Blind buried hole thick copper plate process optimization method based on Bayesian network

Abstract

The invention discloses a Bayesian network-based blind buried hole thick copper plate process optimization method. The method comprises the following steps of S1, collecting historical process parameters and product quality indexes in a blind buried hole thick copper plate manufacturing process; s2, preprocessing the collected data; s3, determining an initial process parameter set, and constructing an initial Bayesian network structure; s4, establishing a directed acyclic graph structure to form a complete Bayesian network; s5, executing reverse reasoning path analysis along the Bayesian network, and identifying a key process parameter set; s6, on the basis of the key process parameter set, optimizing the parameter combination by adopting a two-stage optimization strategy to generate an optimal parameter set; and S7, applying the optimal parameter set to the manufacturing process. According to the method, the Bayesian network and target oriented optimization are fused, intelligent screening and combination of the technological parameters of the blind buried hole thick copper plate are realized, and the method has the advantages of being high in causal interpretability, high in optimization efficiency and remarkable in yield improvement.

Description

Technical Field

[0001] The present invention relates to the field of optimizing the manufacturing process of circuit boards, and particularly to a method for optimizing the process of blind buried via thick copper clad boards based on Bayesian networks. Background Art

[0002] Blind buried via thick copper clad boards are a type of multi-layer printed circuit board with complex structures and extremely high process requirements. They are widely used in scenarios with high reliability and high-density wiring, such as server motherboards, automotive electronics, industrial control, 5G communication devices, and military electronics. Their structures usually include multiple inner-layer blind vias and buried via channels, the copper layer thickness is much higher than that of conventional PCBs, there are dozens of manufacturing process steps, and there are numerous parameter variables. The difficulty and risk of manufacturing process control are significantly higher than those of ordinary PCBs.

[0003] In the traditional manufacturing process, the yield of blind buried via thick copper clad boards highly depends on the experience accumulation of process engineers. Specifically, engineers need to complete process optimization by manually setting parameters, trial-and-error adjustment, batch evaluation, etc. based on past product experience and limited data analysis methods. However, due to the high coupling characteristics of blind buried vias and thick copper structures, there are highly non-linear and multi-level causal relationships between parameters, resulting in it being difficult for manual parameter adjustment to capture the deep influence mechanism, thus causing the following significant problems: First, the test cost is high and the trial-and-error cycle is long, severely restricting the new product development rhythm; second, the subjectivity of parameter configuration is strong and the reusability is poor, making it difficult to form a highly generalizable process decision support system; third, the product quality fluctuates greatly and the defect rate is high, manifested in common abnormal phenomena such as hole wall rupture, conduction failure, delamination and peeling off, seriously affecting the consistency and reliability of the final product.

[0004] To solve the above problems, in recent years, some PCB manufacturing enterprises and research institutions have tried to introduce data-driven modeling technologies, such as using methods like regression analysis, principal component analysis, neural networks, etc. to analyze and model the process data. These methods have achieved the identification of relationships between complex variables to a certain extent, but still face the following limitations: First, most methods only have the ability of "correlation modeling" and lack a causal relationship expression mechanism, unable to provide interpretable reasoning support for the impact of parameter changes on quality results; second, the model structure is often of a black box nature, making it difficult to reflect the explicit expression of process knowledge and having a high threshold for engineering applications; third, there is a lack of dynamic update and closed-loop feedback capabilities, and once the model is established, it is difficult to adapt to product model changes or process condition changes.

[0005] In addition, in terms of parameter combination optimization, traditional techniques mostly adopt exhaustive methods, empirical methods or heuristic tuning algorithms. Although these methods can improve the local parameter configuration efficiency to a certain extent, their optimization objectives are mostly based on simple functions (such as minimum error, optimal fitting), lacking a direct correlation with the actual product quality indicators. In the blind buried via thick copper clad board process, there are often synergistic or contradictory relationships between different product quality indicators. For example, there is a trade-off coupling behavior between the conduction rate and the hole wall defect rate. Optimization under a single objective often cannot ensure the balance and improvement of the overall quality indicators.

[0006] To sum up, the existing optimization techniques for blind buried via thick copper clad board manufacturing processes mainly have the following core defects: First, they rely on manual experience or experimental data for process parameter setting, lacking systematic and intelligent method support; second, data modeling methods cannot express the complex causal structure between process parameters, lacking process interpretability; third, in the process of parameter combination optimization, product quality indicators are not effectively introduced into the objective function, and true "quality-oriented process configuration" cannot be achieved; fourth, there is a lack of model adaptability and result stability evaluation mechanism for dynamic manufacturing environments, restricting the popularization and application of the methods. Summary of the Invention

[0007] An object of the present invention is to propose a blind buried via thick copper clad board process optimization method based on Bayesian network. The present invention integrates Bayesian network modeling and target-oriented optimization strategies, and aims at the multi-variable and highly coupled process problems in the manufacturing of blind buried via thick copper clad boards, realizes the intelligent screening and combined configuration of key process parameters, constructs an optimization mechanism with clear causality and strong stability, and has the advantages of strong inference interpretability, high parameter regulation efficiency, and significant improvement in process yield. It is applicable to high-density and high-reliability PCB manufacturing scenarios, and has broad engineering application value and industrial promotion prospects.

[0008] A blind buried via thick copper clad board process optimization method based on Bayesian network according to an embodiment of the present invention includes the following steps: S1. Collect historical process parameters and product quality indicators in the manufacturing process of blind buried via thick copper clad boards, and perform associated identification; S2. Preprocess the collected historical process parameters and product quality indicators to generate a standardized data set; S3. Based on a preset process optimization objective, use a target-oriented strategy to screen the standardized data set, determine an initial process parameter set, and construct an initial Bayesian network structure; S4. Establish a directed acyclic graph structure between the initial process parameter set and the product quality indicators, and calculate the conditional probabilities of each node in the network to form a complete Bayesian network; S5. Starting from the process optimization goal as the reasoning starting point, perform reverse reasoning path analysis along the Bayesian network to identify the set of key process parameters that affect the quality indicators of the target product; S6. Based on the set of key process parameters, optimize the parameter combination using a two-stage optimization strategy. The first stage is global rough search, using Latin hypercube sampling to generate candidate parameter combinations, and the second stage generates the optimal parameter set based on the candidate parameter combinations; S7. Apply the optimal parameter set to the manufacturing process to complete the optimization process.

[0009] Optionally, the process parameters include drilling-related parameters, lamination parameters, electroplating parameters, etching parameters, and surface treatment parameters.

[0010] Optionally, the product quality indicators include conduction rate, hole wall defect rate, interlayer bond strength, and overall defect rate.

[0011] Optionally, the preprocessing includes missing value filling, outlier removal, normalization processing, and discretization processing.

[0012] Optionally, the goal-oriented strategy is guided by the product quality objective function, screens out influential process parameters by calculating correlation and mutual information, and constructs an initial Bayesian network structure.

[0013] Optionally, the specific steps of S3 are as follows: S31. Preset the process optimization objective function based on the product quality indicators: ; Wherein, represents the value of the process optimization objective function, represents the conduction rate, represents the hole wall defect rate, represents the interlayer bond strength, represents the overall defect rate, , , , represents the preset weight, reflecting the priority of each indicator in the process optimization objective function; S32. Extract the process parameter matrix of all samples from the standardized dataset: ; Wherein, represents the process parameter matrix, represents the standardized value of the th process parameter in the th sample, represents the total number of samples, represents the total number of process parameters; Extract the quality index vector corresponding to each sample and calculate the corresponding process optimization objective function value; S33. Calculate the Pearson correlation coefficient between each process parameter and the objective function value: ; where, represents the Pearson correlation coefficient between the -th process parameter and the objective function, represents the sample mean of the -th process parameter, represents the mean of the objective function among all samples, represents the -th sample's corresponding objective function value; S34. Screen out the set of process parameters whose absolute value of the Pearson correlation coefficient exceeds the correlation threshold: ; where, represents the set of process parameters highly correlated with the objective function screened out, represents the correlation threshold, represents the -th process parameter; S35. Calculate the mutual information value between each process parameter in the set of process parameters and each product quality index: ; where, represents the mutual information value between the process parameter and the product quality index, represents the joint probability distribution of the process parameter and the product quality index, and represent their respective marginal probability distributions; S36. Combine correlation and mutual information to define the comprehensive evaluation score for each process parameter: ; where, represents the comprehensive evaluation score of the -th process parameter, and represent the weighting coefficients, represents the mutual information value between the process parameter and the product quality index, represents the -th Pearson correlation coefficient between the process parameter and the objective function, represents the -th process parameter, represents the -th product quality index; S37. Sort the process parameters of all process parameter sets in descending order according to the comprehensive evaluation score, and select the top process parameters as the initial process parameter set in the initial Bayesian network; S38. Combine the initial process parameter set with the corresponding product quality index to form the initial node set of the Bayesian network, and generate the initial Bayesian network structure.

[0014] Optionally, the S4 specifically includes: S41. Obtain the initial node set of the Bayesian network; S42. Use the hill-climbing algorithm to model the initial node set based on the standardized data set, and construct a directed acyclic graph structure; S43. On the premise that the directed acyclic graph structure is fixed, perform parameter learning on the conditional probability distribution of all nodes in the Bayesian network, and use the maximum likelihood estimation method to calculate the conditional probability of each node under the condition of its parent node: ; where represents the conditional probability value of node under the condition of its parent node , represents any node in the Bayesian network, represents the parent node of node , represents the number of occurrences of the joint state in the standardized data set, represents the total number of occurrences of the parent node combination, represents the number of times; S44. Combine the directed acyclic graph structure with the conditional probability values to form a complete Bayesian network.

[0015] Optionally, each edge of the directed acyclic graph structure represents the causal dependence relationship from the process parameter node to the product quality index node.

[0016] Optionally, the S5 specifically includes: S51. Use the set process optimization objective function to determine the target product quality index node concerned in the current optimization task of the target product quality index; S52. Starting from each target product quality index node, perform reverse path tracing in the directed acyclic graph structure, identify all process parameter nodes that have a directed path connection with it, and record the corresponding path set at the same time. Each path in the path set consists of multiple intermediate nodes; S53. Extract the forward conditional probability from the parameter part of the Bayesian network, and combine it with the marginal probability value, and use Bayes' formula to calculate the posterior probability: ; Among them, represents the posterior probability of process parameters when the target product quality index node is known, represents the conditional probability of the target index node under the condition of the process parameter node, represents the marginal probability of the process parameter node, represents the marginal probability of the target product quality index node, represents the target product quality index node, represents the process parameter node; S54. Construct a marginal influence degree function, calculate the marginal influence intensity of process parameters on the target product quality index, and construct a comprehensive causal scoring function to calculate the comprehensive causal score of each process parameter: ; Among them, represents the marginal influence intensity of process parameters on the target product quality index, represents the partial derivative operation, represents the comprehensive causal score of process parameters, represents the weight of the target product quality index; S55. For all process parameters participating in causal path reasoning, sort them in descending order according to the comprehensive causal score, and select the top process parameters with high scores to form a set of key process parameters.

[0017] Optionally, the S6 specifically includes: S61. Construct a parameter combination space based on the set of key process parameters: ; Among them, represents the parameter combination space, represents the process parameter the value set of, represents the number of key process parameters, represents the Cartesian product operator to generate parameter combinations; S62. Conduct a global rough adjustment search based on the parameter combination space, and use Latin hypercube sampling to generate candidate parameter combinations: ; Among them, represents the candidate parameter combination, represents the th parameter combination, represents the number of candidate parameter combinations; S63. Use the complete Bayesian network as the inference model, define the set of target product quality indices, and construct a target comprehensive scoring function combined with path complexity: ; Among them, represents the target comprehensive scoring function of the parameter combination ; represents the th parameter combination, represents the conditional probability of the target product quality index under the parameter combination, represents the th target product quality index of the parameter combination, represents the weight value of the target product quality index; ; Among them, represents the path information entropy of the target product quality index node of the th parameter combination, represents the set of parent nodes of the target product quality index node, represents the marginal probability of the target product quality index node, represents the set of target product quality indexes, represents the th target product quality index of the parameter combination, represents the th path information entropy of the target product quality index node of the parameter combination; S64. Rank the candidate parameter combinations according to the target comprehensive scoring function, and select the top optimal parameter combinations to form a fine-tuning parameter set; S65. Construct a sensitivity-driven neighborhood set around the parameter combinations in each fine-tuning parameter set: ; Among them, represents the neighborhood set centered on the parameter combination, represents the Euclidean distance between parameter combinations, represents the dynamic neighborhood radius calculated based on sensitivity, represents the parameter combination; S66. For the parameter combinations in each neighborhood, calculate the target stability index based on the output results of multi-round inference of the Bayesian network: ; Among them, represents the target stability index of the parameter combination, represents the number of target product quality indexes, represents the standard deviation of the prediction probability of the parameter combination for the target product quality index under multi-round inference; S67. Modify the comprehensive optimization objective function: ; in, represents the comprehensive optimization objective function; S68. Perform optimal search for each neighborhood and obtain the local optimal parameter combination by maximizing the comprehensive optimization objective function: ; in, Indicates The local optimal parameter combination in the neighborhood is represented by Indicates the variable value when the function value is maximum; S69. Summarize all local optimal parameter combinations into an optimal parameter set.

[0018] The beneficial effects of the present invention are: First of all, the present invention provides a blind and buried via thick copper plate process optimization method based on Bayesian network, which can realize causal modeling, goal-oriented parameter screening and combination optimization between process parameters and product quality in a multi-dimensional and multi-stage PCB manufacturing process, and overcomes the core technical difficulties such as traditional process reliance on experience and trial and error, black box optimization process, fuzzy objective function and difficulty in adapting to complex manufacturing scenarios. By collecting and preprocessing historical process data during the manufacturing process of blind and buried via thick copper plates, and constructing a standardized data set in combination with product quality indicators, the method can systematically identify key process variables that have a high influence on the target quality results, effectively reduce invalid parameter interference, and improve modeling efficiency and process interpretability.

[0019] Secondly, the present invention introduces a goal-oriented strategy with product quality indicators as the optimization target. By combining correlation analysis with mutual information calculation, the initial process parameter set is scientifically determined, and then a Bayesian network structure based on a directed acyclic graph is constructed, which realizes the modeling and quantitative expression of the causal dependency relationship between complex variables in the thick copper blind buried via process. By combining Bayesian network structure learning with parameter learning, a complete Bayesian model that can be used for reasoning is formed, and reverse reasoning path analysis is performed with the target quality indicator node as the starting point, which accurately identifies the key process parameter set, ensures that the optimization action is strongly correlated with product performance, and significantly improves the engineering value and effectiveness of the optimization direction.

[0020] Finally, in the parameter optimization stage, the present invention proposes a two-stage parameter optimization strategy. First, Latin hypercube sampling is used to generate candidate combinations in the global space. Then, based on the candidate set, an objective scoring function weighted by Bayesian inference values and path information entropy is constructed. A parameter perturbation stability evaluation mechanism and a sensitivity-driven neighborhood construction method are introduced to achieve double optimization of the candidate parameter combinations in terms of the objective comprehensive score and prediction stability. Further, by combining the joint scoring function and the heuristic search strategy, the problem of getting stuck in local optimal solutions is avoided, while taking into account the logical continuity of the process stage and the practical feasibility of parameter combinations, greatly improving the credibility, stability, and implementability of the parameter combination optimization results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of a blind buried via thick copper clad laminate process optimization method based on a Bayesian network proposed by the present invention; Figure 2 is a schematic diagram of the target-oriented process parameter screening and initial Bayesian network structure construction process of a blind buried via thick copper clad laminate process optimization method based on a Bayesian network proposed by the present invention; Figure 3 is a flowchart of parameter combination optimization based on a two-stage optimization strategy for a blind buried via thick copper clad laminate process optimization method based on a Bayesian network proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0023] Refer to Figures 1-3 , a blind buried via thick copper clad laminate process optimization method based on a Bayesian network, includes the following steps: S1. Collect historical process parameters and product quality indicators in the manufacturing process of blind buried via thick copper clad laminates and perform associated identification; S2. Preprocess the collected historical process parameters and product quality indicators to generate a standardized data set; S3. Based on a preset process optimization goal, use a target-oriented strategy to screen the standardized data set, determine an initial process parameter set, and construct an initial Bayesian network structure; S4. Establish a directed acyclic graph structure between the initial process parameter set and the product quality indicators, and calculate the conditional probabilities of each node in the network to form a complete Bayesian network; S5. Take the process optimization goal as the starting point of reasoning, perform reverse reasoning path analysis along the Bayesian network, and identify the set of key process parameters that affect the quality indicators of the target product; S6. Based on the set of key process parameters, optimize the parameter combination using a two-stage optimization strategy. The first stage is global rough search, using Latin hypercube sampling to generate candidate parameter combinations, and the second stage generates the optimal parameter set based on the candidate parameter combinations; S7. Apply the optimal parameter set to the manufacturing process to complete the optimization process.

[0024] In this embodiment, the process parameters include drilling-related parameters, lamination parameters, electroplating parameters, etching parameters, and surface treatment parameters.

[0025] In this embodiment, the product quality indicators include conduction rate, hole wall defect rate, interlayer bonding strength, and overall defect rate.

[0026] In this embodiment, the preprocessing includes missing value filling, outlier removal, normalization processing, and discretization processing.

[0027] In this embodiment, the target-oriented strategy is guided by the product quality objective function, screens out influential process parameters by calculating correlation and mutual information, and constructs an initial Bayesian network structure.

[0028] In this embodiment, the specific steps of S3 are as follows: S31. Preset the process optimization objective function based on the product quality indicators: ; Among them, represents the value of the process optimization objective function, represents the conduction rate, represents the hole wall defect rate, represents the interlayer bonding strength, represents the overall defect rate, , , , represents the preset weight, reflecting the priority of each indicator in the process optimization objective function; S32. Extract the process parameter matrix of all samples from the standardized dataset: ; Among them, represents the process parameter matrix, represents the standardized value of the th process parameter in the th sample, represents the total number of samples, represents the total number of process parameters; Extract the quality index vector corresponding to each sample, and calculate the corresponding process optimization objective function value; S33. Calculate the Pearson correlation coefficient between each process parameter and the objective function value: ; where, represents the Pearson correlation coefficient between the th process parameter and the objective function, represents the sample mean of the th process parameter, represents the mean of the objective function among all samples, represents the th sample corresponding objective function value; S34. Screen out the set of process parameters whose absolute value of the Pearson correlation coefficient exceeds the correlation threshold: ; where, represents the set of process parameters highly correlated with the objective function screened out, represents the correlation threshold, represents the th process parameter; S35. Calculate the mutual information value between each process parameter in the set of process parameters and each product quality index: ; where, represents the mutual information value between the process parameter and the product quality index, represents the joint probability distribution of the process parameter and the product quality index, and represent their respective marginal probability distributions; S36. Define the comprehensive evaluation score of each process parameter by integrating correlation and mutual information: ; where, represents the comprehensive evaluation score of the th process parameter, and represent the weighting coefficients, represents the mutual information value between the process parameter and the product quality index, represents the Pearson correlation coefficient between the th process parameter and the objective function, represents the th process parameter, represents the th product quality index; S37. Sort the process parameters of all process parameter sets in descending order according to the comprehensive evaluation score, and select the first process parameters as the initial process parameter set in the initial Bayesian network; S38. Combine the initial process parameter set with the corresponding product quality indicators into the initial node set of the Bayesian network, and generate the initial Bayesian network structure.

[0029] In this embodiment, the S4 specifically includes: S41. Obtain the initial node set of the Bayesian network; S42. Use the hill-climbing algorithm to model the initial node set based on the standardized data set, and construct a directed acyclic graph structure; S43. On the premise that the directed acyclic graph structure is fixed, perform parameter learning on the conditional probability distributions of all nodes in the Bayesian network, and use the maximum likelihood estimation method to calculate the conditional probability of each node under the condition of its parent nodes: ; Wherein, represents the conditional probability value of node under the condition of its parent node , represents any node in the Bayesian network, represents the parent node of node , represents the number of occurrences of the joint state in the standardized data set, represents the total number of occurrences of the parent node combination, represents the number of times; S44. Combine the directed acyclic graph structure with the conditional probability values to form a complete Bayesian network.

[0030] In this embodiment, each edge of the directed acyclic graph structure represents the causal dependence relationship from the process parameter node to the product quality indicator node.

[0031] In this embodiment, the S5 specifically includes: S51. Use the set process optimization objective function to determine the target product quality indicator node concerned in the current optimization task among the target product quality indicators; S52. Starting from each target product quality indicator node, perform reverse path tracing in the directed acyclic graph structure, identify all process parameter nodes that have a directed path connection with it, and record the corresponding path sets at the same time. Each path in the path sets consists of multiple intermediate nodes; S53. Extract the forward conditional probability from the parameter part of the Bayesian network, and combine it with the marginal probability value, and use Bayes' formula to calculate the posterior probability: ; Among them, represents the posterior probability of process parameters when the target product quality index node is known, represents the conditional probability of the target index node under the condition of the process parameter node, represents the marginal probability of the process parameter node, represents the marginal probability of the target product quality index node, represents the target product quality index node, represents the process parameter node; S54. Construct a marginal influence degree function, calculate the marginal influence intensity of process parameters on the target product quality index, and construct a comprehensive causal scoring function to calculate the comprehensive causal score of each process parameter: ; Among them, represents the marginal influence intensity of process parameters on the target product quality index, represents the partial derivative operation, represents the comprehensive causal score of process parameters, represents the weight of the target product quality index; S55. For all process parameters participating in causal path reasoning, sort them in descending order according to the comprehensive causal score, and select the first process parameters with high scores to form a set of key process parameters.

[0032] In this embodiment, the S6 specifically includes: S61. Construct a parameter combination space based on the set of key process parameters: ; Among them, represents the parameter combination space, represents the process parameter value set, represents the number of key process parameters, represents the Cartesian product operator to generate parameter combinations; S62. Conduct a global rough adjustment search based on the parameter combination space, and use Latin hypercube sampling to generate candidate parameter combinations: ; Among them, represents the candidate parameter combination, represents the th parameter combination, represents the number of candidate parameter combinations; S63. Using the complete Bayesian network as the inference model, define the set of target product quality indicators, and construct a target comprehensive scoring function that combines path complexity: ; Among them, represents the target comprehensive scoring function of the parameter combination , represents the th parameter combination, represents the conditional probability of the target product quality indicators under the parameter combination, represents the th target product quality indicator of the parameter combination, represents the weight value of the target product quality indicators; ; Among them, represents the path information entropy of the target product quality indicator node of the th parameter combination, represents the set of parent nodes of the target product quality indicator node, represents the marginal probability of the target product quality indicator node, represents the set of target product quality indicators, represents the th target product quality indicator of the parameter combination, represents the th path information entropy of the target product quality indicator node of the parameter combination; S64. Rank the candidate parameter combinations according to the target comprehensive scoring function, and select the top optimal parameter combinations to form a fine-tuning parameter set; S65. Construct a sensitivity-driven neighborhood set around the parameter combinations in each fine-tuning parameter set: ; Among them, represents the neighborhood set centered on the parameter combination, represents the Euclidean distance between parameter combinations, represents the dynamic neighborhood radius calculated by sensitivity, represents the parameter combination; S66. For the parameter combinations in each neighborhood, calculate the target stability index based on the output results of multiple rounds of inference of the Bayesian network: ; Among them, represents the target stability index of the parameter combination, represents the number of target product quality indicators, It represents the standard deviation of the predicted probability of the target product quality index under multiple rounds of inference for the parameter combination. S67. Modify the comprehensive optimization objective function: ; Among them, represents the comprehensive optimization objective function; S68. Perform an optimal search for each neighborhood, and obtain the local optimal parameter combination by maximizing the comprehensive optimization objective function: ; Among them, represents the th local optimal parameter combination within the neighborhood, and represents the variable value when taking the maximum value of the function value;

[0033] Example 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to the optimization process of the batch production process of blind buried via thick copper plates for a newly developed high-layer and high-reliability server motherboard in a communication equipment manufacturing enterprise in 2024. The designed number of layers of this batch of products is 18 layers, the copper thickness requirements are 70μm for the inner layer and 100μm for the outer layer, the blind holes penetrate through layers 4 - 6 and 12 - 14, the manufacturing process is complex, and the required finished product yield is ≥98%. However, the yield in the initial trial production stage is only 93.6%. The main defects are rough blind hole walls, through hole cracks, and inner layer delamination, which seriously affect the mass production and delivery progress.

[0034] In this actual manufacturing environment, first, data on the same type of blind buried via thick copper plate production line within the past 6 months is collected, and the process data of 12,000 products is cumulatively summarized, covering 45 process parameter information of the main processes such as drilling, lamination, electroplating, etching, and surface treatment, as well as 4 main product quality indicators, including conduction rate, hole wall defect rate, interlayer bond strength, and overall defect rate. Missing value filling, outlier removal, normalization, and discretization processing are performed on the collected data, and finally a standardized data set is generated.

[0035] Subsequently, based on the preset process optimization objectives, that is, maximizing the conduction rate, minimizing the hole wall defect rate and the overall defect rate, and at the same time ensuring that the interlayer bond strength is higher than 1200 N / cm², an objective function model is established. Through the Pearson correlation coefficient, initially relevant process parameters are screened out, and further combined with mutual information analysis, the influence degree of each parameter on the process objective is comprehensively evaluated. Finally, 22 initial process parameters are determined as the initial Bayesian network node set.

[0036] Using the method proposed by the present invention, a hill-climbing algorithm is adopted for structure learning to construct a preliminary directed acyclic graph. Then, parameter learning of the conditional probabilities between nodes is carried out by means of maximum likelihood estimation to form a complete Bayesian network. Through reverse path inference analysis, with the conduction rate and the hole wall defect rate as the inference starting points, 10 key process parameters are identified, including the drilling diameter, the drill speed, the lamination temperature, the electroplating current density, the etching solution temperature, etc., which have a significant impact on the target indicators.

[0037] In the parameter combination optimization stage, a combination space is constructed based on the key process parameters. 3000 groups of candidate parameter combinations are generated by Latin hypercube sampling. The target comprehensive score of each group of parameters is calculated using the Bayesian network inference model. Combining with the dynamic weight adjustment of path entropy, and at the same time introducing the target stability index for combination screening. In the local fine-tuning stage, based on the sensitivity adaptive neighborhood search and comprehensive heuristic search strategies, 8 groups of optimal parameter combinations are finally screened out.

[0038] To verify the actual effect of the method of the present invention, the 8 groups of optimal parameter combinations screened out are applied to actual production verification, and batch trial productions are carried out respectively.

[0039] Table 1 Comparison table of the overall performance before and after the optimization of the manufacturing process of blind buried via thick copper plates ; From the perspective of product quality indicators, the average conduction rate of the samples before optimization was 96.7%. After implementing the method of the present invention, the conduction rate increased to 99.5%, with an increase amplitude of 2.8%, effectively reducing the problem of unqualified electrical performance. The hole wall defect rate decreased significantly from 3.2% before optimization to 0.8%, with a decrease amplitude of 2.4%, indicating that through the precise control of key process parameters, the occurrence of common defects such as rough hole walls and cracks in the processing of blind holes and buried holes is inhibited. The interlayer bonding strength increased from 1208 N / cm² before optimization to 1320 N / cm², an increase of 112 N / cm², fully proving that the optimized process parameter combinations such as lamination and electroplating in the present invention have a significant improvement effect on the inner layer bonding reliability. The overall defect rate decreased from 5.5% to 1.1%, with a decrease amplitude of 4.4%, reflecting a qualitative leap in the overall process control level and product consistency.

[0040] In terms of process effects, the finished product yield before optimization was only 93.6%. After being optimized by the method of the present invention, the yield increased to 98.7%, with an increase amplitude of 5.1%. It not only meets the requirements of high yield delivery for high-end server motherboards, but also significantly reduces production losses and rework costs. At the same time, through the data-driven parameter screening and combination strategy of the present invention, the process development cycle is shortened from about 90 days originally to 52 days, with a shortening ratio of more than 42%, greatly accelerating the new product introduction and mass production rhythm and improving the manufacturing response speed.

[0041] In addition, the present invention also significantly improves the stability of the manufacturing process. Through Bayesian network inference path analysis, stability index calculation, and multiple rounds of inference optimization, the fluctuation range of the batch yield has decreased from ±3.5% before optimization to ±1.2%, with a 2.3% reduction in fluctuations. This reflects that the optimized process parameter configuration can still maintain high consistency and robustness under different batches and different equipment conditions, significantly reducing the abnormal risks during the production process.

[0042] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A method for optimizing the process of blind buried via thick copper clad plates based on Bayesian network, characterized in that, It includes the following steps: S1. Collect the historical process parameters and product quality indicators during the manufacturing process of blind buried via thick copper plates, and perform associated identification; S2. Preprocess the collected historical process parameters and product quality indicators to generate a standardized data set; S3. Based on a preset process optimization goal, use a goal-oriented strategy to screen the standardized data set, determine the initial process parameter set, and construct an initial Bayesian network structure; S4. Establish a directed acyclic graph structure between the initial process parameter set and the product quality indicators, and calculate the conditional probabilities of each node in the network to form a complete Bayesian network; S5. Taking the process optimization goal as the starting point of reasoning, perform reverse reasoning path analysis along the Bayesian network to identify the key process parameter set that affects the target product quality indicators; S6. Based on the key process parameter set, use a two-stage optimization strategy to optimize the parameter combination. The first stage is global rough adjustment search, using Latin hypercube sampling to generate candidate parameter combinations, and the second stage is to generate the optimal parameter set based on the candidate parameter combinations; S7. Apply the optimal parameter set to the manufacturing process to complete the optimization process.

2. The method for optimizing the process of blind buried via thick copper clad plate based on Bayesian network according to claim 1, wherein The process parameters include drilling-related parameters, lamination parameters, electroplating parameters, etching parameters, and surface treatment parameters.

3. The blind buried via thick copper clad laminate process optimization method based on Bayesian network according to claim 1, characterized in that, The product quality indicators include conduction rate, hole wall defect rate, interlayer bonding strength, and overall defect rate.

4. A blind buried via thick copper clad laminate process optimization method based on Bayesian network according to claim 1, characterized in that, The preprocessing includes missing value filling, outlier removal, normalization processing, and discretization processing.

5. A method for optimizing the process of blind buried via thick copper clad laminate based on Bayesian network according to claim 1, characterized in that, The goal-oriented strategy is guided by the product quality objective function, screens out influential process parameters by calculating correlation and mutual information, and constructs an initial Bayesian network structure.

6. The blind buried via thick copper clad laminate process optimization method based on Bayesian network according to claim 1, characterized in that The specific steps of S3 are as follows: S31. Preset a process optimization objective function based on the product quality indicators; S32. Extract the process parameter matrix of all samples from the standardized data set; At the same time, extract the quality indicator vector corresponding to each sample, and calculate the corresponding process optimization objective function value; S33. Calculate the Pearson correlation coefficient between each process parameter and the objective function value; ； Among them, represents the Pearson correlation coefficient between the th process parameter and the objective function, represents the sample mean of the th process parameter, represents the mean of the objective function among all samples, represents the objective function value corresponding to the S34. Screen out the process parameter set whose absolute value of the Pearson correlation coefficient exceeds the correlation threshold; S35. Calculate the mutual information value for each process parameter and each product quality indicator in the process parameter set; S36. Combine correlation and mutual information to define the comprehensive evaluation score of each process parameter; ； Among them, represents the comprehensive evaluation score of the th process parameter, and represent the weighting coefficients, represents the mutual information value between the process parameter and the product quality index, represents the Pearson correlation coefficient between the th process parameter and the objective function, represents the th process parameter, represents the th product quality index; S37. Sort the process parameters of all process parameter sets in descending order according to the comprehensive evaluation score, and select the first process parameters as the initial process parameter set in the initial Bayesian network; S38. Combine the initial process parameter set with the corresponding product quality indicators into the initial node set of the Bayesian network to generate an initial Bayesian network structure.

7. A blind buried via thick copper clad laminate process optimization method based on Bayesian network according to claim 1, characterized in that The specific steps of S4 are as follows: S41. Obtain the initial node set of the Bayesian network; S42. Use the hill climbing algorithm to model the initial node set based on the standardized data set to construct a directed acyclic graph structure; S43. On the premise that the directed acyclic graph structure is fixed, perform parameter learning on the conditional probability distribution of all nodes in the Bayesian network, and use the maximum likelihood estimation method to calculate the conditional probability of each node under the condition of its parent nodes; S44. Combine the directed acyclic graph structure with the conditional probability values to form a complete Bayesian network.

8. A method for optimizing the process of blind buried via thick copper clad plate based on Bayesian network according to claim 7, characterized in that, Each edge of the directed acyclic graph structure represents a causal dependency relationship from a process parameter node to a product quality indicator node.

9. A blind buried via thick copper clad laminate process optimization method based on Bayesian network according to claim 1, characterized in that, The S5 specifically includes: S51, using the set process optimization objective function, determining the target product quality indicator node of the target product quality indicator that the current optimization task focuses on; S52, taking each target product quality indicator node as a starting point, performing reverse path tracing in the directed acyclic graph structure, identifying all process parameter nodes connected to it by directed paths, and recording a corresponding path set, wherein each path in the path set is composed of a plurality of intermediate nodes; S53, extracting the forward conditional probability from the parameter part of the Bayesian network, and combining the marginal probability value, using the Bayesian formula to calculate the posterior probability; S54, constructing a marginal influence function to calculate the marginal influence intensity of the process parameters on the target product quality indicators, and constructing a comprehensive causal scoring function to calculate the comprehensive causal score of each process parameter; S55. Sort all the process parameters involved in causal path reasoning in descending order according to the comprehensive causal score, and select the top process parameters with high scores to form a set of key process parameters.

10. A method for optimizing the process of blind buried via thick copper clad plate based on Bayesian network according to claim 1, characterized in that, The S6 specifically includes: S61. Constructing a parameter combination space based on a set of key process parameters; S62, performing a global coarse adjustment search based on the parameter combination space, and using Latin hypercube sampling to generate candidate parameter combinations; S63. Using the complete Bayesian network as the inference model, define the target product quality indicator set and construct a target comprehensive scoring function combining the path complexity: ； Among them, represents the target comprehensive scoring function of the parameter combination represents the th parameter combination represents the conditional probability of the target product quality index under the parameter combination represents the th target product quality index of the parameter combination represents the weight value of the target product quality index; ； Among them, represents the path information entropy of the target product quality index node of the th parameter combination, represents the set of parent nodes of the target product quality index node, represents the marginal probability of the target product quality index node, represents the set of target product quality indexes, represents the th target product quality index of the parameter combination, represents the th path information entropy of the target product quality index node of the parameter combination; S64. Sort the candidate parameter combinations according to the target comprehensive scoring function, and select the top optimal parameter combinations to form a fine-tuning parameter set; S65, constructing a sensitivity-driven neighborhood set around the parameter combination in each fine-tuning parameter set; S66. For each parameter combination in the neighborhood, the target stability index is calculated based on the output results of multiple rounds of Bayesian network reasoning; S67, modifying the comprehensive optimization objective function; S68, performing an optimal search for each neighborhood, and obtaining a local optimal parameter combination by maximizing the comprehensive optimization objective function; S69. Summarize all local optimal parameter combinations into an optimal parameter set.

Images