation between templates is realized, and a path similarity score matrix is obtained; Further, through the working condition label matching filtering algorithm (parameter: label intersection threshold), the path similarity score matrix is filtered, so that only the template pairing set with high working condition label matching rate is retained, and a working condition label filtered template pairing table is generated; Further, through the hierarchical clustering algorithm (parameters: similarity threshold, correlation weight), clustering grouping of the working condition label filtered template pairing table is realized, and the template set with high overlap and path similarity score higher than the set threshold is divided into the same causal pattern cluster; Further, through the repeated pattern elimination algorithm (parameter: hash identification code conflict detection rule), the redundant templates in each causal pattern cluster are processed to remove the redundancy, the representative templates with more complete information and higher path integrity are retained, and the optimized causal pattern cluster set is output; Through the knowledge graph construction algorithm, the optimized causal pattern cluster set is converted into a domain causal knowledge graph with node attributes, edge weights and working condition labels, realizing the sustainable evolution and efficient retrieval of causal knowledge; Exemplary, in the PCB manufacturing risk attribution scenario, 1000 historical templates are input into the causal knowledge memory bank, each template containing root cause variables, a set of weighted directed edges (edge weight range 0.05 to 0.95), time delay parameters (0 to 5 seconds), and a set of working condition labels (including board thickness, material type, equipment model). Using the path similarity matrix calculation algorithm, set the node set difference weight to 0.4, the edge weight difference weight to 0.4, and the time delay vector difference weight to 0.2, a 1000x1000 similarity score matrix is calculated, with a score range of 0 to 1. Through the working condition label matching filtering algorithm, set the label intersection threshold to 80%, filter out the template pairing table with similarity ≥0.85 and label matching degree ≥0.8 from the score matrix, and reduce the number of pairings to 250. Using the hierarchical clustering algorithm, set the similarity threshold to 0.9 and the correlation weight to 0.6, and divide the high similarity templates into 68 causal pattern clusters. Execute the repeated pattern elimination algorithm on the templates in each cluster, and the hash identification code conflict detection finds that 9 repeated templates are eliminated. Finally, the optimized 68 pattern clusters are converted into knowledge graph nodes and edge structures, each cluster node number is between 5 to 15, the edge weight is normalized according to the average value within the cluster, and the working condition label is used as a node additional attribute in the graph, which realizes the significant improvement of the causal knowledge graph in the subsequent matching retrieval, and the retrieval response time is significantly shortened from the original hundreds of milliseconds to less than 20 milliseconds.

[0026] The step S7: In the subsequent risk analysis process, when similar working condition labels or similar abnormal patterns are detected, the matching historical causal path templates are retrieved from the causal knowledge memory bank, and the retrieval results are used as prior guidance information to initialize the structure search space of the current dynamic causal graph model, improving the convergence speed and stability of the attribution reasoning. Specifically includes: S7.1: Based on the structured causal path template set stored in the causal knowledge memory bank, the template contains historical root cause variable names, corresponding weighted directed causal edge sequences, time delay parameters and impact intensity weights of each edge, and associated working condition labels and abnormal pattern feature vectors, an index construction algorithm is used to generate an inverted index table based on the working condition-pattern double key to support fast matching retrieval under high concurrency conditions, and a semantically searchable causal knowledge index is output; S7.2: Encode the real-time working condition label output by the current AI analysis module, and extract the multi-dimensional feature vector of the latest detected manufacturing risk abnormal event, including abnormal variable type, confidence score distribution, time series deviation pattern and hierarchical attribution score vector initial value, convert it to a unified dimension query vector representation based on joint embedding space mapping algorithm, generate a real-time query factor carrying context semantics; S7.3: Based on the real-time query, a hybrid metric strategy is adopted to combine the Jaccard similarity of the working condition label and the cosine similarity of the abnormal pattern vector, and the Top-K most matched historical causal path templates are screened out to form a candidate template set, where each template is attached with a matching score to quantify its applicability, and the historical causal path candidate factors with confidence ranking are output; S7.4: The structure of each historical causal path template in the candidate template set is analyzed, the node set and edge connection relationship of the weighted directed causal graph are extracted, and the common substructure of multiple high-matching-degree templates is combined into an initial causal graph skeleton based on graph fusion rules, which is used as a priori guiding constraint for the structure search of the current dynamic causal graph model, thereby narrowing the search space range and generating a graph structure priori factor with domain experience guidance; S7.5: The graph structure priori factor is injected into the time-series causal discovery engine as the initialization input of the dynamic causal structure learning process, which limits the variable pair combination range of local conditional independence test and provides candidate edge priority ranking for delay mutual information analysis, so that the evolution starting point of the new causal graph is closer to the real physical mechanism, significantly reduces the probability of false connection and accelerates the convergence speed, and outputs a robust dynamic causal structure factor with historical knowledge guidance; Under the condition that the graph structure priori factor is prepared, the time-series causal discovery engine initialization function (parameters: prior node set constraint, prior directed edge set constraint) is used to inject the domain experience guided graph structure skeleton into the starting model of the dynamic causal structure learning process, so as to ensure that the structure search range is limited to variable connection pairs with high credibility and physical mechanism compliance; Further, through the variable pair combination constraint mechanism of local conditional independence test (parameters: candidate variable pair list, exclusion set matrix), the conditional independence test is only performed within the prior skeleton connection relationship or its first-order adjacency relationship, thereby significantly reducing the participation of low-correlation variable pairs without physical driving basis in the test, reducing the computational complexity and improving the statistical efficiency of the test; Further, the delay mutual information analysis priority ranking method (parameters: candidate edge list, edge historical stability score) is used to assign matching score weights to the directed edges entering the time delay analysis link, and the mutual information calculation order is determined based on the weight ranking, so that the edges with high matching and high stability are given priority in direction and time delay identification, thereby improving the robustness of direction inference under noise conditions; Further, based on the priority direction identification result, a dynamic weight adjustment algorithm is used to correct the edge weight initial value (parameters: delay mutual information significance index, stability score), which realizes the initialization of the edge weight consistent with the historical causal path to enhance the convergence speed and physical consistency of the edge weight during subsequent structure evolution; The processing result is input into an incremental update module of the time series causal discovery engine to generate a robust dynamic causal structure factor with historical knowledge guidance characteristics and dynamic adaptation capability, significantly reducing noise interference and misconnection probability. For example, in the diagnosis of PCB impedance stability anomalies, the prior graph structure includes a node set {trace width, dielectric layer thickness, impedance deviation value, copper thickness fluctuation} and a directed edge set {trace width→impedance deviation value (time lag 2), dielectric layer thickness→impedance deviation value (time lag 3), copper thickness fluctuation→impedance deviation value (time lag 4)}, and the stability scores of the prior edges are 0.92, 0.88 and 0.85 respectively. The variable pair restriction of local conditional independence test is within the prior edge and its first-order neighbor combination, for example, only the independence relationship between trace width and dielectric layer thickness under the condition of controlling copper thickness fluctuation is tested. Delayed mutual information analysis is sorted by stability score from high to low, and the mutual information value of the trace width→impedance deviation edge is calculated first, and its significance is evaluated at different time lags τ, and τ is set to 1-5, and the mutual information value is maximum at τ=2 and the significance index is 0.87. The edge weight initial value is adjusted by the weighted average of the delayed mutual information significance index and the stability score, for example, the weighted coefficients are 0.6 and 0.4 respectively, and the calculation formula is:

[0027] wherein, The adjusted edge weight is 0.89. After the above initialization processing, the time series causal discovery engine keeps the edge weight stable at more than 0.88 in the subsequent data stream evolution process, the misconnection edge ratio is significantly reduced, the attribution convergence speed is improved to 1.7 times of the original model, and the fast and accurate causal structure update under complex working conditions is realized.

[0028] The step S8: generating an explainability report based on the attribution tracing result, the report containing the root cause variable name, the causal influence path, the intervention verification result and the applicable working condition range, feeding back to the design end or the manufacturing control system through the man-machine interaction interface, triggering the preventive process adjustment or design rule correction action. Specifically including: S8.1: Obtain the root cause confirmation result and the corresponding causal path pattern from the output of S6, the root cause confirmation result including the verified effective candidate root cause variable name and its node identifier in the dynamic causal graph, based on which the causal influence path segment composed of its upstream input edge and downstream influence edge is extracted, and a causal link sequence factor with time series direction mark is generated; S8.2: The causal link sequence is subjected to semantic enhancement processing, the original parameter name is converted into readable description text with engineering meaning by combining the pre-defined variable label mapping table in the process knowledge base, a set of causal declarative sentences containing subject-relation-object structure is generated based on a natural language template engine, and a risk attribution explanation paragraph with domain semantic expression capability is output; S8.3: The intervention response curve set generated from S5 is integrated, trend feature extraction is performed on each intervention response curve, the risk posterior probability decline inflection point and time lag interval are identified by using piecewise linear fitting algorithm, the risk probability difference before and after intervention is calculated as an effectiveness quantitative indicator, and an intervention verification result summary table containing change amplitude, response time delay and confidence level is generated; S8.4: Based on the matched historical working condition label in S7 and the production environment conditions where the current detected abnormality is located, a suitable working condition range description box is constructed, the description box contains constraint dimensions such as laminated material type, plate thickness specification, mass production batch size and equipment model combination, and the effective boundary of this attribution conclusion is determined by using set intersection operation, and a working condition coverage domain declaration labeled with applicable scene restriction is output; S8.5: The risk attribution explanation paragraph, the intervention verification result summary table and the working condition coverage domain declaration are integrated, structured layout is performed according to a preset report template, an interpretable report document containing a title area, a causal graph visualization area, an evidence support area and a suggestion measure area is generated, and the document is pushed to a PCB design end EDA tool plug-in or a manufacturing execution system MES through a man-machine interaction interface, triggering a design rule check DRC update prompt or a process parameter optimization instruction issuing action.

[0029] The application also provides an artificial intelligence-based PCB manufacturability intelligent analysis and early warning system, which uses the above-mentioned artificial intelligence-based PCB manufacturability intelligent analysis and early warning method to intelligently analyze and early warn the PCB manufacturability.

[0030] So far, the technical solutions of the application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the application, and the technical solutions after the changes or replacements will fall within the protection scope of the application.

[0031] The above description is only the preferred embodiments of the application and is not used to limit the application; for those skilled in the art, the application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the application shall be included in the protection scope of the application.

Claims

1. A method for intelligent analysis and early warning of PCB manufacturability based on artificial intelligence, characterized in that, Includes the following steps: S1: Acquire multi-source timing process data during the PCB design and manufacturing process, and attach operating condition labels to each type of data to identify its corresponding production environment conditions. S2: Perform sliding window segmentation and standardization on the multi-source time-series process data, and based on local conditional independence test and delayed mutual information analysis, identify potential causal connections between variables in each time window to generate an initial causal edge set. S3: Based on the initial set of causal edges, construct a dynamic causal graph model, evaluate the consistency of each causal edge under different operating conditions, filter out pseudo-causal connections, and output a robustly updated dynamic causal structure. S4: When a manufacturing risk anomaly is detected, the observed value of the anomaly variable is injected as an evidence node into the dynamic causal graph model, the confidence update algorithm is started, the anomaly support is passed backward along the causal edge, the attribution contribution of each upstream variable to the anomaly event is calculated, and a hierarchical attribution score vector is generated. S5: Based on the attribution score vector, high-contribution candidate root variables are selected, and do-operation intervention simulation is performed on the candidate root variables to form a set of intervention response curves; S6: Determine the validity of the candidate root cause based on the set of intervention response curves. If a candidate variable leads to a significant decrease in the probability of downstream risks after intervention, it is confirmed as the root cause. The causal path pattern, along with the corresponding working condition label, is stored in the causal knowledge memory bank to form the output of this attribution tracing.

2. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 1, characterized in that, Following step S6, the following is also included: S7: In the subsequent risk analysis process, when similar working condition labels or similar abnormal patterns are detected, the matching historical causal path templates are retrieved from the causal knowledge memory, and the retrieval results are used as prior guidance information to initialize the structural search space of the current dynamic causal graph model. S8: Generate an interpretable report based on the attribution traceability results, and feed it back to the design end or manufacturing control system through the human-machine interface to trigger preventive process adjustments or design rule correction actions.

3. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 1, characterized in that, Step S1 specifically includes: Obtain line stack-up parameter data from the PCB design system, perform automatic extraction based on the design file parsing interface, and generate a standardized stack-up structure description table; Collect borehole density sequence data, use the CAM process decomposition module to perform coordinate mapping and region division, calculate the local borehole density moving average, and output the borehole load distribution factor; The copper thickness distribution time series data was collected and continuously monitored during the manufacturing process. Outlier removal and trend smoothing were performed, and a copper thickness fluctuation spectrum was generated based on the moving average filtering algorithm. Obtain the impedance control parameter sequence, extract the effective dielectric constant time series curve based on the transmission line model, mark the deviation range in combination with the target impedance tolerance range, and generate impedance compliance criteria. It integrates equipment status log data from different process nodes, uses a unified log protocol for timestamp alignment and semantic annotation, adds the operating condition label corresponding to the current production batch, and outputs a multi-source heterogeneous process data stream with environmental context identifiers.

4. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 3, characterized in that, The line stacking parameter data includes the thickness, dielectric constant, copper foil type, and stacking order of each signal layer and dielectric layer.

5. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 1, characterized in that, Step S2 specifically includes: The sliding window segmentation process is applied to the multi-source timing process data streams from the PCB design and manufacturing stages. Based on the preset time window length and step size parameters, the continuous timing data is divided into overlapping time segment sequences to generate segmented data block sets. The segmented data block set is subjected to zero-mean standardization and range normalization. Based on the statistical characteristics within the sliding window of each process parameter variable, the standardized local time-series segment data matrix is ​​output. Based on the standardized local time-series data matrix, a local conditional independence test algorithm is executed to detect whether there is a significant direct dependency between any two process parameter variables, while controlling for the influence of other covariates, and to generate a preliminary set of undirected dependent edges. For each pair of variables in the undirected dependent edge set factor, the delayed mutual information analysis method is applied to calculate the mutual information value under different time lags. Based on the time lag corresponding to the maximum mutual information, the direction and time lag parameters of the causal effect are determined, and a directed causal candidate edge set factor with direction label and time lag label is generated. The directed causal candidate edge set factors are filtered by a saliency threshold and verified by a minimum description length criterion, and the final initial causal edge set factors with direction and time delay labels are output.

6. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 1, characterized in that, Step S3 specifically includes: Based on the initial causal edge set with direction and time delay labels in the output, a node set of the dynamic causal graph model is defined, where each node corresponds to a process parameter variable in multi-source time-series process data. A unique identifier is generated using a variable namespace mapping mechanism to form a standardized node registry causal model. For each causal edge in the initial causal edge set, the influence intensity quantification process is performed. The edge weight value is calculated based on the Granger causal intensity coefficient and the standardized regression weight fusion algorithm, and the weighted directed edge set is output. A sliding history window comparison mechanism is constructed. The sliding history window maintains the existence records of causal edges under N similar working conditions in batches. Based on the working condition label matching strategy, the frequency of occurrence of the current causal edge in different windows is statistically analyzed. When the existence consistency of a certain causal edge is lower than the preset confidence threshold, it is determined to be caused by accidental correlation or noise interference, and a pseudo causal screening marker is generated. A pseudo-causal filtering operation is performed on the weighted directed edge set. Low-consistency causal edges are removed based on the pseudo-causal screening flag, while high-frequency and stable causal connections are retained. The weights of the remaining edges are then normalized by variance weighting, and a clean and updated dynamic causal graph structure is output.

7. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 6, characterized in that, The Granger causality strength coefficient is determined by the residual ratio of the multivariate vector autoregressive model, and the standardized regression weights are corrected through a cross-time window parameter stability test.

8. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 1, characterized in that, Step S4 specifically includes: Based on the manufacturing risk anomaly event identifiers output by the AI ​​analysis module, the names of the associated anomaly variables and their observation sequence within the current time window are obtained. The anomaly detection result parsing interface is used to extract the exceedance type, occurrence time, and spatial location information of the anomaly variables, and anomaly evidence factors are generated. The anomalous evidence factors are used as observation evidence nodes and injected into the corresponding variable nodes in the constructed evolvable dynamic causal graph model. Evidence fixing operations are performed to form conditional probability constraints, update the joint probability distribution of the entire graph network, and generate local belief state factors. Based on the local belief state factor, a confidence update algorithm based on Bayesian backpropagation is initiated. The upstream parent nodes are traversed backward along the directed edges in the dynamic causal graph. The conditional probability table on each path is recursively derived, and the posterior probability increment of each upstream variable for the current abnormal event is calculated to generate the backpropagation support factor. The backpropagation support factor is normalized and weighted, and combined with the historical stability score and operating condition consistency index of each causal edge, the comprehensive attribution contribution of each upstream process parameter variable is calculated to generate a hierarchical attribution score vector factor. Based on the hierarchical attribution scoring vector, a contribution threshold is set, and a set of upstream variables with high contribution above the threshold is selected. The variable name, the process stage to which they belong, and the path with the greatest influence are marked, and a list of candidate root causes with high contribution is generated.

9. The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence according to claim 1, characterized in that, The intervention response curve set includes variable names, intervention start time, posterior probability dynamic response curves, and effectiveness determination results.

10. An AI-based intelligent analysis and early warning system for PCB manufacturability, characterized in that: The PCB manufacturability intelligent analysis and early warning method based on artificial intelligence as described in any one of claims 1-9 is used to perform intelligent analysis and early warning of PCB manufacturability.

Images