Circuit board production whole-process quality regulation system

Abstract

The application discloses a line board production whole-process quality regulation system and belongs to the technical field of line board quality control, comprising a data acquisition module, a cause-effect knowledge graph construction module, a cause-effect reasoning and prediction module, a cross-process feedforward compensation module and a self-evolution learning module.The application breaks through the limitations of fragmentation and passive response of the prior art in quality control, realizes a leap change from post-detection to pre-prevention and from single-point control to global coordination through a complete closed loop of data acquisition, cause-effect graph construction, reasoning and prediction, feedforward compensation and self-evolution learning, and realizes synchronous evolution of the cause-effect knowledge graph and the compensation strategy through a self-evolution learning mechanism driven by reinforcement learning, thereby forming a mutually enhanced positive cycle, so that the regulation precision and self-adaptive ability of the system are continuously improved with the extension of the running time, and the system is significantly superior to the static system in the prior art which needs artificial regular maintenance and parameter adjustment.

Description

Technical Field

[0001] This invention relates to the field of circuit board quality control technology, and in particular to a quality control system for the entire circuit board production process. Background Technology

[0002] As the support structure and electrical interconnection carrier for electronic components, the manufacturing quality of printed circuit boards (PCBs) directly determines the reliability and lifespan of electronic products. A typical PCB production line involves dozens of processes, including inner layer fabrication, lamination, drilling, hole metallization, pattern plating, outer layer etching, solder masking, and surface treatment. Each process contains multiple process parameters, and there are complex causal relationships between these processes.

[0003] Existing technologies mostly focus on local optimization of a single process (such as electroplating hole filling and lamination white spot prediction). Although they can improve the quality performance of the process to a certain extent, they cannot solve the problem of causal transmission across processes. The parameter deviations of the upstream process are accumulated and amplified through multiple processes, and often only manifest as defects after several processes. By then, the root cause of the upstream process is difficult to trace.

[0004] To address this issue, a root cause analysis method based on Granger causality testing and structural equation modeling is proposed. This method can extract causal relationships between process parameters from historical data and output a ranking list of influencing factors for process engineers to refer to. However, this type of method stops at the level of "diagnostic suggestions" and relies on manual judgment and experience to implement process adjustments. It cannot form an automated closed loop from "anomaly detection, root cause location, parameter adjustment, and effect feedback". In the high-speed production environment of circuit board manufacturing, the response delay of manual intervention is usually several hours or even days. During this period, a large number of work-in-process products have already flowed to subsequent processes, and batch scrapping is difficult to avoid. The application of knowledge graph technology in the PCB field remains at the stage of static knowledge management. Although some patents have attempted to use knowledge graphs for defect tracing, solidifying the defect-cause relationships from historical experience into a graph structure for maintenance personnel to query and reference, these knowledge graphs are mostly built once and updated periodically, lacking adaptive evolution capabilities driven by real-time data. Currently, there is exploration in the industry of collaborative control methods for multi-process PCB manufacturing, which forms a process status baseline through cross-verification of redundant sensors and locates the defect transmission link through bidirectional verification combining forward tracking and reverse backtracking. However, its core is still the "tracking" and "location" of existing defects, that is, a passive response after a defect occurs, lacking an active cross-process feedforward compensation mechanism based on causal reasoning.

[0005] Based on this, a quality control system for the entire circuit board production process is proposed. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the prior art and to propose a quality control system for the entire circuit board production process.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The PCB manufacturing process quality control system includes a data acquisition module, a causal knowledge graph construction module, a causal reasoning and prediction module, a cross-process feedforward compensation module, and a self-evolutionary learning module. The data acquisition module is configured to collect process parameter data streams, equipment operating status data, environmental variable data, and quality inspection result data of multiple process nodes in the circuit board manufacturing process in real time, and generate a multi-dimensional time series database. The causal knowledge graph construction module is configured to construct a cross-process dynamic causal knowledge graph based on a multi-dimensional time-series database and using a combination of nonlinear Granger causality test and dynamic Bayesian network. The nodes of the dynamic causal knowledge graph include at least process nodes, process parameter entity nodes, and quality index nodes. The edges represent the direction, strength, and delay time of the causal association. The dynamic causal knowledge graph supports adaptive updates of the causal structure triggered by KL divergence detection of newly acquired data and historical data within a sliding time window. The causal reasoning and prediction module is configured to input the real-time data stream of the multi-dimensional time series database into the dynamic causal knowledge graph, use a multi-hop graph attention network to perform multi-hop message passing and attention weight allocation, generate an updated graph structure containing long-range dependencies, and perform quality prediction and anomaly root cause localization based on the updated graph structure, and output reasoning results containing anomaly type, root cause process and predicted quality trend. The cross-process feedforward compensation module is configured to identify the causal path from the upstream process parameter node to the downstream process quality indicator node in the dynamic causal knowledge graph based on the reasoning result. When the upstream process parameter deviates from the preset threshold range, a compensation strategy for the downstream process is generated in advance based on the causal strength weight and path delay time of the causal path, and is automatically sent to the corresponding equipment execution parameter adjustment through the manufacturing execution system interface. The self-evolutionary learning module is configured to track and evaluate the actual quality effect after the compensation strategy is executed, and to perform closed-loop correction and self-evolutionary update of the causal edge weights and compensation strategy of the dynamic causal knowledge graph based on the reinforcement learning mechanism.

[0008] As a preferred embodiment, the construction of the dynamic causal knowledge graph adopts a dual-constraint mechanism that integrates process mechanism knowledge with data-driven discovery of causal relationships: The process mechanism knowledge is derived from failure mode and effects analysis documents and expert rules, and is used to pre-set the prior constraints on the existence of causal edges. The data-driven discovery of causal relationships is extracted from the multidimensional time-series database based on nonlinear Granger causality tests and structural equation modeling. The causal relationships discovered through data-driven discovery are filtered and corrected by the prior constraints, generating a hybrid causal knowledge graph that integrates mechanisms and data.

[0009] As a preferred embodiment, the cross-process feedforward compensation module is further configured as follows: The confidence of each candidate compensation path is weighted and calculated using a path confidence decay function, which uses the product of the path length and the confidence of each causal edge in the path as the decay factor. Based on the weighted calculation results, high-confidence transmission paths are activated first to generate a multi-level compensation strategy set; A multi-objective optimization algorithm is used to select the optimal compensation strategy from the set of multi-level compensation strategies.

[0010] As a preferred embodiment, the causal reasoning and prediction module is further configured as follows: A pre-trained trend semantic encoder is introduced to encode the trend pattern of process parameters within the current time window and generate a trend semantic vector. Based on trend semantic vectors, the candidate anomaly attribution paths output by the multi-hop graph attention network are reordered, and the root cause process list is output in descending order of causal influence.

[0011] As a preferred embodiment, the adaptive update of the causal structure in the causal knowledge graph construction module adopts the following mechanism: Set the sliding time window length and sliding step size, and perform KL divergence detection on newly acquired data and historical data within each sliding window; When the KL divergence exceeds a preset threshold, a nonlinear Granger causality test is triggered on the data within the current sliding window to re-evaluate the strength of the affected causal edges and dynamically adjust the edge weights and structural topology of the causal knowledge graph.

[0012] As a preferred embodiment, the reinforcement learning-based closed-loop correction in the self-evolutionary learning module includes: A deep Q-network is constructed as the policy network. The state space of the deep Q-network is a snapshot vector of the current dynamic causal knowledge graph, and the action space is the range of process parameters that can be adjusted by the cross-process feedforward compensation module. The reward function is constructed based on the absolute value of the deviation between the actual quality result after the compensation strategy is executed and the predicted quality result output by the causal inference and prediction module. The reward function guides the adaptive update of the policy network.

[0013] As a preferred embodiment, the data acquisition module is further configured to: set a unique physical identification code on each circuit board unit, wherein the physical identification code is a QR code, barcode or laser-engraved code; All data in the multidimensional time-series database are associated and stored using the physical identifier as an index, enabling precise binding of full-process data with specific circuit board units and supporting single-board-level quality traceability.

[0014] As a preferred embodiment, the plurality of process nodes include at least drilling process nodes, electroplating process nodes, lamination process nodes, and etching process nodes; The compensation strategy generated by the cross-process feedforward compensation module includes at least one or more of the following combinations: current density compensation value for electroplating process, temperature curve compensation parameter for lamination process, and etching solution concentration compensation value for etching process.

[0015] As a preferred embodiment, the system further includes a digital twin simulation verification module, configured as follows: Before the compensation strategy generated by the cross-process feedforward compensation module is issued and executed through the manufacturing execution system interface, the compensation strategy is input into a pre-built digital twin model for virtual simulation verification. If the simulation verification results meet the preset quality indicators, then execution is allowed. If the conditions are not met, the causal reasoning and prediction module is triggered to re-perform the reasoning, and the cross-process feedforward compensation module regenerates the compensation strategy.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention breaks through the limitations of fragmented and passive response in quality control in existing technologies by establishing a complete closed loop of data acquisition, causal graph construction, inference prediction, feedforward compensation, and self-evolutionary learning. It achieves a leapfrog transformation from post-event detection to pre-event prevention and from single-point control to global collaboration. Based on causal path identification and path credibility assessment in dynamic causal knowledge graph, it generates feedforward compensation strategies for downstream processes and automatically issues them through the MES system, filling the gap in cross-process collaborative control in existing technologies.

[0017] 2. This invention achieves synchronous evolution of causal knowledge graph and compensation strategy through a reinforcement learning-driven self-evolutionary learning mechanism, forming a mutually reinforcing positive cycle. This enables the system's control accuracy and adaptive capability to continuously improve with the extension of running time, which is significantly better than the static systems in the prior art that require regular manual maintenance and parameter adjustment. Through proactive and intelligent quality control, this invention can significantly reduce the defect rate in the circuit board manufacturing process, improve the first-pass yield of products, and reduce the economic losses caused by rework and scrap. Attached Figure Description

[0018] Figure 1 This is a block diagram of the overall architecture of the circuit board production process quality control system proposed in this invention. Figure 2 This is a schematic diagram of the node and edge relationships in the dynamic causal knowledge graph of this invention; Figure 3 This is a flowchart of the adaptive update process for causal structure based on KL divergence detection in this invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] Example, refer to Figures 1 to 3 The circuit board production process quality control system includes a data acquisition module, a causal knowledge graph construction module, a causal reasoning and prediction module, a cross-process feedforward compensation module, and a self-evolutionary learning module. The data acquisition module is configured to collect process parameter data streams, equipment operating status data, environmental variable data, and quality inspection result data from multiple process nodes during the circuit board manufacturing process in real time, and generate a multi-dimensional time-series database. The data acquisition module is further configured to: set a unique physical identification code on each circuit board unit, which can be a QR code, barcode, or laser-engraved code; all data in the multi-dimensional time-series database is stored in association with the physical identification code as an index, realizing the precise binding of the entire process data with the specific circuit board unit, and supporting single-board-level quality traceability.

[0024] Multiple process nodes include at least drilling process nodes, electroplating process nodes, lamination process nodes, and etching process nodes; the compensation strategy generated by the cross-process feedforward compensation module includes at least one or more of the following combinations: current density compensation value for electroplating process, temperature curve compensation parameter for lamination process, and etching solution concentration compensation value for etching process.

[0025] The causal knowledge graph construction module is configured to be based on a multi-dimensional time-series database and adopts a method combining nonlinear Granger causality test and dynamic Bayesian network to construct a dynamic causal knowledge graph across processes. The nodes of the dynamic causal knowledge graph include at least process nodes, process parameter entity nodes and quality index nodes. The edges represent the direction, strength and delay time of causal association. The dynamic causal knowledge graph supports adaptive updates of the causal structure triggered by KL divergence detection of newly collected data and historical data within a sliding time window. The construction of the dynamic causal knowledge graph adopts a dual constraint mechanism that integrates process mechanism knowledge and data-driven discovery of causal relationships: the process mechanism knowledge comes from Failure Mode and Effects Analysis (FMEA) documents and expert rules, which are used to pre-set the prior constraints on the existence of causal edges; the data-driven discovery of causal relationships is extracted from multi-dimensional time series databases based on nonlinear Granger causality tests and structural equation models; the data-driven discovery of causal relationships is filtered and corrected through prior constraints to generate a hybrid causal knowledge graph that integrates mechanisms and data.

[0026] The causal reasoning and prediction module is configured to input real-time data streams from a multi-dimensional time-series database into a dynamic causal knowledge graph, use a multi-hop graph attention network for multi-hop message passing and attention weight allocation, generate an updated graph structure containing long-range dependencies, and perform quality prediction and anomaly root cause localization based on the updated graph structure, outputting reasoning results including anomaly type, root cause process and predicted quality trend. The adaptive update of the causal structure in the causal knowledge graph construction module adopts the following mechanism: set the sliding time window length and sliding step size, and perform KL divergence detection on the newly collected data and historical data in each sliding window; when the KL divergence exceeds the preset threshold, trigger a nonlinear Granger causality test on the data in the current sliding window, re-evaluate the strength of the affected causal edges, and dynamically adjust the edge weights and / or structural topology of the causal knowledge graph.

[0027] The causal reasoning and prediction module is further configured as follows: a pre-trained trend semantic encoder is introduced to encode the trend pattern of process parameters within the current time window to generate a trend semantic vector; based on the trend semantic vector, the candidate anomaly attribution paths output by the multi-hop graph attention network are reordered, and a list of root cause processes arranged in descending order of causal influence is output.

[0028] The cross-process feedforward compensation module is configured to identify the causal path from the upstream process parameter node to the downstream process quality indicator node in the dynamic causal knowledge graph based on the reasoning results. When the upstream process parameter deviates from the preset threshold range, a compensation strategy for the downstream process is generated in advance based on the causal strength weight and path delay time of the causal path, and is automatically sent to the corresponding equipment execution parameter adjustment through the manufacturing execution system interface. The cross-process feedforward compensation module is further configured as follows: the confidence of each candidate compensation path is weighted using a path confidence decay function, with the path confidence decay function using the product of the path length and the confidence of each causal edge in the path as the decay factor; based on the weighted calculation results, high-confidence propagation paths are activated first to generate a multi-level compensation strategy set; and a multi-objective optimization algorithm is used to select the optimal compensation strategy from the multi-level compensation strategy set.

[0029] The self-evolutionary learning module is configured to track and evaluate the actual quality effect after the compensation strategy is implemented. Based on the reinforcement learning mechanism, it performs closed-loop correction and self-evolutionary update of the causal edge weights and compensation strategies of the dynamic causal knowledge graph.

[0030] The closed-loop correction based on reinforcement learning in the self-evolutionary learning module includes: constructing a deep Q-network as the policy network, where the state space of the deep Q-network is a snapshot vector of the current dynamic causal knowledge graph, the action space is the range of process parameters that can be adjusted by the cross-process feedforward compensation module, and the reward function is constructed based on the absolute value of the deviation between the actual quality result after the execution of the compensation policy and the predicted quality result output by the causal inference and prediction module; and guiding the adaptive update of the policy network through the reward function.

[0031] The system also includes a digital twin simulation verification module, configured to: input the compensation strategy into a pre-built digital twin model for virtual simulation verification before the compensation strategy generated by the cross-process feedforward compensation module is sent out for execution through the manufacturing execution system interface; if the simulation verification result meets the preset quality indicators, execution is allowed; if not, the system returns to the causal reasoning and prediction module to regenerate the compensation strategy.

[0032] The system also includes a visualization monitoring and early warning module, configured to generate a monitoring dashboard in real time based on a multi-dimensional time-series database and a dynamic causal knowledge graph. This dashboard includes the real-time status of each process node, the causal propagation path, anomaly warning information, and process optimization suggestions, presented graphically through a human-machine interface. When an anomaly type exceeds a preset risk level, a tiered early warning is triggered. The data acquisition module is configured to collect process parameter data streams, equipment operating status data, environmental variable data, and quality inspection result data from multiple process nodes during circuit board manufacturing in real time, generating a multi-dimensional time-series database.

[0033] Data types and sources collected: A distributed edge acquisition architecture is adopted, with edge acquisition nodes deployed at each key process node to achieve real-time, high-frequency data acquisition, avoiding the bandwidth and latency bottlenecks of a single central server; The collected data includes at least the following four categories: (1) Process parameter data: including but not limited to spindle speed, feed rate, and drop ratio in the drilling process; current density, bath temperature, pH value, and additive concentration in the electroplating process; heating rate, pressure curve, and vacuum degree in the lamination process; and etching solution concentration, temperature, and spray pressure in the etching process. The sampling frequency shall not be less than 1Hz, and high-frequency sampling of 100Hz or higher shall be used for key parameters (such as electroplating current density).

[0034] (2) Equipment status data: including equipment vibration amplitude, running time of key components (such as cumulative number of drill bits used, electroplating anode life), servo motor torque feedback value, etc., which are used to build equipment health assessment model.

[0035] (3) Environmental variable data: including workshop temperature and humidity, cleanliness level (characterized by the number of ≥0.5μm particles per cubic meter of air), used to eliminate the interference of environmental factors on process stability.

[0036] (4) Quality inspection results data: including defect type and location coordinates output by automatic optical inspection (AOI), network continuity / insulation status output by electrical performance testing (flying probe test / special fixture test), characteristic impedance value output by impedance tester, hole wall coating thickness output by X-ray inspection, etc.

[0037] Data association storage mechanism: The data acquisition module is further configured to assign a unique physical identifier to each circuit board unit. This identifier can be a QR code, barcode, or laser-engraved code. By using OCR recognition or a barcode scanner to read the physical identifier, data collected at each process node of each circuit board unit is linked and stored using this physical identifier as an index, forming a complete "one board, one file" data archive that supports single-board-level quality traceability. This single-board-level traceability mechanism allows for precise location of the specific circuit board unit and the parameter status of each process it passed through when a quality anomaly occurs in a batch, providing accurate data support for root cause analysis.

[0038] Data transfer relationships between modules: The multidimensional time-series data collected by the data acquisition module serves two purposes: firstly, as input to the causal knowledge graph construction module for offline or near-online causal structure learning; and secondly, as a real-time input stream for the causal inference and prediction module for online quality prediction and anomaly root cause localization. Data flow between the data acquisition module and the causal knowledge graph construction and prediction modules is asynchronously transmitted using a message queue mechanism (such as Apache Kafka) to ensure no data backlog or loss occurs in high-frequency data acquisition scenarios.

[0039] Technical solution for the causal knowledge graph construction module: The causal knowledge graph construction module is configured based on a multi-dimensional time-series database and employs a combination of nonlinear Granger causality testing and dynamic Bayesian networks to construct a dynamic causal knowledge graph across processes. The nodes of the dynamic causal knowledge graph include at least process nodes, process parameter entity nodes, and quality indicator nodes. Edges represent the direction, strength, and delay time of causal associations. Furthermore, the dynamic causal knowledge graph supports adaptive updates of the causal structure triggered by KL divergence detection between newly acquired data and historical data within a sliding time window.

[0040] In industrial manufacturing scenarios, the relationships between process parameters often exhibit nonlinear characteristics (e.g., the relationship between current density and coating thickness shows nonlinear saturation in the high current region, and the relationship between lamination temperature and resin flow characteristics shows abrupt changes near the glass transition temperature). Traditional linear Granger causality tests cannot effectively identify such nonlinear causal relationships. Therefore, this solution adopts a kernel-based nonlinear Granger causality test.

[0041] Let time series variables and Let these represent the time-series measurements of the two process parameters. Mapping the sequence to a high-dimensional regenerative kernel Hilbert space (RKHS), performing a linear Granger test in this space is equivalent to performing a nonlinear Granger test in the original space. Specifically, define the kernel function... Map the original data to the feature space ,satisfy .

[0042] The core idea of ​​the nonlinear Granger causality test is: if incorporating historical information about X can significantly reduce the prediction error of Y, then X is considered a Granger cause of Y.

[0043] The specific testing model is as follows: Unconstrained model: Wherein is the unknown function defined in RKHS, and p and q are the lag orders of autoregression and cross-regression, respectively (typical values ​​are determined according to the characteristics of the process; for electroplating, p=3~5, q=2~4; for lamination, p=4~6, q=3~5). This is the residual term.

[0044] Constraint Model: The F-statistic is constructed by comparing the sum of squared residuals of the two models: in To constrain the sum of squared residuals of the model, This is the sum of squared residuals of the unconstrained model. If the value of the F-statistic is greater than the critical value at a given significance level (typically α=0.05), then the null hypothesis (X is not a Granger cause of Y) is rejected, and a causal relationship from X to Y is determined to exist.

[0045] Simultaneously, a dynamic Bayesian network is used to model the causal structure across processes, with each process parameter serving as a network node, and conditional probability distribution is applied. Describes the temporal dependencies between nodes, where This indicates the current state of each node. This indicates the node state at the previous time step. It represents external condition variables (such as ambient temperature and humidity, equipment running time, etc.).

[0046] Data structure of dynamic causal knowledge graph: Dynamic causal knowledge graphs utilize graph databases (such as...) ) is stored, where: Node types include process nodes (such as drilling, electroplating, lamination, and etching), process parameter entity nodes (such as current density, bath temperature, lamination pressure, and etching solution concentration), and quality indicator nodes (such as hole copper thickness, impedance deviation, and interlayer bonding strength). Each node stores its unique identifier, node type label, and a sliding window cache of historical feature vectors.

[0047] Edge properties include: causal direction (indicated by an arrow pointing from cause to effect); causal strength. (Values ​​range from 0 to 1, obtained by normalizing the F-statistic of the nonlinear Granger test); Delay time (This represents the time interval from a change in the causal parameter to a change in the result parameter, expressed in seconds or steps, determined by the location of the maximum peak in cross-correlation analysis); Confidence level (The reliability is calculated based on historical sample size and statistical significance, with a value range of 0-1).

[0048] A dual-constraint construction mechanism that integrates FMEA mechanistic knowledge with data-driven approaches: To further improve the reliability of causal knowledge graphs, this scheme introduces a dual-constraint mechanism that integrates process mechanism knowledge with data-driven discovery of causal relationships.

[0049] Mechanistic knowledge constraints: These are derived from Failure Mode and Effects Analysis (FMEA) documents and expert rules containing causal chain reasoning rules, used to pre-define the existence of causal edges. For example, according to electrochemical principles, current density is a direct cause of coating thickness, while the effect of workshop temperature on coating thickness is indirectly transmitted through bath temperature; according to polymer science principles, lamination temperature directly affects the melt viscosity of the prepreg, thus affecting resin flow and interlayer adhesion. This prior knowledge can constrain the direction of causal discovery and avoid false causal associations.

[0050] The severity level (usually 1-10) recorded in the FMEA document is used to assign initial confidence to the preset causal edges. The causal edges corresponding to the failure modes with higher severity are given higher initial confidence weights, ensuring that the system prioritizes the causal paths that have the greatest impact on product quality.

[0051] Data-driven discovery: Based on the aforementioned nonlinear Granger causality test, causal relationships extracted from a multidimensional time-series database are used to generate initial causal edges and their strength scores.

[0052] Integration method: The causal relationships discovered by data-driven methods are screened and corrected by using mechanistic knowledge, retaining only the causal edges that conform to the physicochemical mechanism, or the data-driven results are used as the main approach in areas not covered by mechanistic knowledge.

[0053] Specifically, the integration process includes three steps: The first step is to generate a "whitelist" (allowed causal edges) and a "blacklist" (prohibited causal edges) of candidate causal edges based on the failure mode analysis in the FMEA document. The second step is to match the causal edges discovered by data-driven discovery with the blacklist and whitelist, and filter out the causal edges in the blacklist. The third step is to assign a higher overall confidence level to causal edges that appear in both the whitelist and the data-driven results; and to reduce the confidence level and mark causal edges that appear only in the data-driven results but not in the whitelist as "pending verification".

[0054] Causal structure adaptive update mechanism based on KL divergence detection: A key innovation of the causal knowledge graph construction module in this solution is that it supports adaptive updates of the causal structure triggered by KL divergence detection of newly collected data and historical data within a sliding time window. This mechanism enables the causal knowledge graph to automatically adjust its structure and weights according to changes in real-time data, which is different from the static knowledge graph update schemes in existing technologies.

[0055] Let the historical data distribution be... The newly collected sliding window data is distributed as follows The KL divergence between two distributions is defined as: In practical engineering implementation, a discretization approximation is used: in Let be the number of samples falling into the i-th interval within the new sliding window. This represents the total number of samples in the new window. Let be the number of samples falling into the i-th interval in the historical data. The total number of historical data samples is denoted as n. The interval division adopts the equal frequency binning method, and the typical value of the number of bins n is 10 to 20.

[0056] Set the length of the sliding window (Typical values ​​are 500–2000 sample points, corresponding to approximately 1–4 hours of production data) and sliding step size (Typical value) At the end of each sliding window, perform the following update process: Calculate the KL divergence between the new window data and the historical data. .

[0057] like The preset threshold, typically ranging from 0.05 to 0.15 (the specific value can be adjusted based on the sensitivity of the process), triggers a re-evaluation of the causal structure.

[0058] Re-perform the nonlinear Granger causality test on the data within the current sliding window to re-evaluate the strength of the affected causal edges.

[0059] Based on the reassessment results, the edge weights of the causal knowledge graph are dynamically adjusted. The specific update rules are as follows: Where α is the update rate (typically ranging from 0.3 to 0.7). The causal strength is recalculated based on the current window data. For the updated weights, The weight is the weight from the previous time step.

[0060] If a new causal relationship is discovered (i.e., an edge that did not exist before, but the current window data shows a significant causal relationship), then an edge is added, with an initial weight set to... If the original causal edge is no longer significant in the current window of data (F statistic is below the threshold), then the weight of the edge is multiplied by the decay factor β (β < 1, typically 0.5), and the edge is deleted if it is not significant in multiple consecutive windows.

[0061] The physical significance of this mechanism is that when equipment ages, material batches are changed, or the ambient temperature and humidity change significantly, the causal relationship pattern between process parameters may drift. KL divergence detection can detect this drift in time and trigger the adaptive update of the causal structure, ensuring that the causal knowledge graph is always synchronized with the current production conditions.

[0062] Data transfer relationships between modules: The dynamic causal knowledge graph output by the causal knowledge graph construction module serves as the core knowledge base, simultaneously supporting the causal reasoning and prediction module (the inference engine for quality prediction and root cause localization) and the cross-process feedforward compensation module (for causal path identification and compensation strategy generation). When the causal knowledge graph undergoes adaptive updates, the updated graph structure is automatically synchronized to the causal reasoning and prediction module and the cross-process feedforward compensation module, ensuring that subsequent inference and compensation decisions are based on the latest causal knowledge.

[0063] Technical solution for causal reasoning and prediction module: The causal reasoning and prediction module is configured to input real-time data streams from a multi-dimensional time-series database into a dynamic causal knowledge graph. It employs a multi-hop graph attention network (Multi-hopGAT) for multi-hop message passing and attention weight allocation, generating an updated graph structure containing long-range dependencies. Based on this updated graph structure, it performs quality prediction and anomaly root cause localization, outputting inference results that include anomaly type, root cause process, and predicted quality trend.

[0064] Mathematical architecture of multi-hop graph attention networks: This solution introduces a multi-hop propagation mechanism based on graph attention networks to effectively model long-range causal dependencies across processes. In PCB manufacturing, parameter deviations in upstream processes may be transmitted through multiple processes before ultimately affecting product quality—for example, drill bit wear in the drilling process, after passing through multiple processes such as desmearing, hole metallization, and pattern electroplating, ultimately manifests as insufficient copper thickness in the holes.

[0065] Let G = (V, E) be a causal knowledge graph, where V is the set of nodes and E is the set of edges. For node i, its l-hop neighborhood is defined as... , where d(j, i) represents the shortest path length from node j to node i.

[0066] For node i, the l-th hop feature update at the h-th attention head, the attention coefficients... The calculation method is as follows: in The feature representation of node i in the (l-1) hop layer (node ​​features include real-time values ​​of current process parameters, equipment status features, and statistical features within the most recent time window). The weight matrix is ​​a learnable matrix. Let be the attention vector, || denotes the vector concatenation operation, and LeakyReLU is a non-linear activation function (with a negative slope of 0.2).

[0067] The fusion feature of node i in the l-th hop layer is represented as: Where H is the number of attention heads (typically 4 to 8). For nonlinear activation functions (such as ReLU or ELU), residual connections The introduction of this method effectively alleviates the gradient vanishing problem in multi-hop propagation.

[0068] Through L-hop message passing (L is determined by the length of the process route; for typical multilayer PCB manufacturing processes, L is 3 to 5 hops), each node can perceive the state information of other nodes within a L-hop distance, thereby enabling the modeling of long-range causal dependencies across processes. For example, in the process chain of drilling → desmearing → hole metallization → pattern plating → outer layer etching, the drill wear signal in the drilling process can be perceived by the final plating quality node after propagating through 5 hops, allowing the system to predict its impact on the final plating quality in the early stages of drill wear.

[0069] Trend Semantic Encoder and Root Cause Reordering: To further improve the accuracy of root cause localization, this scheme introduces a pre-trained trend semantic encoder to encode the trend patterns of process parameters within the current time window. Suppose a certain process parameter within the time window... The sequence of measured values ​​within is The trend semantic encoder encodes it into a trend semantic vector. : in For a time-series encoder based on the Transformer architecture, it is pre-trained on historical defect data to learn how to map parameter trend patterns to the semantic space. Employing a standard Transformer encoder architecture, comprising 6 encoder layers and 8 attention heads, with a feedforward network dimension of 2048, the algorithm is trained end-to-end on historical defect data using cross-entropy loss. The trend semantic encoder can identify different trend patterns such as "slow rise," "step jump," "periodic fluctuation," and "high-frequency noise," and perform pattern matching with historical defect cases.

[0070] Finally, the candidate anomaly attribution paths output by the multi-hop graph attention network are reordered using a trend semantic encoder, resulting in a root cause process list sorted in descending order of causal influence. in This refers to the comprehensive attention weights from node i to the target quality node j, calculated using an L-jump graph attention network. The trend semantic similarity function uses cosine similarity, where the closer the cosine value is to 1, the more similar the trend patterns are.

[0071] Components of the output The output of the causal reasoning and prediction module includes information in at least three dimensions: Anomaly Types: Based on the deviation between quality inspection data and predicted values, defect types are identified. Typical defect types include, but are not limited to: no copper inside the hole (referring to an open circuit inside the metallized hole, which loses electrical connection performance during continuity testing and is one of the fatal quality defects of PCB), thin copper in the hole (insufficient copper layer thickness on the hole wall, affecting conductivity and mechanical strength), lamination white spots, open circuit (broken circuit), and short circuit (unexpected connection between circuits).

[0072] Root Cause Processes: A list of root cause processes arranged in descending order of causal influence, with each process's contribution probability and confidence level indicated. Each root cause entry in the list includes the process name, causal influence score (0-1 normalized), confidence interval (95% confidence interval), and a simplified causal path diagram from that process to the abnormal quality indicator.

[0073] Predicting Quality Trends: Based on the current process parameters and causal knowledge graph, predict the changing trends of various quality indicators over the next 3-5 time steps. The prediction results are presented as time-series curves, with confidence intervals for the predicted values ​​clearly marked, enabling early warning of quality risks.

[0074] Data transfer relationships between modules: The causal reasoning and prediction module receives real-time data streams from the data acquisition module and dynamic causal knowledge graphs from the causal knowledge graph construction module as inputs, and outputs reasoning results. These results are simultaneously transmitted to two downstream modules: firstly, to the cross-process feedforward compensation module as the decision-making basis for compensation strategy generation; and secondly, to the self-evolutionary learning module as a reference benchmark for the prediction quality results in the reward function.

[0075] Technical solution for cross-process feedforward compensation module: The cross-process feedforward compensation module is one of the core innovations of this solution. This module is configured to identify causal paths from upstream process parameter nodes to downstream process quality indicator nodes in a dynamic causal knowledge graph based on the inference results. When the upstream process parameters deviate from the preset threshold range, a compensation strategy for the downstream process is generated in advance based on the causal strength weight and path delay time of the causal path, and is automatically sent to the corresponding equipment to adjust the execution parameters through the Manufacturing Execution System (MES) interface.

[0076] Traditional feedback control strategies adjust only after quality indicators deviate from target values. This "post-hoc compensation" is often too late in PCB manufacturing—by the time defects are detected in downstream processes, the root cause in the upstream processes has already disappeared, and a large number of work-in-process products have been affected. Feedforward compensation, on the other hand, operates on a completely different principle: it is based on causal relationships in a causal knowledge graph. When upstream process parameters deviate, it compensates for downstream process parameters before actual deviations occur, thus nipping defects in the bud.

[0077] Causal path identification and path credibility evaluation: Let the parameters of the currently detected upstream process node i be... Deviating from its preset threshold range Deviation The system identifies all causal paths originating from node i and ultimately pointing to quality indicator node j from a dynamic causal knowledge graph. .

[0078] For each causal path Define path credibility for: in Let e ​​be the causal strength of the causal edge e (value range 0-1). λ is the path length (i.e., the number of steps from the source node to the target node), and λ is the attenuation coefficient (reflecting the rate attenuation of causal effect as the path length increases, typically ranging from 0.3 to 0.7, obtained by fitting the causal effect attenuation curve of historical data).

[0079] Path credibility decay function The introduction of this concept has a clear physical significance: In the PCB manufacturing process, the parameter deviations of upstream processes are often diluted or corrected by the control effect of intermediate processes after being transmitted through multiple processes. The longer the path, the greater the uncertainty. Therefore, a lower confidence weight is assigned to ensure that the system does not make overly aggressive compensation decisions due to excessively long causal paths.

[0080] Compensation strategy generation and multi-objective optimization Based on the identified causal path, the system generates a set of compensation strategies. Each of the compensation strategies Includes the following elements: Target process: The downstream process corresponding to the end point of the path.

[0081] Compensation parameters: The process parameters that need to be adjusted in this process.

[0082] Compensation amount: based on causal strength and deviation calculate: in The causal delay time from process i to process j (in steps or seconds). causal path The overall causal strength is defined as the product of the strengths of all causal edges on the path.

[0083] Execution sequence: based on delay time Determine the timing for issuing the compensation strategy. Specifically, the system sets a timer based on the expected output time of the current process i and the expected arrival time of the downstream process j. Compensation instructions are issued at all times, among which For the current time, This represents the equipment response time margin.

[0084] When multiple causal paths point to the same downstream process, a weighted summation method is used to merge the compensation amounts. To avoid overcompensation leading to new quality problems, this scheme employs a multi-objective optimization algorithm to select the optimal compensation strategy from the set of compensation strategies. The optimization objectives include at least: Minimize the predicted quality deviation ; Minimize compensation costs (Excessive parameter adjustment may introduce new instabilities). Maximize path credibility ; The NSGA-II algorithm was used for the solution. The NSGA-II parameters were set as follows: population size 100, number of generations 50, crossover probability 0.9, mutation probability 0.1, generating a set of optimal compensation strategies on the Pareto front. Finally, the compensation strategy that best matches the current operating condition was selected—by calculating the Euclidean distance between each strategy and successful strategies under similar historical operating conditions, the strategy with the smallest distance was chosen as the final execution strategy.

[0085] Automatic distribution and execution: After selecting the optimal compensation strategy, the system automatically distributes compensation parameter commands to the corresponding devices through the MES system's standard API interface (such as OPCUA, MQTT, or RESTful API). The distributed content includes: Target equipment identification (such as electroplating tank number, etching line number); The names and target values ​​of the process parameters that need to be adjusted (e.g., "current density: 22.5A / dm²"). Execution timestamp (calculated based on causal delay time, with millisecond-level precision); After receiving the instruction, the equipment controller automatically adjusts the corresponding process parameters to complete the closed-loop adjustment. The entire process requires no manual intervention, and the end-to-end response time from anomaly detection to parameter issuance is in the millisecond to second range, far superior to the hours of delay caused by manual intervention.

[0086] The beneficial effects of this module in various application scenarios are illustrated below through two typical scenarios: Scenario 1: Feedforward compensation for drill bit wear during the drilling process.

[0087] When the data acquisition module detects that the cumulative number of drill bit uses exceeds a threshold and the drill bit wear index rises, the system identifies a causal path based on a causal knowledge graph: drill bit wear → increased hole wall roughness → decreased adhesion of the hole metallization coating → risk of copper-free holes. Based on the path's reliability and causal strength, the system issues an early warning to the downstream electroplating process and automatically generates electroplating parameter compensation strategies (such as extending the micro-etching time by 3-5 seconds and increasing the current density by 0.5-1.0 A / dm²), effectively preventing the occurrence of copper-free defects in the holes. Compared to existing technologies that only detect copper-free holes during final electrical testing, this solution achieves intervention before the defect is exposed.

[0088] Scenario 2: Cross-process collaborative compensation for temperature deviations in the lamination process.

[0089] When the temperature of the lamination process deviates from the set value, the system identifies the causal path: lamination temperature deviation → change in the flow characteristics of the prepreg → uneven resin thickness between layers → uniform linewidth after etching. The system not only adjusts the temperature control parameters of the lamination process itself, but also predicts the impact of this deviation on the uniformity of the downstream etching process based on the causal graph, and simultaneously adjusts the spray pressure and etching solution concentration compensation value of the etching process to achieve coordinated control across processes.

[0090] Data transfer relationships between modules: The cross-process feedforward compensation module receives inference results (including anomaly type, root cause process, and predicted quality trend) from the causal reasoning and prediction module and a dynamic causal knowledge graph from the causal knowledge graph construction module as input. The generated compensation strategy is executed through the MES interface and passed to the self-evolutionary learning module as a compensation execution record for subsequent self-evolutionary learning.

[0091] The technical solution of the self-evolutionary learning module: The self-evolutionary learning module is configured to track and evaluate the actual quality effect after the compensation strategy is executed, and to perform closed-loop correction and self-evolutionary update of the causal edge weights and compensation strategy of the dynamic causal knowledge graph based on the reinforcement learning mechanism.

[0092] Reinforcement Learning Architecture Based on Deep Q-Network: This scheme adopts a Deep Q-Network (DQN) as the core architecture of the policy network. The Deep Q-Network approximates the state-action value function Q(s, a) through a deep neural network, realizing the value mapping from the state space to the action space.

[0093] Define the following core elements: State space S: A snapshot vector of the current dynamic causal knowledge graph, containing the feature values ​​of all nodes in the graph (real-time values ​​of process parameters, statistical features such as mean and standard deviation within the most recent time window), and the edge weight matrix (causal strength matrix). The state space has approximately [dimensions] and a path credibility vector. , where |V| is the number of nodes (typical value 50-200), d is the node feature dimension (typical value 64-256), and |E| is the number of edges (typical value 100-500).

[0094] Action Space A: The adjustable range of process parameters for the cross-process feedforward compensation module, including current density adjustment (-2 to +2 A / dm², step size 0.1), temperature compensation (-3 to +3°C, step size 0.5), pressure correction coefficient (0.8 to 1.2, step size 0.02), and etchant concentration increment (-5% to +5%, step size 0.5%). The action space is a continuous value space, approximated by discretization sampling (evenly dividing the value range of each parameter into 20 to 50 discrete actions).

[0095] The reward function R(s, a, s') is constructed based on the deviation between the actual quality result after the compensation strategy is implemented and the predicted quality result output by the causal inference and prediction module. Specifically, it is defined as follows: in This is the actual quality indicator vector (feedback from the subsequent quality inspection module, smoothed by Kalman filtering to reduce measurement noise). Let the target quality index vector be... The L2 norm penalty term for parameter adjustment range (to avoid over-adjustment), and I_anomaly is an anomaly indicator function (takes 1 when a new quality anomaly occurs after adjustment, otherwise takes 0). and This is the weighting coefficient (typically 0.01 and 0.1). The reward value ranges from [-10, 0], with values ​​closer to 0 indicating better control.

[0096] The loss function of a deep Q-network is defined as: in These are the parameters of the current Q-network (network structure: 3 fully connected hidden layers, 256 neurons per layer, activation function is ReLU). The parameters of the target Q network are copied from θ every fixed number of steps C=100. is the discount factor (typically 0.95 to 0.99), and D is the experience replay buffer (capacity set to 10,000 experience tuples, using a priority experience replay strategy).

[0097] Dual-loop correction mechanism of causal knowledge graph and compensation strategy: The self-evolutionary learning module of this scheme implements a dual-loop correction mechanism of causal knowledge graph and compensation strategy. The first closed loop—policy update: After the compensation policy is executed, the system obtains the actual quality results, calculates the reward value, and updates the policy network parameters through gradient descent of the deep Q-network. The policy network is trained using the Adam optimizer with an initial learning rate of 0.001, sampling 64 empirical tuples per batch, allowing the policy network to gradually converge to the optimal control policy.

[0098] The second closed loop—map update: If there is a significant deviation between the actual quality result and the predicted quality result (i.e.) δ is a preset deviation threshold, typically set at 30% of the tolerance of the target quality index, indicating that the strength of causal relationships in the causal knowledge graph may have errors.

[0099] The system triggers a re-evaluation process for causal edges, adjusting their weights based on actual observation data: in The learning rate is typically 0.05 to 0.2, corr(·) is the Pearson correlation coefficient between the observed causal parameter and the effect quality, and sign(·) takes a value of +1 (positive correlation) or -1 (negative correlation).

[0100] This update rule ensures that the direction of adjustment of causal weights is consistent with the direction of the actual observed correlation, and the adjustment magnitude is proportional to the magnitude of the prediction bias.

[0101] The model's adaptive evolution effect: Through the above mechanism, this system achieves a complete autonomous evolutionary cycle of "production → evaluation → learning → optimization": In the initial stage, the causal knowledge graph was constructed based on historical data and mechanistic knowledge, and the policy network was pre-trained on offline data (using the Proximal Policy Optimization (PPO) algorithm for 100 rounds of pre-training).

[0102] During continuous operation, the system learns from actual production data, and the causal graph gradually converges to a state that highly matches the actual working conditions, while the strategy network gradually approaches the optimal control strategy.

[0103] When production conditions drift (such as equipment aging or material batch changes), the system automatically senses the changes through the reward function and adaptively adjusts the causal weights and strategy parameters without the need for manual recalibration.

[0104] Data transfer relationships between modules: The self-evolutionary learning module receives compensation execution records (including compensation strategy content, execution time, and execution parameters) from the cross-process feedforward compensation module and actual quality detection results after compensation from the data acquisition module. The causal weight updates output by the module are fed back to the causal knowledge graph construction module, while the policy network parameter updates are fed back to the policy generation unit of the cross-process feedforward compensation module, forming a complete closed loop.

[0105] Digital twin simulation verification module: In some implementations of this solution, the system also includes a digital twin simulation verification module, configured to: input the compensation strategy into a pre-built digital twin model for virtual simulation verification before the compensation strategy generated by the cross-process feedforward compensation module is sent out for execution through the manufacturing execution system interface; if the simulation verification result meets the preset quality indicators, then the execution is allowed; if not, the system returns to the causal reasoning and prediction module to regenerate the compensation strategy.

[0106] The digital twin model is trained and generated based on historical data. It employs a hybrid architecture combining Long Short-Term Memory (LSTM) networks and Physical Information Neural Networks (PINNs) to provide high-fidelity simulation of the circuit board manufacturing process, demonstrating product quality outcomes under different combinations of process parameters. The training data for the digital twin model comes from complete sequences of process parameters and corresponding quality inspection results collected from historical production batches. After data cleaning and time-series alignment, it is trained using supervised learning. The simulation accuracy of the digital twin model is evaluated using the root mean square error (RMSE) metric. Retraining is triggered when the RMSE exceeds a preset threshold.

[0107] Performing virtual simulation verification before adjusting parameters on real equipment effectively avoids the risk of batch quality accidents that may be caused by improper adjustments, providing a second layer of protection for the system's security.

[0108] Visual monitoring and early warning module: In some implementations of this solution, the system also includes a visual monitoring and early warning module, which is configured to generate a monitoring dashboard in real time based on a multi-dimensional time-series database and a dynamic causal knowledge graph. The dashboard includes the real-time status of each process node, the causal impact propagation path, abnormal early warning information and process optimization suggestions. The dashboard is presented in a graphical manner through a human-computer interaction interface. When the abnormality type exceeds the preset risk level, a graded early warning push is triggered.

[0109] The tiered early warning mechanism includes three levels: Level 1 Warning (Alert, green indicator) – The predicted value of the quality indicator is close to the threshold but still within the tolerance range, requiring attention; Level 2 Warning (Warning, yellow indicator) – The predicted value of the quality indicator exceeds 80% of the threshold but is still within the threshold, requiring preparation for intervention; Level 3 Warning (Alarm, red indicator) – The predicted value of the quality indicator exceeds the threshold, immediately triggering automatic compensation or notifying operators to intervene. Warning information is simultaneously pushed to relevant personnel through enterprise instant messaging tools (such as DingTalk and WeChat Work) and email systems.

[0110] Taking quality control applied to the manufacturing process of multilayer circuit boards as an example, the specifics are as follows: Using the manufacturing process of an eight-layer FR-4 rigid circuit board as an application scenario, the specific implementation process of this solution system is illustrated by way of example.

[0111] Step 1: System deployment and data acquisition module configuration.

[0112] Sensors and data acquisition terminals are deployed at key nodes in the drilling, descaling, hole metallization, pattern electroplating, outer layer etching, and electrical testing processes. Parameters such as current density, bath temperature, and pH value are collected at a frequency of 100Hz, while ambient temperature and humidity data are collected at a frequency of 1Hz. AOI defect detection results and electrical test results are collected on a cycle based on each output. A QR code is laser-engraved on each circuit board unit to establish a single-board-level data archive, ensuring full-process data traceability from material input to finished product.

[0113] Step 2: Initialization of causal knowledge graph and construction of dual constraints.

[0114] Based on 12 months of historical production data (covering approximately 50,000 production batches), an initial causal knowledge graph was constructed using nonlinear Granger causality tests and dynamic Bayesian networks. Mechanistic constraints were applied to the graph using FMEA documents and expert rules, and a pre-defined "whitelist" and "blacklist" of causal edges were established to exclude physically impossible causal edges, such as those related to electroplating solution temperature affecting drill bit lifespan. The sliding window length L_window was set to 1000 sample points (corresponding to approximately 2 hours of production data), the sliding step size S_step was set to 250, and the KL divergence threshold was set. The severity level of FMEA is used as the initial weighting factor for the confidence of causal edges. The initial confidence of causal edges corresponding to failure modes with a severity level ≥ 8 is set to 0.9, that of severity levels 5-7 is set to 0.7, and that of severity levels ≤ 4 is set to 0.5.

[0115] Step 3: Real-time execution and causal reasoning prediction.

[0116] The system entered real-time operation. In a certain production batch, the data acquisition module detected that the current density of the pattern electroplating process continuously decreased from the preset value of 20.0 A / dm² to 19.2 A / dm², deviating from the threshold range [19.5, 20.5]. The causal reasoning and prediction module input the real-time data into a multi-hop graph attention network (set L=4 hops, H=8 attention heads), identified that the decrease in electroplating current density was the main reason for the decrease in the predicted value of the hole copper thickness from 25.0 μm to 21.8 μm, and output the root cause as "the decrease in effective current density caused by the consumption of electroplating anodes", with a contribution probability of 0.82 and a confidence interval of [0.78, 0.86]. It predicted that without intervention, a copper-free defect would occur in the hole, with a confidence level of 87%.

[0117] Step 4: Generation and execution of cross-process feedforward compensation.

[0118] The cross-process feedforward compensation module identifies the causal path: decreased electroplating current density → decreased deposition rate → insufficient copper thickness in the vias. The path reliability C(P) = 0.85, and the delay time τ = 1 process step (i.e., the current electroplating tank can be adjusted). The system generates a compensation strategy through multi-objective optimization: increasing the current density from 19.2 A / dm² to 21.0 A / dm², and simultaneously increasing the tank temperature from 28.0°C to 28.5°C to improve deep plating capability. After the compensation strategy is verified by digital twin simulation, it is automatically sent to the electroplating equipment controller for execution via the MES interface. After execution, the actual copper thickness in the vias recovers to 24.8 μm, meeting the IPC-6012 Class 2 standard requirement (≥20 μm).

[0119] Step 5: Self-evolutionary learning and closed-loop correction.

[0120] After compensation, the system tracked a deviation of 0.3 μm between the actual quality result (hole copper thickness 24.8 μm) and the predicted result (24.5 μm), with a reward value R = -0.09. The deep Q-network updated its strategy parameters accordingly. Meanwhile, since the deviation between the actual and predicted results was less than the threshold δ = 1.0 μm, the system determined the causal weights were accurate and did not trigger a map update.

[0121] After six months of continuous operation, the system's average yield increased from 94.2% before implementation to 97.8%, the scrap rate due to lack of copper in the holes decreased by 63%, and the number of manual interventions due to process abnormalities decreased from about 15 times per week to about 2 times per week.

[0122] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A circuit board manufacturing process-wide quality control system, characterized in that, It includes a data acquisition module, a causal knowledge graph construction module, a causal reasoning and prediction module, a cross-process feedforward compensation module, and a self-evolutionary learning module: The data acquisition module is configured to collect process parameter data streams, equipment operating status data, environmental variable data, and quality inspection result data of multiple process nodes in the circuit board manufacturing process in real time, and generate a multi-dimensional time series database. The causal knowledge graph construction module is configured to construct a cross-process dynamic causal knowledge graph based on a multi-dimensional time-series database and using a combination of nonlinear Granger causality test and dynamic Bayesian network. The nodes of the dynamic causal knowledge graph include at least process nodes, process parameter entity nodes, and quality index nodes. The edges represent the direction, strength, and delay time of the causal association. The dynamic causal knowledge graph supports adaptive updates of the causal structure triggered by KL divergence detection of newly acquired data and historical data within a sliding time window. The causal reasoning and prediction module is configured to input the real-time data stream of the multi-dimensional time series database into the dynamic causal knowledge graph, use a multi-hop graph attention network to perform multi-hop message passing and attention weight allocation, generate an updated graph structure containing long-range dependencies, and perform quality prediction and anomaly root cause localization based on the updated graph structure, and output reasoning results containing anomaly type, root cause process and predicted quality trend. The cross-process feedforward compensation module is configured to identify the causal path from the upstream process parameter node to the downstream process quality indicator node in the dynamic causal knowledge graph based on the reasoning result. When the upstream process parameter deviates from the preset threshold range, a compensation strategy for the downstream process is generated in advance based on the causal strength weight and path delay time of the causal path, and is automatically sent to the corresponding equipment execution parameter adjustment through the manufacturing execution system interface. The self-evolutionary learning module is configured to track and evaluate the actual quality effect after the compensation strategy is executed, and to perform closed-loop correction and self-evolutionary update of the causal edge weights and compensation strategy of the dynamic causal knowledge graph based on the reinforcement learning mechanism.

2. The circuit board production process quality control system according to claim 1, characterized in that, The construction of the dynamic causal knowledge graph adopts a dual-constraint mechanism that integrates process mechanism knowledge with data-driven discovery of causal relationships: The process mechanism knowledge is derived from failure mode and effects analysis documents and expert rules, and is used to pre-set the prior constraints on the existence of causal edges. The data-driven discovery of causal relationships is extracted from the multidimensional time-series database based on nonlinear Granger causality tests and structural equation modeling. The causal relationships discovered through data-driven discovery are filtered and corrected by the prior constraints, generating a hybrid causal knowledge graph that integrates mechanisms and data.

3. The circuit board production process quality control system according to claim 1, characterized in that, The cross-process feedforward compensation module is further configured as follows: The confidence of each candidate compensation path is weighted and calculated using a path confidence decay function, which uses the product of the path length and the confidence of each causal edge in the path as the decay factor. Based on the weighted calculation results, high-confidence transmission paths are activated first to generate a multi-level compensation strategy set; A multi-objective optimization algorithm is used to select the optimal compensation strategy from the set of multi-level compensation strategies.

4. The circuit board production process quality control system according to claim 1, characterized in that, The causal reasoning and prediction module is further configured as follows: A pre-trained trend semantic encoder is introduced to encode the trend pattern of process parameters within the current time window and generate a trend semantic vector. Based on trend semantic vectors, the candidate anomaly attribution paths output by the multi-hop graph attention network are reordered, and the root cause process list is output in descending order of causal influence.

5. The circuit board production process quality control system according to claim 1, characterized in that, The adaptive update of the causal structure in the causal knowledge graph construction module adopts the following mechanism: Set the sliding time window length and sliding step size, and perform KL divergence detection on newly acquired data and historical data within each sliding window; When the KL divergence exceeds a preset threshold, a nonlinear Granger causality test is triggered on the data within the current sliding window to re-evaluate the strength of the affected causal edges and dynamically adjust the edge weights and structural topology of the causal knowledge graph.

6. The circuit board production process quality control system according to claim 1, characterized in that, The closed-loop correction based on reinforcement learning in the self-evolutionary learning module includes: A deep Q-network is constructed as the policy network. The state space of the deep Q-network is a snapshot vector of the current dynamic causal knowledge graph, and the action space is the range of process parameters that can be adjusted by the cross-process feedforward compensation module. The reward function is constructed based on the absolute value of the deviation between the actual quality result after the compensation strategy is executed and the predicted quality result output by the causal inference and prediction module. The reward function guides the adaptive update of the policy network.

7. The circuit board production process quality control system according to claim 1, characterized in that, The data acquisition module is further configured to: set a unique physical identification code on each circuit board unit, wherein the physical identification code is a QR code, barcode or laser-engraved code; All data in the multidimensional time-series database are associated and stored using the physical identifier as an index, enabling precise binding of full-process data with specific circuit board units and supporting single-board-level quality traceability.

8. The circuit board production process quality control system according to claim 1, characterized in that, The plurality of process nodes include at least drilling process nodes, electroplating process nodes, lamination process nodes and etching process nodes; The compensation strategy generated by the cross-process feedforward compensation module includes at least one or more of the following combinations: current density compensation value for electroplating process, temperature curve compensation parameter for lamination process, and etching solution concentration compensation value for etching process.

9. The circuit board production process quality control system according to claim 1, characterized in that, The system also includes a digital twin simulation verification module, configured as follows: Before the compensation strategy generated by the cross-process feedforward compensation module is issued and executed through the manufacturing execution system interface, the compensation strategy is input into a pre-built digital twin model for virtual simulation verification. If the simulation verification results meet the preset quality indicators, then execution is allowed. If the conditions are not met, the causal reasoning and prediction module is triggered to re-perform the reasoning, and the cross-process feedforward compensation module regenerates the compensation strategy.

Images