PCB impedance characteristic prediction and evaluation method based on neural network

Abstract

The invention discloses a PCB impedance characteristic prediction and evaluation method based on a neural network. The method comprises the steps of collecting and standardizing multi-batch PCB impedance sample data, normalizing and cleaning multi-dimensional characteristics, dividing characteristic subspaces based on a parameter attribution algorithm, constructing a multi-branch neural network with physical prior regularization, and realizing task specificity and collaborative decoupling. And through collaborative optimization of a multi-task loss function, the precision and consistency of the model are dynamically monitored, and multi-task prediction and feature interpretation are output. According to the scheme, the prediction accuracy of the model under complex working conditions is improved, and an intelligent data basis and performance monitoring capability are provided for PCB design and process optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of printed circuit board (PCB) impedance characteristic modeling and multi-objective neural network prediction, and particularly relates to a PCB impedance characteristic prediction and evaluation method based on a neural network. BACKGROUND

[0002] With the continuous progress of printed circuit board (PCB) design and manufacturing technology, and the increasing demand of electronic products for high-speed signals and high-reliability interconnections, the accurate prediction, control and collaborative optimization of PCB impedance characteristics have become a key technical problem in the industry. For impedance modeling and performance prediction of multiple types of transmission lines (such as microstrip lines, stripline lines, and differential lines) under multiple process and environmental conditions, data-driven methods such as neural networks have gradually become mainstream technical means. Currently, under high-dimensional parameter space, multi-objective and multi-task impedance characteristic prediction requirements are increasingly prominent, involving complex correlations of multiple types of data such as structural parameters, process parameters, and environmental conditions. The industry generally uses methods such as multi-branch neural networks, joint loss function optimization, and feature engineering attribution to reduce the complexity and human dependence of traditional physical modeling, and to improve the generalization and automatic modeling capabilities of the model.

[0003] In the prior art, multi-objective PCB impedance modeling is usually based on a multi-layer fully connected neural network structure (MLP), a shared bottom feature extraction plus multi-branch output architecture, or a general method integrating multi-task learning (MTL). Such solutions rely heavily on normalized feature processing and parameter attribution sorting to strengthen automatic feature extraction and key parameter aggregation. Some literature or patents also introduce regularization loss terms to improve the overall prediction performance of multi-objective tasks. In recent years, in response to high-dimensional parameter and multi-process label practical application scenarios, techniques such as task collaboration, hierarchical feature sharing, and parameter sensitivity analysis under multi-branch architecture have continued to develop, gradually evolving towards model interpretability, task collaborative optimization, elastic decoupling, and industry feedback.

[0004] However, the above-mentioned traditional multi-objective neural network modeling scheme still has the following outstanding problems and limitations in the field of high-dimensional PCB impedance prediction:

[0005] (1) When modeling multiple objectives / multi-tasks jointly, there is a significant overlap of features and coupling of parameters between tasks, which causes mutual interference of gradients between branch outputs, ultimately leading to a decrease in the prediction accuracy of one or more objectives, an imbalance in the prediction distribution, and possibly physically unreasonable outputs in some branches;

[0006] (2) Existing parameter attribution and feature allocation schemes are mostly static or manually set, lacking a parameter-task mapping mechanism based on physical causality and dynamic optimization, making it difficult to adapt to task-specific differences caused by process variations and environmental interweaving in a multi-dimensional parameter space;

[0007] (3) Lack of systematic physical constraints and explainable mechanisms, parameter paths are often mixed across target branches, the internal decision logic of the model is seriously "black-boxed", which is not conducive to high credibility application and process closed-loop feedback in the field of PCB design, and some models are difficult to give parameter / process improvement suggestions;

[0008] (4) Lack of task coordination consistency, existing loss functions and training processes cannot effectively measure and adjust the physical consistency relationship between multi-objective prediction results, which easily leads to uncontrolled physical redundancy items in the output results or task conflicts;

[0009] (5) Most of the disclosed technical solutions fail to support adaptive dynamic optimization of feature subspace and regular weight structure parameters during the training process, making it difficult to ensure long-term stability of model performance and progress of precision under actual parameter perturbation, distribution drift and new process import. SUMMARY

[0010] The present application provides a PCB impedance characteristic prediction and evaluation method based on a neural network to solve the above technical problems.

[0011] The technical solution of the present application is as follows: a PCB impedance characteristic prediction and evaluation method based on a neural network, comprising:

[0012] S1: Collect high-dimensional PCB impedance historical data samples containing different structural parameters, process parameters and environmental parameters, and record the corresponding task label information of the samples for multiple types of transmission lines, frequency bands and temperature conditions.

[0013] S2: Perform normalization, outlier rejection and condition label division processing on the collected high-dimensional PCB impedance historical data samples to generate data preprocessing results that meet the needs of multi-task collaborative modeling.

[0014] S3: Based on the normalized high-dimensional PCB impedance historical data samples, use a parameter attribution algorithm to divide the feature subspace, group the multi-dimensional features that dominate the impedance output of different transmission line types, frequency bands and environmental conditions, and form a parameter allocation scheme for task-specific modeling.

[0015] S4: Input the parameter allocation scheme into the multi-branch neural network modeling architecture, configure a shared bottom feature extraction layer, and for each type of transmission line and target impedance output, build a multi-objective prediction branch with a task-specific decoupling layer.

[0016] S5: Apply a physically prior-guided explainable regularization term in the task-specific decoupling layer, control the regularization strength to limit some process parameters to only affect the specified target impedance output, and realize feature decomposition and correlation constraint of each target prediction branch.

[0017] S6: The regularized multi-branch neural network modeling architecture is jointly trained using a multi-task loss function and an inter-task collaborative consistency supervision mechanism, the physical consistency or known redundant relationship between different target prediction results is measured and adjusted, and the training process is dynamically optimized.

[0018] S7: During the joint training process, the target prediction accuracy and the inter-task collaborative consistency index are monitored in real time based on the validation set, the feature subspace decoupling ratio and the regularization constraint strength are dynamically adjusted, and the accuracy collaboration between tasks and the model stability are adaptively optimized.

[0019] S8: Using the optimized multi-branch neural network modeling architecture, for new input of different structural parameters, process parameters and multi-condition data, the target impedance prediction results and their uncertainty intervals under each task are inferred and output.

[0020] S9: Based on the target impedance prediction results, feature importance decomposition information and physical consistency analysis index output by inference, the model performance is evaluated, and knowledge feedback and multi-target performance monitoring report for PCB design or manufacturing process improvement are generated.

[0021] The application provides a PCB impedance characteristic prediction and evaluation method based on a neural network, which has the following specific beneficial effects:

[0022] (1) The existing PCB impedance multi-target modeling often uses parameter full set / shared feature input, which easily leads to gradient conflict and irrelevant feature interference between target branches, causing prediction accuracy to decrease and even model training to be unstable. The application groups the input features in fine granularity through a parameter attribution algorithm, and designs a "shared-specific-decoupled" multi-branch neural network architecture according to the physical influence degree of each target task. On the basis of shared bottom feature extraction, through the task-specific decoupling layer, highly correlated parameters are accurately allocated to each prediction branch, and physical prior regularization constraints are used to effectively suppress irrelevant disturbances between tasks. Experiments show that this scheme can reduce the prediction error between multi-targets by 15% to 35%, greatly improving the overall stability of model training and inference;

[0023] (2) For the "black box" characteristics of traditional multi-target networks, the application introduces physical prior knowledge in the feature decomposition and branch mapping stage, deeply integrates the physical coupling relationship of process parameters / structural parameters and artificial intelligence modeling. Through adjustable design of the regularization term, some features are forced to act only on the physically related target impedance output, so that the model has the ability of parameter tracing and mechanism explanation. Finally, the prediction results of each target can be output while tracking the contribution and importance of key parameters, significantly enhancing the transparency and credibility of model output for engineering users, and realizing "explainable AI" in the impedance prediction field;

[0024] (3) Existing multi-objective / multi-task neural network solutions often face typical problems such as gradient resonance between tasks and mixed use of parameters. Through the adaptive optimization of the feature subspace distribution, the supervision of task coordination consistency, and the dynamic adjustment of the loss function in this invention, it is possible to monitor and automatically suppress conflicts between multi-objective outputs caused by redundancy or coupling of parameters, process design, etc., so that gradient updates have stronger directionality and anti-interference ability;

[0025] (4) The data preprocessing, grouping, and attribution design of this invention are highly automated, which can efficiently adapt to the actual application needs of large-scale, multi-type PCB impedance samples and complex working conditions. The solution has strong parameter space expansion capability, can adapt to multi-target impedance prediction under different transmission line types / frequency bands / environments, and supports engineering integration needs such as batch archiving of samples and structured knowledge feedback. Compared with traditional methods, the model generalization error is reduced, effectively supporting diverse scenarios such as product design and development, quality monitoring, and process optimization;

[0026] In summary, this invention systematically solves the problems of model accuracy bottleneck, physical consistency and interpretability in high-dimensional multi-objective collaborative impedance prediction, significantly improves prediction performance and industry trust, expands the practicality and intelligence level of modeling scenarios, and has extremely high technological innovation value and industrial application prospects. Attached Figure Description

[0027] Appendix Figure 1 This is the main flowchart of a neural network-based method for predicting and evaluating PCB impedance characteristics.

[0028] Appendix Figure 2 This is a sub-flowchart of a neural network-based method for predicting and evaluating PCB impedance characteristics.

[0029] Appendix Figure 3 This is a sub-flowchart of a neural network-based method for predicting and evaluating PCB impedance characteristics. Detailed Implementation

[0030] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0031] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0032] As used herein, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms "comprises", "comprising", "includes" and / or "including", or the like, when used in this specification, specify the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof. Also, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0033] As shown in the accompanying drawings, Figure 1 The application provides a neural network-based PCB impedance characteristic prediction and evaluation method, specifically comprising:

[0034] S1: Collect high-dimensional PCB impedance historical data samples containing different structural parameters, process parameters and environmental parameters, and record the corresponding task label information of the samples for multiple types of transmission lines, frequency bands and temperature conditions.

[0035] S2: Perform normalization, outlier rejection and condition label division processing on the collected high-dimensional PCB impedance historical data samples to generate data preprocessing results that meet the multi-task collaborative modeling requirements.

[0036] S3: Based on the normalized high-dimensional PCB impedance historical data samples, use a parameter attribution algorithm to divide the feature subspace, group the multi-dimensional features that dominate the impedance output of different transmission line types, frequency bands and environmental conditions, and form a parameter allocation scheme for task-specific modeling.

[0037] S4: Input the parameter allocation scheme into a multi-branch neural network modeling architecture, configure a shared bottom feature extraction layer, and for each type of transmission line and target impedance output, build a multi-target prediction branch designed with a task-specific decoupling layer.

[0038] S5: Apply an interpretable regularization term with physical prior guidance in the task-specific decoupling layer, control the regularization strength to limit the influence of some process parameters on only the specified target impedance output, and realize feature decomposition and correlation constraint of each target prediction branch.

[0039] S6: Jointly train the regularized multi-branch neural network modeling architecture using a multi-task loss function and an inter-task collaborative consistency supervision mechanism, measure and adjust the physical consistency or known redundant relationship between different target prediction results, and dynamically optimize the training process.

[0040] S7: During the joint training process, the target prediction accuracy and inter-task coordination consistency indicators are monitored in real time based on the validation set, and the feature subspace decoupling proportion and regularization constraint strength are dynamically adjusted to achieve adaptive optimization of inter-task accuracy coordination and model stability.

[0041] S8: Using the optimized multi-branch neural network modeling architecture, the target impedance prediction results and their uncertainty intervals under each task are inferred and output for new input data of different structural parameters, process parameters and multi-condition data.

[0042] S9: Based on the target impedance prediction results, feature importance decomposition information and physical consistency analysis indicators output by inference, the model performance is evaluated, and knowledge feedback and multi-target performance monitoring reports for PCB design or manufacturing process improvement are generated.

[0043] The step S1: Collect high-dimensional PCB impedance historical data samples containing different structural parameters, process parameters and environmental parameters, and record the corresponding task label information of the samples for multiple types of transmission lines, working frequency bands and temperature conditions. Specifically, it includes:

[0044] S1.1: Obtain historical PCB impedance measurement records as original data set input, collect high-precision impedance parameters of multiple batches of PCB samples with different structural parameters based on professional PCB impedance test system, to obtain complete data set including structural parameters such as transmission line type, width, thickness, dielectric constant, as well as manufacturing process parameters and typical environmental parameters (such as temperature, humidity).

[0045] For PCB samples containing different structural parameters, process parameters and environmental parameters, the input object is the historical PCB impedance measurement records and the original information of the process, structure and environment of the corresponding samples.

[0046] Using professional PCB impedance automatic test system, high-precision impedance data of mass batches of PCB physical samples are collected, and the structural parameters of each sample are collected simultaneously, including transmission line type, width, thickness, number of dielectric layers and dielectric constant.

[0047] Further, through the process parameter archiving management module, the manufacturing process parameter information of each sample is systematically obtained, including etching process batch number, pressing temperature, pressing pressure, soldering process category, etc., to ensure that each data entry has complete production history.

[0048] Using industrial environment monitoring and sampling equipment, real-time measurement and recording of environmental parameters corresponding to each collection process are carried out, including temperature, humidity, air pressure and other environmental variables, and all parameters are integrated into the same original data structure through data acquisition card or digital interface.

[0049] Through the data synchronization archiving algorithm (parameters: measurement timestamp, sample number mapping, parameter collection frequency), the impedance test results are automatically associated with the corresponding structure, process, and environmental parameters one by one, and a complete original data set for multiple types of tasks is generated.

[0050] Through high-precision impedance measurement and multi-dimensional parameter synchronization collection, complex and diverse sample physical information, manufacturing process information, and environmental condition parameters are integrated into a high-dimensional original data set in a structured manner, achieving full-quantity data coverage of impedance characteristics covering multiple types of transmission lines, multiple frequency bands, and multiple environmental / technological batches.

[0051] For example, data collection was performed on batch multi-layer board samples produced by a PCB manufacturing enterprise in the second quarter of 2023. Each group of samples was subjected to 4-port impedance measurement using an Agilent E5071C vector network analyzer. The measurement parameters included 50Ω, 90Ω target lines for specific types of microstrip lines, striplines, and differential lines, with measurement width values ranging from 0.2 to 1.2 mm, thickness ranging from 0.015 to 0.035 mm, and dielectric constant between 4.2 and 4.8. For each sample, the etching process batch number, pressing temperature (140-180℃), pressing pressure (10-15MPa), and measurement station room temperature (22-28℃), humidity (30-60%RH) were recorded. All parameters were automatically numbered and stored by the data collection terminal. A high-precision impedance sample data set of 2000 different structures, processes, and environmental parameters was obtained, greatly enriching the basis database for subsequent feature engineering and multi-task neural network collaborative modeling. Each data item has traceability, full-quantity, and multi-dimensional expansion capabilities.

[0052] S1.2: For the obtained structure parameter, process parameter, and physical environment parameter data set, parameter standard coding rules are used for format standardization to ensure that each dimension parameter (such as transmission line width / thickness / material, interlayer dielectric type, etching process batch number, environmental temperature / humidity, etc.) is stored under a professional unified naming system, laying the foundation for subsequent feature consistency processing and batch data processing output structured data objects.

[0053] The obtained structural parameters, process parameters and physical environment parameter data sets are standardized by parameter standard coding rules to realize the format standardization of multi-source raw data, and the numerical representation of each physical quantity and process parameter is standardized among different collection batches, devices and systems. The parameter standard mapping dictionary (parameters: transmission line type code, width specification value, thickness specification value, medium type standard number, etching process batch number, environment temperature and humidity code, etc.) is used to renumber all structural parameters (such as LineType, W, T, Er), process parameters (such as EtchBatch, LaminateTemp, LaminatePress), and environmental parameters (such as EnvTemp, EnvRH) according to the international or industry standard naming system, so as to conform to the unified data label template.

[0054] Further, through the structured data analysis and reconstruction algorithm (parameters: multi-level parameter field mapping table, abnormal field verification rule), the non-standardized naming, inconsistent numerical types and unit mismatch phenomena caused by heterogeneous collection systems, manual entry or collection batch changes in the above data fields are automatically identified, and automatically converted to standardized format according to the standard parameter interface protocol. For example, for the width (W) or thickness (T) field, the unit (such as mm, mil, μm) is automatically checked, and unit conversion and decimal precision uniformity are performed according to the preset target unit, and a parameter list with unified dimensions is output.

[0055] Through the data field uniqueness verification and consistency identification generation algorithm (parameters: sample number specification, field uniqueness constraint), a unique structured primary key is generated for each parameter set to realize cross-platform and cross-batch data traceability, and a unique code index is configured for each parameter data.

[0056] Further, based on the parameter standardization processing result, a multi-dimensional parameter structured object output algorithm is used to encode all structural parameters, process parameters and environmental parameters into standardized data objects (such as JSON, table or database entry), and generate a structured parameter set for subsequent batch feature processing and efficient data interface docking.

[0057] Through the above parameter standard coding standardization processing, the multi-source and multi-batch structural parameters, process parameters and physical environment parameters are converted into completely consistent, dimensionally uniform and nominally standardized structured data objects, which lay a solid data foundation for feature consistency processing and subsequent batch data automation processing flow, and realize the structured, standardized management and efficient expansion capability of large-scale high-dimensional PCB impedance data.

[0058] Exemplarily, in some embodiments, for 2000 groups of multi-layer PCB sample data collected in the second quarter of 2023, the parameter standard encoding dictionary is used to convert the structure parameters (such as the transmission line type is uniformly coded as “LineType” field, and the numerical value is coded as an industry standard short code, for example, MBL represents a microstrip line), width and thickness parameters of all samples into decimal form (with 3 decimal places) in millimeters (mm), and the dielectric constant parameter is stored as a “four decimal place” floating point. For process parameter fields such as etching process batch number, pressing temperature, pressing pressure, all are archived by standard database main index numbering method, the process batch number uses 8-digit code, the temperature is uniformly Celsius (℃), the pressure is uniformly megapascal (MPa), and is bound one by one with the sample number. The environmental parameters (such as temperature, humidity) are standardized as decimal type, taking three decimal places, named “EnvTemp” and “EnvRH” respectively. All parameters are output as structured JSON objects by an automated data processing engine and imported into a MySQL database for unified storage. During this process, the system automatically detects that about 30 parameter items have been confused (such as the width is mistakenly filled as inches), and automatically corrects and standardizes through unit conversion rules. The final output of the structured parameter data set is automatically numbered, thereby supporting all batch data interfaces and task calls such as subsequent feature engineering and neural network model task allocation, achieving a data standardization rate of 100%, a parameter consistency rate of more than 99.9%, and providing a solid data support and technical guarantee for the consistency and controllability of multi-target high-dimensional neural network collaborative modeling.

[0059] S1.3: Based on the standardized structured data objects of structure parameters and process parameters, combined with the impedance values and their accuracy information output by the impedance measurement equipment, data label mapping is performed to associate each group of data samples with their corresponding impedance measurement types (such as specific transmission line types, frequency band distribution, temperature / humidity test conditions), generating labeled high-dimensional PCB impedance historical sample data set with task labels, laying a template foundation for multi-task classification and parameter attribution analysis.

[0060] S1.4: For the labeled high-dimensional PCB impedance historical sample data set, collect and summarize multi-label multi-task attributes such as different transmission line types (such as microstrip line, strip line, differential line), multi-frequency bands (such as low frequency, medium frequency, high frequency) and environmental conditions (such as normal temperature, low temperature, high temperature, high humidity), and use hierarchical classification method in professional PCB design field to group samples according to task labels to form multi-task label distribution reference set for subsequent feature attribution analysis.

[0061] S1.5: For the multi-label grouped high-dimensional PCB impedance historical sample data set, a data consistency verification algorithm and a missing value completion strategy are used to conduct integrity audit on key structural parameters, process parameters and task label items, automatically correct or fill in missing items, and output a high-dimensional, open multi-task impedance historical sample basic data set without missing, accurate label, and standardized parameters, which is directly input to the downstream normalization and feature engineering link and establishes a reliable data source for the whole process chain conversion.

[0062] The step S2: performing normalization, outlier rejection and working condition label division processing on the collected high-dimensional PCB impedance historical data samples to generate data preprocessing results meeting the multi-task collaborative modeling requirements. Specifically, it includes:

[0063] S2.1: Apply standardization normalization algorithm to the collected high-dimensional PCB impedance historical data samples to normalize the multi-dimensional input features such as structural parameters, process parameters and environmental parameters, to obtain data standardization results with unified numerical domain and no dimension interference, and ensure balanced influence of parameters on model convergence in subsequent parameter attribution and feature extraction steps.

[0064] For the high-dimensional PCB impedance historical sample data set after standardization encoding, a standard normalization algorithm (parameters: full-amount structural parameters, process parameters, environmental parameters) is used to realize unified adjustment of the numerical range of each input feature dimension.

[0065] Further, the normalization process is to convert each parameter to the [0, 1] interval by using the Min-Max normalization method, and then to use Z-score standardization for abnormal distribution parameters, and to output a normalized feature matrix in a unified format for all input features;

[0066] Further, by using the Min-Max normalization method (Min-Max Scaling, standard interval [0, 1]), each structural parameter (such as LineType, W, T, Er, etc.), process parameter (such as EtchBatch, LaminateTemp, LaminatePress) and environmental parameter (such as EnvTemp, EnvRH) is respectively transformed as follows:

[0067]

[0068] where x i is the original value of the i-th feature, min(x i ) and max(x i ) are the minimum and maximum values of the feature in the sample space, and x' i is the normalized dimensionless feature value.

[0069] Further, for the feature parameters that still have extreme outliers after normalization, the Z-score standardization algorithm (parameters: mean μ, standard deviation σ) is used to realize the standard normal distribution mapping of the sample features:

[0070]

[0071] wherein μ and σ are the mean and standard deviation of the feature F i , and x i is the standardized feature score.

[0072] Further, through the normalization processing module, the normalization mapping table of different dimension parameters under multiple batches and multiple task labels is integrated, and the processing result is output as a normalized feature matrix in a unified format, which is used for further feature analysis and modeling of the standardized data.

[0073] Through the layer-by-layer processing of normalization and standardization algorithms, the absolute comparability and dimensionless interference of the structure parameters, process parameters and environmental parameters in the numerical domain are ensured, the numerical stability of the parameter attribution algorithm and feature engineering processing in the downstream is guaranteed from the source, and the convergence efficiency and generalization ability of the model for high-dimensional feature space are improved.

[0074] For example, for 2000 groups of multilayer PCB impedance samples collected in the second quarter of 2023, the structure parameters (width W value 0.2-1.2mm, thickness T 0.015-0.035mm, dielectric constant Er 4.2-4.8), process parameters (pressing temperature 140-180℃, pressure 10-15MPa), and environmental parameters (measurement temperature 22-28℃, humidity 30-60%RH), the Min-Max normalization is used to linearly convert each field to the [0,1] interval. Taking the width parameter as an example, the original minimum value is 0.2mm and the maximum value is 1.2mm. After normalization conversion, 0.2mm is mapped to 0 and 1.2mm is mapped to 1, and other values are linearly mapped between them. Further, for the distribution skewness characteristics such as pressing temperature, the sample mean μ=160℃ and the standard deviation σ=14.14℃ are calculated, and the Z-score normalization is used to convert all parameters to standard distribution type data. Finally, a high-dimensional structured normalized feature matrix is output as a standardized input, which provides sufficient guarantee for data consistency and convergence in the subsequent anomaly value detection and multi-task neural network modeling stage.

[0075] S2.2: Based on the normalization result, use statistical outlier detection algorithm (such as IQR method or Z-score threshold method) to perform outlier rejection processing on each feature dimension to identify and exclude abnormal data points caused by equipment errors, abnormal working conditions or data acquisition defects, and generate a high-dimensional PCB impedance historical data cleaning set with high confidence, which eliminates abnormal interference for subsequent working condition label grouping and feature division.

[0076] S2.3: Input the high-confidence high-dimensional PCB impedance history data cleaning set into the label automatic mapping module, use the multi-label hierarchical mapping algorithm to standardize the coding and normalize the integration of the transmission line type label, frequency band label and temperature working condition label of each data sample, to obtain an information label matrix with strong consistency and adaptive multi-task task structure, supporting efficient task division and downstream label dependence elimination in the multi-task modeling process.

[0077] S2.4: Based on the label matrix and sample main index, perform sample multi-dimensional grouping processing algorithm, and automatically map the sample data with high confidence and label mapping completion to the multi-task collaborative modeling input structure body according to multiple dimensions such as transmission line type, frequency band category, temperature working condition, to generate data input subsets adaptive to multi-objective neural network common parameters and individual feature mining.

[0078] S2.5: Apply data consistency test and sample balance discrimination method to the aforementioned multi-task modeling input structure body, detect the distribution balance and coverage of the grouped samples in each task label dimension, and automatically generate data consistency feedback indicators to guide subsequent feature subspace division and task-specific network structure optimization, and realize quality closed-loop protection of data preprocessing results.

[0079] The step S3: based on the normalized high-dimensional PCB impedance history data samples, using parameter attribution algorithm for feature subspace division, grouping the multi-dimensional features dominated by different transmission line types, frequency bands and environmental condition impedance outputs, forming a parameter allocation scheme for task-specific modeling. Specifically includes:

[0080] S3.1: Based on the normalized high-dimensional PCB impedance history data samples, based on the process and structure parameter category information, perform parameter preliminary screening classification operation to obtain a preliminary parameter category set, and establish a basic feature distribution for subsequent parameter attribution analysis.

[0081] The input is the normalized high-dimensional PCB impedance history data samples, including the standardized structure parameters, process parameters and environmental parameter feature matrix.

[0082] The parameter category information screening method (parameters: structure parameter list, process parameter list, environmental parameter list) is used to realize the preliminary category classification of the input features.

[0083] Further, a structure parameter mapping dictionary is established through a category mapping table, and the normalized features are summarized into a structure parameter subset according to fields such as transmission line type, line width, line thickness, dielectric constant, etc., to realize the structure attribute clustering of high-dimensional features.

[0084] Furthermore, a process parameter regularization matching algorithm (parameters: process type identifier, batch number, temperature control parameter) is adopted to assign parameters such as etching batch number, pressing temperature, and pressurization pressure to a subset of process parameters and mark their process layer index.

[0085] Furthermore, the environmental parameter domain aggregation method (parameters: temperature range, humidity range) is executed to merge all environmentally related features (such as measured temperature and measured humidity) into the environmental parameter classification set, thereby achieving the isolation of environmental factors across the entire domain.

[0086] The system automatically detects and labels other unclassified parameters in samples using a category attribution list. An anomaly detection mechanism is employed to temporarily classify these parameters as features to be determined and output anomaly alerts to ensure the integrity of parameter classification.

[0087] Through the initial parameter screening and classification operation, the normalized feature matrix of the previous step is transformed into a preliminary set of parameter categories labeled with structural parameters, process parameters and environmental parameters, thus realizing the basic feature distribution for subsequent parameter attribution analysis.

[0088] For example, 2000 sets of impedance-normalized samples of multilayer PCBs collected in 2023 were analyzed. The structural parameters included the fields LineType, W (normalized from 0 to 1), T (0 to 1), and Er (0 to 1); the process parameters included EtchBatch, LaminateTemp, and LaminatePress (all normalized); and the environmental parameters included EnvTemp and EnvRH (all normalized). A parameter category segmentation script automatically aggregated LineType, W, T, and Er into a subset of structural parameters, categorized EtchBatch, LaminateTemp, and LaminatePress into a subset of process parameters, and aggregated EnvTemp and EnvRH into a group of environmental parameters. The detected abnormal parameter field "TestTime" was marked as a pending parameter feature because it did not belong to any of the above three categories. The parameter category set output by the classification results covered 8 structural parameters, 15 process parameters, and 6 environmental parameters. If TestTime was determined to be relevant by subsequent attribution algorithms, it could be re-merged or deleted to simplify the input. The above classification operation achieved an accuracy rate of 99.5%, establishing a unified and structured data foundation for subsequent parameter attribution analysis and feature subspace partitioning, greatly improving the scientificity and efficiency of multi-objective task-specific parameter allocation.

[0089] S3.2: Based on the obtained set of parameter categories, use parameter attribution algorithms (such as SHAP, feature importance ranking) to calculate the influence strength between each structural parameter, process parameter and each impedance output label, so as to obtain the specific attribution index of each parameter on the target impedance component.

[0090] For the normalized and labeled high-dimensional PCB impedance history data samples, a parameter attribution algorithm is used (parameter category set: structural parameters, process parameters, environmental parameters) to calculate the characteristic attribution indicators of each input parameter to the target impedance output.

[0091] Using the SHAP (SHapley Additive exPlanations) method, for each structural parameter, process parameter, and environmental parameter in the normalized feature matrix, the SHAP value contribution weight is calculated for each impedance output label, realizing the specific influence measurement of single parameter to multiple target outputs.

[0092] Through the feature importance sorting algorithm (such as Feature Importance statistics based on tree model), the ranking scores of all parameters in the parameter category set to each impedance component (such as specific impedance targets under different transmission line types and different frequency bands) are calculated, which supplements and verifies the SHAP attribution results.

[0093] Further, the joint influence strength between each feature parameter and multiple impedance outputs is modeled and estimated using multiple correlation analysis algorithms (such as multi-target correlation coefficient R and multi-variable linear regression coefficient), realizing the comprehensive measurement of the interaction between parameters and between parameters and multiple target outputs.

[0094] Based on the above SHAP value contribution weight, feature ranking score, and multi-target correlation analysis results, a structured feature attribution indicator matrix is generated, where the matrix elements represent the influence strength and direction of each parameter to each target impedance component.

[0095] Through the feature attribution indicator matrix, the attribution influence of each parameter in the parameter category set is represented as a vector, providing a numerical quantification basis for the subsequent optimization of multi-dimensional feature subspace, realizing the specific parameter identification for the target impedance output.

[0096] Through the parameter attribution algorithm chain, the structured feature input of the previous parameter category classification step is transformed by the system into the specific attribution indicators of each parameter to the target impedance component, realizing the multi-target fine measurement of feature attribution, and laying a quantitative foundation for the subsequent feature subspace grouping and decoupling modeling.

[0097] For example, for the 2000 groups of normalized PCB impedance samples, input the structure parameters (such as W = 0.45, T = 0.86), process parameters (such as LaminateTemp = 0.62, EtchBatch = 0.35), and environmental parameters (such as EnvTemp = 0.55, EnvRH = 0.44), and use the LightGBM regression model to perform multi-objective fitting on each impedance output label (such as microstrip line / high frequency / normal temperature, etc.), and calculate the SHAP value for the structure, process, and environmental parameters under different targets. For microstrip line high frequency impedance prediction, the SHAP mean value of width W is 0.54, the thickness T is 0.18, the laminate temperature LaminateTemp is 0.12, and the environmental temperature EnvTemp is 0.09, indicating that the width is the dominant feature. Further using the feature importance ranking based on decision tree, W ranks first, T ranks second, and LaminateTemp ranks third. The Pearson correlation analysis is applied between the parameters and the target impedance, and the correlation coefficient R between the width W and the high frequency microstrip line impedance is 0.81, and the LaminateTemp is 0.28. The final attribution result is the attribution index matrix, such as [W-0.54, T-0.18, LaminateTemp-0.12, EnvTemp-0.09]. When applied to stripline low frequency impedance prediction, the contribution of process parameters is significantly improved, represented by the attribution index matrix [W-0.27, T-0.31, LaminateTemp-0.33, EnvTemp-0.21]. The above attribution analysis realizes parameter-specific measurement, provides quantitative allocation basis for subsequent feature subspace division, and improves the input explanation and attribution accuracy of the multi-objective prediction process.

[0098] S3.3: Perform multi-dimensional feature subspace division operation on the normalized parameter space using specific attribution indicators, group the parameter set with the highest feature influence according to each target task (such as transmission line type, frequency band, and environmental working condition), and form a task-specific feature subspace set.

[0099] Based on the high-dimensional PCB impedance historical data samples after normalization, the input is the standardized feature matrix of structure parameters, process parameters, and environmental parameters, and the specific attribution indicator matrix between each parameter and each target impedance output generated by the parameter attribution algorithm (such as SHAP value, feature importance ranking, correlation coefficient analysis, etc.).

[0100] The multi-target specific attribution index screening method (parameters: attribution contribution threshold λ, attribution ranking TOP-N) is used to realize the influence measurement of each feature on each target task in the normalized parameter space. For each target task (including the impedance prediction components corresponding to the transmission line type, frequency band category and environmental working condition label), the attribution intensity of each feature parameter in the attribution index matrix is ranked, and the TOP-N feature set with significant influence is screened, and the value of N depends on the complexity of the target output and the modeling requirements.

[0101] Further, through feature influence aggregation analysis, the SHAP value, feature importance score and multi-target correlation parameter are integrated, and a weighted normalization method is used to quantitatively integrate various attribution indexes to form the joint influence score S of each feature parameter on each target output i,j :

[0102] S i,j =α1·SHAP i,j +α2·Imp i,j +α3·|Corr i,j |

[0103] Where SHAP i,j is the average SHAP value of feature i for target j; Imp i,j is the importance score of feature i for target j based on a specific model (such as a tree model); Corr i,j is the correlation coefficient between feature i and target j; α1, α2, α3 are the weighted coefficients set by experience, and satisfy α1+α2+α3=1. The above formula allows flexible tuning of the attribution evaluation system according to the actual project.

[0104] Further, for each target task, all normalized feature parameters are ranked from high to low according to the joint score S i,j , and the parameter set higher than the preset threshold λ is selected, or the top N items are truncated as the specific feature set F j of the target task, to realize the preliminary division of the feature subspace. Each F j set is the key influence parameter set of the dominant target j (such as a specific transmission line type / frequency band / working condition impedance output).

[0105] Further, through feature intersection and mutual information analysis, the obtained specific feature sets F j of each target are compared with each other to identify the completely specific parameter subset and the common parameter set with high overlap between multiple targets. The task-specific parameter independence measurement formula is used to calculate the parameter similarity matrix M j,k between tasks:

[0106]

[0107] where M j,k The closer to 0, the more independent the parameter subspace of task j and task k, the closer to 1, the more overlap. This index provides data basis for subsequent parameter coordination and decoupling analysis between tasks.

[0108] Through attribution strength screening, joint score aggregation, intersection analysis and subspace independence measurement, the normalized parameter space system is converted into a task-specific feature subspace set {F1, F2,... F t}, providing quantitative and structured support for subsequent independent modeling of each target task and task-specific network branch configuration.

[0109] Through multi-dimensional feature subspace division driven by specific attribution indicators, the parameter category set and attribution matrix are converted into a target-specific feature set that can be input into a multi-task neural network architecture, achieving optimal segmentation of the multi-target task parameter space, effectively reducing feature interference between tasks, and improving prediction accuracy and stability of the model in high-dimensional complex scenarios.

[0110] For example, for the microstrip line high-frequency impedance prediction task collected in the second quarter of 2023, the SHAP value of the structural parameter "line width W" in the input parameter attribution matrix is 0.52, the feature importance score is 0.58, and the correlation coefficient is 0.79. After joint scoring formula weighting (α1=0.4, α2=0.4, α3=0.2), S W,HF-MTL =0.4*0.52+0.4*0.58+0.2*0.79=0.589. Among all parameters, the key parameters with S i,HF-MTL >0.2 include W (0.589), T (0.37), LaminateTemp (0.21), forming the specific feature set F HF-MTL ={W, T, LaminateTemp} for the high-frequency microstrip line task. For the stripline low-frequency impedance prediction task, the top parameters are T (0.41), LaminateTemp (0.36), and EnvTemp (0.23). The intersection of the two task parameter sets F HF-MTL and F LF-SL is calculated, and the intersection ratio of the union M j,k is 0.5, indicating that the feature subspace partially overlaps but has task-specific parameters in this scenario. Using the above processing, the optimal specific feature parameter set corresponding to each type of transmission line / frequency band / working condition is finally output, achieving fine-grained feature configuration for downstream multi-branch modeling, and verifying the significant prediction coordination and accuracy improvement effect between tasks in actual model training.

[0111] S3.4: For the parameter subset with high overlap or coupling between tasks, use collaborative feature analysis (such as multi-target correlation coefficient, collaborative redundancy analysis) to determine the common parameter subspace, and implement attribution decoupling labeling for redundant or mutually exclusive features, ensuring that the parameter allocation scheme takes into account specificity and collaboration.

[0112] S3.5: Organize the final grouped task-specific feature subspace set and the common parameter subspace to form a parameter allocation scheme, and output it in the form of a structured parameter mapping table, providing a unified standard for subsequent feature input configuration and task-specific modeling of multi-branch neural network architecture.

[0113] The step S4: input the parameter allocation scheme into the multi-branch neural network modeling architecture, configure the shared bottom layer feature extraction layer, and for each type of transmission line and target impedance output, build a multi-target prediction branch with task-specific decoupling layer. Specifically includes:

[0114] S4.1: Analyze the high-dimensional parameter grouping results output by the parameter allocation scheme, and based on the parameter subspace obtained by the feature attribution algorithm, determine the feature set corresponding to each transmission line type and target impedance output to form the subspace feature mapping structure of the network input end.

[0115] The input is the high-dimensional parameter grouping result output by the parameter attribution algorithm, including the specific parameter subspace set allocated for each transmission line type and target impedance output, and the related structured feature category mapping table.

[0116] The parameter subspace analysis method (parameters: specific attribution index matrix, parameter category mapping table) is used to realize the automatic identification and classification of high-dimensional parameter set under each task target.

[0117] Further, through the feature subspace mapping algorithm (parameters: task label set, target impedance output type), the clustering and attribution of each parameter subspace to specific transmission line type and impedance output are realized, and the parameter space is divided into multiple data subsets for different tasks, and the input feature set corresponding to each target impedance branch is determined.

[0118] Further, the parameter-task allocation index matrix construction method (parameters: parameter allocation scheme, attribution score threshold) is used to automatically generate the mapping relationship table between parameters and target tasks, representing the attribution of each type of parameter to the corresponding transmission line type and impedance output, and realizing the consistency and traceability of parameter mapping.

[0119] Further, the application feature structure labeling mechanism (parameters: parameter ID, attribution level label, action classification code) is applied to assign structural attribute labels to each parameter feature in the mapping table, indicate the specific index and participation branch in the network input structure, and provide clear instructions for the task-specific feature path configuration of the network input end.

[0120] Through the parameter subspace mapping matrix and feature classification label, the high-dimensional parameter grouping result is normalized and converted to the subspace feature mapping structure of the network input end, providing a standardized feature entry for the subsequent multi-branch neural network modeling architecture, and realizing the unambiguous configuration of task-specific input.

[0121] For example, for 3240 multi-layer PCB impedance historical samples collected in the first half of 2023, the feature attribution algorithm is used to extract the structural parameter set {W(0.52), T(0.46), Er(0.17)} under the microstrip line-high frequency category and the process parameter set {LaminateTemp(0.21)}, and to distinguish the specific parameter set {T(0.41), LaminateTemp(0.36); EnvTemp(0.23)} under the stripline-low frequency category. According to the parameter attribution matrix, the parameter category is structured and coded as F_{HF-MTL}={W,T,Er,LaminateTemp}, F_{LF-SL}={T,LaminateTemp,EnvTemp}. Using the parameter mapping table, the following is generated: parameter W→ microstrip line-high frequency branch input, T→ all branch input, LaminateTemp→ high frequency microstrip line and low frequency stripline branch input, EnvTemp→ stripline-low frequency branch input. Through the structured label mechanism, the task classification and input index of each parameter are labeled, such as [W, HF-MTL, idx=1], [T, ALL, idx=2]. The above processing forms a multi-branch subspace feature mapping structure for the network input end, realizes the accurate grouping and unambiguous input of the key parameters required by each prediction branch of the network, improves the specificity and synergy of multi-branch modeling, and supports the efficient implementation of the downstream shared feature extraction and decoupled modeling link.

[0122] S4.2: Take the parameter set of each feature subspace in the parameter allocation scheme as input, and perform a unified feature extraction process on the high-dimensional PCB parameter features through the deep neural network bottom weight sharing mechanism to generate a shared bottom feature tensor as the common input basis for subsequent multi-objective decoupled modeling.

[0123] The input is the high-dimensional parameter grouping result output by the parameter allocation scheme, which includes the parameter set of each feature subspace after being filtered and structured based on the task-specific attribution index, covering information for different transmission line types and target impedance outputs, and all parameters have been uniformly formatted and normalized.

[0124] The deep neural network bottom layer weight sharing mechanism (parameters: feature subspace set, shared convolution or full connection module depth L, activation function type, normalization strategy) is adopted to realize unified feature extraction of high-dimensional PCB parameter features.

[0125] Further, through the feature vector splicing algorithm (parameters: feature subspace input matrix), the structure parameters, process parameters and environment parameter subspaces allocated by each task are assembled in the main index order to form a multi-dimensional input tensor as the standard input data format of the bottom neural network.

[0126] Further, the deep feature abstraction algorithm (such as multi-layer nonlinear mapping network, parameters: number of layers L, number of nodes N, activation function ReLU or GELU) is adopted to perform progressive feature mapping and nonlinear activation on the above input tensor, capture the complex interaction between high-dimensional parameters, and output a global shared bottom feature tensor T shared .

[0127] Further, the feature normalization and batch normalization processing mechanism (parameters: normalization layer position, BatchNorm / LayerNorm type) is applied to T shared The normalization and steady-state adjustment are performed in each neural network batch training to improve the consistency and convergence speed of feature extraction.

[0128] Further, through the feature redundancy compression and principal component screening algorithm (parameters: principal component contribution rate threshold, compression ratio), the dimensionality reduction optimization is performed on T shared The main feature vectors with cumulative contribution rate higher than the specified threshold are retained to further improve the efficiency and generalization ability of subsequent multi-objective decoupling modeling.

[0129] Through the above algorithm chain, each subspace parameter set in the parameter allocation scheme is taken as input and efficiently converted into a unified and high-representation shared bottom feature tensor, providing high-quality common input for subsequent multi-objective decoupling branch directional feature mapping, realizing feature collaboration and resource sharing between different task modeling, and reducing the gradient conflict risk caused by feature redundancy.

[0130] For example, the input is the 2023 version of the PCB impedance characteristic data set, the structure parameters (such as W = 0.55, T = 0.76, Er = 0.34) selected by the parameter allocation scheme, the process parameters (such as EtchBatch = 0.41, LaminateTemp = 0.63), and the environmental parameters (such as EnvTemp = 0.48, EnvRH = 0.33) form an input matrix, and each data input has a length of 20 dimensions. A deep neural network with weight sharing structure is used, with three hidden layers, each with 128, 64, and 32 nodes, respectively, and the activation function uses ReLU. All input data batches are first passed through the feature normalization layer (BatchNorm) with a batch training size of 64. After the data enters the three-layer fully connected network, it undergoes nonlinear transformation and multiple layer activation, and the output is a final 32-dimensional shared feature tensor. The feature tensor is further analyzed by PCA, with a principal component contribution rate threshold of 95%, and finally compressed to 16-dimensional principal component features. The test shows that this processing flow can effectively reduce feature redundancy while ensuring that the main physical and process information is not lost, reducing the model convergence round by 30%, and the prediction stability of the subsequent multi-objective neural network branch is significantly improved.

[0131] S4.3: Based on the shared underlying feature tensor, an independent multi-objective prediction branch is constructed for each type of transmission line and target impedance output, and a task-specific decoupling layer is inserted into the branch structure to realize targeted task-specific feature decomposition by mapping different feature subspace tensors.

[0132] S4.4: A multi-head parallel mapping structure is used for the output of each task-specific decoupling layer to further refine the subspace coupling relationship between the transmission line type and impedance output branch and the input parameters, generate sub-target feature representation, and provide a structured basis for subsequent interpretable regularization and task differentiation constraints.

[0133] S4.5: Integrate the task-specific decoupling outputs of the multi-objective prediction branch into the overall network modeling architecture to integrate the cooperative relationship regulation interface, provide a refined structural entry for subsequent physical prior constraint application and task consistency supervision, and realize the organized output of multi-objective feature collaborative reasoning.

[0134] The step S5: Apply the interpretable regularization term of the physical prior guidance in the task-specific decoupling layer to limit the influence of part of the process parameters on the specified target impedance output by adjusting the regularization strength, and realize the feature decomposition and correlation constraint of each target prediction branch. Specifically, it includes:

[0135] S5.1: Based on the feature subspace partitioning results, select a set of process parameters with high physical correlation to the target impedance output as the input factors of the task-specific decoupling layer, construct a feature decomposition input matrix through parameter grouping mapping to support the physical interpretability design of the subsequent decoupling layer structure.

[0136] The input is the parameter grouping structure obtained by the parameter attribution algorithm and feature subspace partitioning, including the high-dimensional structure parameters, process parameters and environmental parameter set corresponding to each target impedance output, as well as their attribution strength and physical correlation label.

[0137] Using the physical correlation screening method (parameters: feature attribution strength threshold, physical prior knowledge base), process parameter screening is realized for each target impedance output, and process parameters with attribution strength higher than the threshold are prioritized to preliminarily select a set of process parameters with significant physical effect on the target impedance.

[0138] Further, through the parameter grouping mapping algorithm (parameters: target branch category, parameter attribution label), the screened process parameters are grouped according to their physical properties and attribution direction, and each type of process parameter is assigned to the corresponding task-specific decoupling layer, providing structured grouping basis for subsequent input matrix assembly.

[0139] Further, the feature decomposition input matrix construction algorithm (parameters: parameter grouping result, sample main index, parameter attribution matrix) is used to vectorize and encode the grouped process parameters, and assemble them into a feature decomposition input matrix with task classification according to the input format of each target impedance output task.

[0140] Further, through the feature label enhancement mechanism (parameters: physical attribution label, impact weight score, parameter unique index), each process parameter feature in the feature decomposition input matrix is attached with physical correlation flag, attribution value, network input index and other attribute labels, realizing the interpretability and traceability enhancement of each field in the input matrix.

[0141] Through feature subspace physical correlation screening, parameter grouping mapping and feature decomposition input matrix construction, the process parameter set to be input into the task-specific decoupling layer is converted into a feature decomposition input matrix that meets the physical interpretation requirements and is structured and clear, providing standardized input for downstream decoupling mapping and interpretable regularization processing, and realizing the physical interpretability design goal of the task-specific decoupling layer.

[0142] Exemplarily, for the precision impedance test samples collected from high multilayer PCB microstrip lines and striplines, the process parameter set {EtchRate (attributed strength 0.41), LaminateTemp (0.38), EtchBatch (0.19)} is highly related to the high-frequency impedance of the microstrip line through characteristic attribution analysis, and the main related parameters for the low-frequency impedance of the stripline are {LaminateTemp (0.36), EnvTemp (0.29)}. The physical correlation screening threshold is set to 0.15, and the above parameters are respectively assigned to the decoupling layer of the microstrip line high-frequency branch and the stripline low-frequency branch. Through the parameter grouping mapping algorithm, EtchRate and LaminateTemp are input into the microstrip line-high-frequency branch, and LaminateTemp and EnvTemp are input into the stripline-low-frequency branch. The characteristic decomposition input matrix is constructed, wherein the parameter field of each sample contains [EtchRate_idx1, LaminateTemp_idx2, EtchBatch_idx3] (microstrip line branch) and [LaminateTemp_idx1, EnvTemp_idx2] (stripline branch), and each field is attached with a physical attribution score and a physical label. After this processing, the characteristic decomposition input matrix of the input decoupling layer has high physical interpretability and clear branch attribution, and the mapping relationship between the upstream physical attribution and the parameter coding process can be traced without ambiguity. The matrix is used for subsequent decoupling mapping and regularization constraints, which can significantly improve the physical consistency of the parameter effect interpretation and the network traceability, and the consistency correlation coefficient between the decoupled parameter allocation and the physical measurement results in the verification scene can be improved to more than 0.93, meeting the high requirements of parameter attribution transparency in engineering applications.

[0143] S5.2: Apply a task-specific decoupling transformation operator to the characteristic decomposition input matrix to realize explicit allocation of the process parameter set in each target impedance prediction branch, map the mixed characteristics of the parameters to the corresponding target branch according to the prior physical relationship, and form a controlled characteristic path matrix between multiple target tasks.

[0144] The input is the characteristic decomposition input matrix obtained according to the parameter attribution algorithm and the physical correlation screening, including the process parameter subset assigned to each target impedance prediction branch and its vectorized coding.

[0145] A task-specific decoupling transformation operator (parameters: target branch category, characteristic decomposition input matrix, decoupling mapping weight matrix) is used to realize the dispatch of the process parameter set in the multi-target task network structure, and the mixed characteristics of each process parameter are mapped from the global input to the specified target impedance output branch according to the physical attribution label and the branch setting.

[0146] Further, through the product operation of the decoupling mapping weight matrix, the process parameter characteristic input of each task branch is calculated:

[0147]

[0148] where F (k) is the task-specific decoupling eigenvector of the kth target branch, is the task-specific decoupling transformation weight matrix of the kth branch, X in is the normalized eigendecomposition input matrix.

[0149] Further, by adjusting the elements in F according to physical priors and attribution weights, only the parameters with physical relevance higher than a threshold have significant impact on F (k) , and the pathway weights of other parameters tend to zero, realizing controlled eigendecomposition.

[0150] Further, a parameter distribution sparsity regulation strategy is adopted, such as imposing an element sparsity regularization term on F to ensure that the parameter cross-weights across target task branches are minimized, avoiding information interference caused by parameter mixing between different prediction branches of the model.

[0151] Further, the mapping results of all target branches are integrated to construct a controlled feature pathway output matrix between multiple target tasks:

[0152] F All = [F (1) , F (2) ,..., F (K) ]

[0153] Through the above processing, the attribution pathway of each process parameter feature in the multi-branch network is clearly defined in the mapping weight matrix, realizing controlled connection between parameters and target branches, and laying a fine structure foundation for subsequent physical constraint imposition and interpretable regularization processing.

[0154] Through the application of task-specific decoupling transformation operators, the structured process parameter set generated in the previous step is mapped, distributed, and converted into specific feature pathways in the multi-target network architecture, outputting a feature pathway matrix with branch mapping weights, realizing the directional contribution of process parameters to target impedance output, and improving the interpretability and controllability of model eigendecomposition.

[0155] For example, for the microstrip line-high frequency impedance prediction branch, the process parameters EtchRate (physical attribution strength 0.41) and LaminateTemp (0.38) are assigned, and for the stripline-low frequency branch, LaminateTemp (0.36) and EnvTemp (0.29) are selected. Set the eigendecomposition input matrix X inwith dimension [batch, 5] (including EtchRate, LaminateTemp, EtchBatch, EnvTemp, PressIndex) to branch weight matrix respectively acting on X in The high-frequency microstrip branch feature vector F (HF-MTL) composed of EtchRate and LaminateTemp, and the low-frequency stripline branch feature vector F (LF-SL) composed of LaminateTemp and EnvTemp, and other parameter channel weights are 0. L1 sparse regularization is used to further constrain the weight distribution, prompting the inactivation of parameter channels unrelated to the target. Model verification shows that after this mapping, the response of the microstrip high-frequency impedance branch to EtchRate parameter changes is significantly improved (the attributable correlation coefficient is increased from 0.74 to 0.92), and the stripline low-frequency branch shows obvious independent sensitivity to EnvTemp. The model branch response has no obvious coupling, and the feature decoupling accuracy reaches 96%, meeting the requirements of engineering explainability and physical controllability in high-dimensional multi-target impedance prediction scenarios.

[0156] S5.3: Introduce a physical prior constraint function to impose soft constraints on the feature channel matrix according to the parameter-impedance physical causal relationship in actual PCB design and manufacturing processes, so that only specified target impedance output branches are activated by partial process parameters, reducing irrelevant feature interference.

[0157] S5.4: Embed an interpretable regularization term in the above soft constraint channel to use L1 regularization or physical prior guided norm penalty to implement sparse modulation of parameter influence weight distribution, prompting the model to automatically exclude process parameters unrelated to the target impedance output, and condense the minimum necessary feature set.

[0158] S5.5: Dynamically adjust the strength parameter of the interpretable regularization term, and real-time optimize the regularization influence based on the loss function feedback during the training process, to achieve the optimal balance between feature interpretability and model prediction accuracy, and output the final task-specific decoupling structure for feature correlation constraint between multiple target tasks.

[0159] The step S6: The regularized multi-branch neural network modeling architecture is jointly trained using a multi-task loss function and a task inter-consistency supervision mechanism to measure and adjust the physical consistency or known redundancy relationship between different target prediction results, and dynamically optimize the training process. Specifically, it includes:

[0160] S6.1: The regularized multi-branch neural network modeling architecture is inputted with multi-task feature samples, and the feature subspace input of each task is mapped to the corresponding prediction branch through the task label routing mechanism to realize the target distribution of multi-task features and lay the foundation for subsequent task-specific loss calculation.

[0161] S6.2: Based on the task-specific decoupling layer output and each target impedance branch, a multi-task loss function is adopted, including mean square error loss, L1 / L2 regularization loss and interpretable regularization term, to jointly optimize each task impedance prediction and feature distribution scheme to obtain multi-target basic prediction ability and parameter constraint ability.

[0162] The input is the high-dimensional feature subspace sample distributed to each target prediction branch through the task label routing mechanism, the regularized multi-branch neural network modeling architecture and the parameter distribution scheme.

[0163] A multi-task loss function design method (parameters: number of each target prediction branch, target type of each branch, loss function weight coefficient) is adopted to realize joint loss measurement of all target impedance outputs. The loss function structure is composed of mean square error loss (MSE), L1 regularization loss, L2 regularization loss and interpretable regularization term to form a multi-target loss total function, which is specifically expressed as:

[0164]

[0165] Where K is the total number of task branches, λ k is the loss weight of the kth target, is the mean square error loss of the kth target branch, L L1 is the global L1 regularization term, L L2 is the global L2 regularization term, L expl is the interpretable regularization term, and μ, γ, δ are the adjustment coefficients of the corresponding terms.

[0166] Further, the basic prediction accuracy of each task branch is evaluated by the mean square error loss (MSE) calculation method (parameters: target branch prediction output, actual measured impedance label), and the specific calculation formula is:

[0167]

[0168] Where N k is the sample size of the kth task branch, is the true impedance output of the ith sample of the kth task branch, is the corresponding predicted value.

[0169] Further, by the global L1 / L2 regular loss calculation method (parameters: model parameter vector, regularization coefficient), the global norm distribution of neural network parameters is respectively constrained to slow down the overfitting phenomenon, wherein

[0170]

[0171] w j represent all trainable weight parameters.

[0172] Further, by the interpretable regularization term design method (parameters: feature influence weight distribution, physical prior sparsity target), the influence weight sparsity modulation of the input parameters of each target prediction branch is realized to promote the weight of the process parameters irrelevant to the target impedance to converge to zero. This term can be defined as:

[0173]

[0174] wherein, S (k) is the parameter index set having a physical direct correlation with the kth impedance output, is the parameter weight corresponding to the kth branch.

[0175] Further, by the joint optimization algorithm (parameters: batch size, learning rate, momentum coefficient, loss function feedback), the joint backpropagation training of the multi-branch neural network is performed to update all parameters and feature allocation weights, and the latest model weight set of each branch and the sub-target loss value are output.

[0176] By the joint optimization and regularization processing of the above multiple loss functions, the prediction error of the model for each target task, the global structure generalization ability, and the interpretable constraint of physical correlation are included in the same evaluation system to realize the collaborative improvement of the multi-target basic prediction ability and the parameter constraint ability.

[0177] For example, in the three-task parallel PCB impedance multi-target prediction scene, the input sample batch is 64, the MSE weighting coefficients of each branch are respectively set as λ1=0.4, λ2=0.3, λ3=0.3, and the regularization term coefficient is set as μ=10 -4 (L1), γ=10 -5 (L2), δ=10 -3 (interpretable regularization). For the kth branch sample in the batch, the actual impedance label is 45.2 and the prediction output is 44.9, and the corresponding MSE is (45.2-44.9) 2 / 64≈1.4×10 -2 . After weight attribution determination, the number of L1 norm activated parameters is reduced to 38% of the initial parameter amount, and the irrelevant parameter influence weight converges to 10 -4The model three-target comprehensive prediction MSE is reduced from the initial 0.187 to 0.022 after 20 rounds of joint training, and the explainability improvement metric (corresponding to the consistency rate of the process dominant parameters and the physical labels) is increased from 0.71 to 0.93, meeting the requirements of high-dimensional multi-target high-precision and feature explainability.

[0178] S6.3: A weight adjustment mechanism is applied to each target loss component of the multi-task loss function, and the task weight is adjusted in real time through a dynamic weight distribution algorithm to balance the prediction accuracy of each target impedance branch and resolve the gradient conflict in the multi-target optimization process, so that the network training maintains consistency and collaboration in multi-target prediction under high-dimensional parameters.

[0179] S6.4: The prediction outputs of all target impedance branches are collected after each training batch, and a task inter-consistency supervision mechanism is used to calculate the inter-task physical consistency evaluation index and conflict loss term based on the known physical coupling relationship or redundant constraints, to flexibly adjust the loss function and guide the model parameters to converge to the optimal direction of physical consistency in real time.

[0180] S6.5: Based on the collaborative consistency evaluation results during training, a multi-task decoupling rate and explainability regularization coefficient automatic adjustment algorithm is executed to dynamically optimize the feature subspace partition ratio and regularization constraint strength, achieving adaptive tuning of the model architecture and continuous improvement of the stability of inter-task prediction.

[0181] S6.6: After completing a complete joint training, based on the model parameters and the output of each task branch, periodic backtracking optimization is performed to assist in stabilizing the sub-tasks that are difficult to converge by replaying the historical gradient trajectory and the evolution process of the inter-task loss, and finally output the overall optimized multi-branch neural network modeling architecture weight parameters.

[0182] The step S7: During the joint training process, the target prediction accuracy and the inter-task collaborative consistency index are monitored in real time based on the validation set, and the feature subspace decoupling ratio and the regularization constraint strength are dynamically adjusted to realize the adaptive tuning of the accuracy collaboration and the model stability between tasks. Specifically, it includes:

[0183] S7.1: Based on the multi-branch neural network joint training stage, the target impedance prediction results of the validation set under the current model parameter state are collected, and the real-time evaluation of the prediction accuracy of each target output is performed using a special performance measurement algorithm to obtain the target impedance prediction accuracy index.

[0184] S7.2: The inter-task consistency determination algorithm is applied to the simultaneously obtained target impedance prediction results to quantify the inter-task collaborative consistency index, measure the physical consistency relationship, redundant verification data or mutual interference behavior of each prediction branch after the task-specific decoupling layer, and obtain the inter-task collaborative consistency criterion.

[0185] S7.3: Based on the target impedance prediction accuracy index obtained by real-time monitoring and the inter-task collaborative consistency criterion, the decoupling ratio of the feature subspace decoupling layer is dynamically optimized and adjusted, the decoupling parameter search algorithm is used to adjust the allocation weight of the high correlation feature, and the feature decomposition structure is optimized to improve the relative independence and accuracy of each target prediction.

[0186] Based on the target impedance prediction accuracy index obtained by the verification set and the inter-task collaborative consistency criterion, the feature subspace decoupling layer dynamic optimization adjustment algorithm is used to realize real-time adjustment of the allocation weight of the high correlation feature. The performance-driven decoupling parameter search method (parameters: task accuracy index, collaborative consistency score, feature influence matrix) is used to perform elastic weight allocation on each feature subspace, automatically search for the optimal attribution ratio of high correlation features among each target task, and realize accurate demarcation of task-specific features and shared features.

[0187] Further, through the adaptive gradient adjustment mechanism (parameters: loss function gradient, feature contribution degree), the parameter weights in the feature subspace are gradually adjusted, the accuracy fluctuation and collaborative consistency change are responded in real time, and the optimal independence of the features between tasks is ensured. The following feature allocation weight update formula is used:

[0188]

[0189] wherein, is the new allocation weight of the kth feature subspace, is the weight of the previous cycle, η is the step size coefficient, E k is the current accuracy of the kth task, C is the collaborative consistency criterion, α and β are the accuracy and consistency adjustment coefficients, respectively, is the gradient of the loss function with respect to w k .

[0190] Further, the momentum adaptive screening algorithm (parameters: momentum term, historical weight fluctuation rate) is used to smooth the weight adjustment trend, suppress the sharp fluctuation of parameters caused by instantaneous noise, and stabilize the feature decomposition process. The following momentum modulation formula is used:

[0191]

[0192] wherein, is the current feature weight increment, μ is the momentum coefficient, is the smoothed feature allocation weight.

[0193] Further, based on the feature correlation threshold screening method, features with high correlation and large task overlap are grouped, merged, and finely redistributed. Parameters with a correlation above the threshold are preferentially assigned to the target branch with the greatest influence. Through the above parameter decoupling adjustment and dynamic feature redistribution, the feature decomposition structure is optimized to minimize the correlation between the input feature subspaces of each target task and optimally suppress information redundancy, while retaining necessary shared features to support task collaboration.

[0194] Through the above chain parameter adjustment and structure optimization, the precision and consistency feedback obtained by real-time monitoring of the validation set directly affect the feature decoupling framework, enabling mutual enhancement and conflict suppression of multi-target model outputs. Ultimately, a dynamically adjusted feature subspace distribution parameter set is output, providing efficient and controllable feature structure configuration for downstream regularization constraint optimization and network-wide collaborative training.

[0195] For example, in the multi-target PCB impedance prediction scenario, a shared feature tensor input with a length of 128 is assigned to three types of tasks: microstrip line (original allocation weight 0.45), stripline (original allocation weight 0.35), and differential line (original allocation weight 0.20). Through the decoupling parameter search algorithm, the allocation ratio is updated every 5 batches. The system detects that the microstrip line task precision E1 is 0.935 and the collaborative consistency criterion C is 0.92. After gradient and historical momentum adjustment, the current round of microstrip line task weight increases from 0.45 to 0.48 (Δ weight is 0.03, momentum coefficient μ = 0.7). According to the feature correlation measure, there are 10 coupled features above 0.85, which are automatically assigned to the microstrip line and stripline tasks for sharing. The remaining features are sorted according to their contribution and assigned to the corresponding branch based on loss feedback. The gradient resonance rate between tasks decreases by 12%, and the prediction accuracy of each target branch improves by 2.1% to 4.3%. The model output feature decomposition parameters are used for subsequent regularization weight intelligent adjustment, establishing a high dynamic matching degree feature decoupling configuration to support subsequent multi-target precision collaborative improvement.

[0196] S7.4: Based on the validation set performance, use the regularization adaptive adjustment strategy to increase or decrease the constraint strength of the interpretable regularization term. Through the regularization weight adjustment algorithm, the discriminative effect of process parameters on the specified impedance target is strengthened, and the influence of irrelevant disturbances between targets on the stability of model training is reduced, thereby optimizing the regularization weight distribution.

[0197] S7.5: With the updated feature subspace decoupling ratio and regularization constraint strength parameters as input, the multi-branch neural network task collaborative optimization training is performed again. Through model adaptive optimization iteration, the prediction accuracy of each target impedance and the inter-task collaborative consistency are simultaneously improved, and the optimal training parameter set is output for the next training cycle.

[0198] The step S8: using the optimized multi-branch neural network modeling architecture, for new input different structure parameters, process parameters and multi-working condition data, respectively infer the target impedance prediction results and their uncertainty intervals under each task.

[0199] Specifically includes:

[0200] S8.1: Perform feature normalization and format adaptation processing on the newly input structure parameters, process parameters and multi-working condition data, use the same normalization parameters and structured data templates as in the model training stage, so that the input feature space of the multi-branch neural network modeling architecture is strictly matched with the to-be-predicted sample, and the adaptability standard input feature matrix is obtained, providing consistent data support for subsequent model inference.

[0201] S8.2: Based on the adaptability standard input feature matrix, input the optimized multi-branch neural network modeling architecture, perform multi-level feature mapping and abstraction on the new input data through the unified bottom feature extraction layer, and obtain the global shared deep feature representation, thereby providing a unified basic feature for decoupling processing of each task branch.

[0202] S8.3: Input the global shared deep feature representation to each task-specific decoupling layer, use the decoupling mapping transformation algorithm to extract the specific feature subspace highly related to each target task (such as different transmission line types, frequency bands, temperature working conditions), generate multi-target subspace feature vectors, and minimize the gradient interference and feature redundancy between different target tasks.

[0203] Input the structure parameters, process parameters and multi-working condition data after standardization and format adaptation processing, obtain the global shared deep feature tensor through the bottom unified feature extraction layer, as the starting input of this step.

[0204] Use the task-specific decoupling mapping algorithm to input the global shared deep feature representation into the decoupling layer corresponding to each task branch, to realize efficient splitting of the feature vector in the parameter subspace.

[0205] Further, through the linear projection or sparse coding method based on the parameter allocation mapping rule, the feature subset with high physical correlation to each target task (such as a specified transmission line type, frequency band, and environmental working condition) is separated from the global feature tensor to form a three-dimensional target task-specific feature subspace.

[0206] Further, through the feature redundancy suppression algorithm (such as maximum relevance minimum redundancy mRMR), perform redundancy discrimination and optimal subset screening on the subspace features extracted by each task branch, to improve the feature independence between tasks and reduce the gradient interference caused by parameter coupling.

[0207] Further, the distribution relationship of each subspace is optimized by using a subspace orthogonal projection formula:

[0208] z k = P k · f g

[0209] wherein z k is a specific feature subspace vector of the kth target task, P k is a parameter projection matrix of the kth task, and f g is a global shared feature representation.

[0210] Through the above decomposition mapping, a highly sparse and optimally correlated multi-target subspace feature vector is finally generated for each target task as an input for downstream impedance prediction reasoning, realizing flexible control of feature modulation and minimization of cross-task interference.

[0211] Through multi-target subspace feature decomposition mapping, the global feature is effectively converted into a data vector serving each target task, realizing parameter interference suppression and feature cross noise reduction in a high-dimensional multi-target scene, and significantly improving the accuracy and stability of multi-task impedance prediction.

[0212] For example, for a certain PCB multi-target impedance prediction scene, the input is the structure parameters (such as transmission line width 0.18mm, thickness 0.035mm, dielectric constant 4.2) and process parameters (batch number A12, etching process PTH-3) and environmental conditions (temperature 85℃, humidity 60%) after standard normalization. The system generates a global feature tensor with a length of 256 through the underlying feature extraction layer, and the task branches include microstrip line, stripline and differential line. For the microstrip line branch, the parameter projection matrix P1 is selected (55 feature channels carrying coupled physical quantities are selected, with a sparsity of 0.8), the stripline branch projection matrix P2 is 63-dimensional sparse feature selection, and the differential line P3 is 92-dimensional feature mapping. Feature optimal subset screening is performed using mRMR, and 18% of the cross-task redundant features are removed. After orthogonal projection, the microstrip line task-specific feature subspace vector z1, the stripline feature subspace z2, and the differential line feature subspace z3 are obtained. In actual testing, the gradient resonance between tasks is reduced to 3.6%, and the correlation of input features of each branch is improved to 92%, laying a foundation for subsequent impedance component accurate reasoning and ensuring physical decoupling and optimal parameter allocation between multi-target prediction.

[0213] S8.4: Apply an output layer mapping function to the multi-target subspace feature vector, combine an interpretable regularization mechanism, and perform impedance prediction reasoning for each task branch, respectively, to obtain a prediction result set containing each target impedance component value and corresponding physical consistency parameters, realize automatic inference of target impedance under multi-task, and output the primary multi-task impedance prediction output.

[0214] S8.5: Based on the primary multi-task impedance prediction output, use uncertainty quantification algorithms (such as Bayesian inference or confidence interval estimation) to perform statistical confidence analysis on each target output, output the impedance prediction value and its uncertainty interval under each task, and provide risk-aware data support for downstream result evaluation and process decision-making.

[0215] The step S9: based on the target impedance prediction results of the inference output, the feature importance decomposition information and the physical consistency analysis indicators, the model performance is evaluated, and the knowledge feedback and multi-target performance monitoring report for PCB design or manufacturing process improvement are generated. Specifically, it includes:

[0216] S9.1: Based on the multi-target collaborative supervision label, the statistical performance evaluation method (such as mean square error, physical consistency metric) is used to calculate the prediction accuracy parameter set of each target impedance output to obtain the accuracy evaluation indicators of multi-dimensional impedance prediction results, and to provide basic data for subsequent performance decomposition analysis.

[0217] S9.2: Based on the target impedance prediction results and the corresponding input parameters, the feature importance decomposition algorithm (such as SHAP value decomposition, multi-task attribution ranking) is used to calculate and statistically analyze the key influence weight of each structure parameter, process parameter and environmental parameter on the target impedance prediction output of each task, and generate a parameter key importance score matrix to establish a parameter action causal chain for knowledge feedback analysis.

[0218] S9.3: Perform physical consistency analysis processing on the target impedance prediction results of each task and the feature importance decomposition information, use the inter-task correlation test and physical constraint checking mechanism to measure the physical redundancy consistency and key parameter reasonable interval of the prediction results, and output the physical consistency criterion of multi-target collaborative modeling to generate the input consistency parameter set for the performance monitoring report.

[0219] S9.4: The target impedance prediction accuracy parameters, feature importance score matrix and physical consistency criterion results are input into the knowledge feedback inference engine module, and based on the professional knowledge rule base, process improvement suggestions, parameter constraint suggestions and model robustness improvement guidance suggestions for different design conditions are generated, and structured knowledge feedback text is output to provide decision support for PCB design and manufacturing process optimization.

[0220] S9.5: Integrate the knowledge feedback text, multi-target performance monitoring parameters and structured evaluation data, and output the multi-target performance monitoring report through the automatic report generation component, which includes the target impedance prediction confidence interval of each transmission line type, the feature parameter contribution ranking, the physical consistency statistical distribution and the process improvement knowledge points, and realizes the full-process performance tracing and knowledge pushing closed loop of high-dimensional multi-target impedance prediction model.

[0221] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0222] The above description is only the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included in the protection scope of the present application.

Claims

1. A neural network-based PCB impedance characteristic prediction evaluation method, characterized by, The method comprises the following steps: S1, collecting high-dimensional PCB impedance historical data samples, recording corresponding task label information of the samples for multiple types of transmission lines, working frequency bands and temperature conditions; S2, performing preprocessing and condition label division processing on the high-dimensional PCB impedance historical data samples to generate a data preprocessing result; S3, based on the preprocessed high-dimensional PCB impedance historical data samples, performing feature subspace division, grouping multi-dimensional features dominated by different transmission line types, frequency bands and environmental condition impedance outputs, and forming a parameter allocation scheme; S4, inputting the parameter allocation scheme into a multi-branch neural network modeling architecture, configuring a shared bottom feature extraction layer, constructing a multi-target prediction branch with a task-specific decoupling layer for each type of transmission line and target impedance output; S5, applying an interpretable regularization term guided by physical prior in the task-specific decoupling layer, limiting part of the process parameters to only affect the specified target impedance output by adjusting the regularization strength; S6, jointly training the regularized multi-branch neural network modeling architecture using a multi-task loss function and a task intercoordination consistency supervision mechanism, measuring and adjusting the physical consistency or known redundancy relationship between different target prediction results, and dynamically optimizing the training process; S7, during the joint training process, based on the validation set, real-time monitoring of the target prediction accuracy and the task intercoordination consistency index, dynamically adjusting the feature subspace decoupling ratio and the regularization constraint strength; S8, using the optimized multi-branch neural network modeling architecture, inputting different structure parameters, process parameters and multi-condition data, and respectively inferring the target impedance prediction results and their uncertainty intervals under each task.

2. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, characterized in that, The step S8 further comprises: S9, based on the target impedance prediction results, feature importance decomposition information and physical consistency analysis index output by inference, evaluating the model performance, and generating knowledge feedback and multi-target performance monitoring report for PCB design or manufacturing process improvement. 3.The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, wherein, The step S1 specifically comprises: Obtain historical PCB impedance measurement records as raw data set input, based on professional PCB impedance test system, collect high-precision impedance parameters of multiple batches of PCB samples with different structure parameters, obtain complete data set of structure parameters, manufacturing process parameters and environmental parameters; Standardize the format of the obtained structure parameters, process parameters and physical environment parameter data set using parameter standard coding rules; Based on the standardized structured data objects of the structure parameters, the process parameters and the physical environment parameters, combined with the impedance value and its accuracy information output by the impedance measurement equipment, perform data label mapping to associate each data sample with its corresponding impedance measurement type one by one, and generate a labeled high-dimensional PCB impedance historical sample data set with task labels; For the labeled high-dimensional PCB impedance historical sample data set, collect and summarize the multi-label multi-task attributes of different transmission line types, multiple frequency bands and environmental conditions, and use the hierarchical classification method in the PCB design field to group the samples according to task labels to form a multi-task label distribution reference set; The data consistency verification algorithm and the missing value completion strategy are used on the high-dimensional PCB impedance historical sample data set after multi-label grouping, and the integrity of the key structure parameters, process parameters and task label items is audited, and the missing items are automatically corrected or filled, and the high-dimensional, open multi-task impedance historical sample basic data set without missing, accurate label and standardized parameters is output.

4. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, characterized in that, The step S2 specifically comprises: The standardization normalization algorithm is applied to the collected high-dimensional PCB impedance historical data samples, the multi-dimensional input features of structure parameters, process parameters and environment parameters are normalized, and a data standardization result is obtained; Based on the data standardization result, the statistical outlier detection algorithm is used to perform outlier rejection processing on each feature dimension, identify and exclude abnormal data points caused by equipment errors, abnormal working conditions or data collection defects, and generate a high-confidence high-dimensional PCB impedance historical data cleaning set; The high-confidence high-dimensional PCB impedance historical data cleaning set is input into the label automatic mapping module, and the multi-label hierarchical mapping algorithm is used to standardize the coding and normalize the integration of the transmission line type label, the frequency band label and the temperature working condition label of each data sample, and an information label matrix with strong consistency and adaptive multi-task task structure is obtained; Based on the label matrix and the sample main index, the sample multi-dimensional grouping processing algorithm is executed, and the sample data with high confidence and completed label mapping is automatically mapped to the multi-task collaborative modeling input structure body according to the transmission line type, the frequency band category and the temperature working condition, and a data input subset is generated; The data consistency verification and sample balance discrimination method is applied to the multi-task collaborative modeling input structure body, the distribution balance and coverage of the grouped samples in each task label dimension are detected, and a data consistency feedback index is generated.

5. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 4, characterized in that, The normalization processing is: the minimum-maximum normalization method is used to convert each parameter to the [0, 1] interval, and then the Z-score standardization is used for the distribution abnormal parameters, and all input features are output in a unified format to form a normalized feature matrix.

6. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, characterized in that, The step S3 specifically comprises: The normalized high-dimensional PCB impedance historical data samples are processed based on the process and structure parameter category information, and a preliminary parameter category set is obtained through parameter preliminary screening classification operation; According to the preliminary parameter category set, the parameter attribution algorithm is used to calculate the influence strength between each structure parameter, process parameter and each impedance output label, and a specific attribution index of each parameter to the target impedance component is obtained; The specific attribution index is used to perform multi-dimensional feature subspace division operation on the normalized parameter space, and the parameter set with the highest feature influence on each target task is grouped to form a task-specific feature subspace set; For the parameter subset with high overlap or coupling between tasks, the common parameter subspace is determined by collaborative feature analysis, and the redundant or exclusive features are marked by attribution decoupling; The final grouped task-specific feature subspace set and the common parameter subspace are organized to form a parameter allocation scheme, and output in the form of a structured parameter mapping table.

7. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, characterized in that, The step S4 specifically comprises: The high-dimensional parameter grouping result output by the parameter allocation scheme is analyzed, and based on the parameter subspace obtained by the characteristic attribution algorithm, the characteristic set corresponding to each transmission line type and target impedance output is determined to form a subspace characteristic mapping structure of the network input end; The parameter set of each characteristic subspace in the parameter allocation scheme is taken as input, and through the deep neural network bottom weight sharing mechanism, a unified feature extraction process is performed on the high-dimensional PCB parameter characteristics to generate a shared bottom feature tensor; Based on the shared bottom feature tensor, for each type of transmission line and target impedance output, an independent multi-objective prediction branch is constructed, and a task-specific decoupling layer is inserted into the branch structure; The output of each task-specific decoupling layer is mapped in a multi-head parallel structure, which further refines the subspace coupling relationship between the transmission line type and impedance output branch and the input parameters to generate a sub-target feature representation; The task-specific decoupling outputs of the multi-objective prediction branch are integrated, and a cooperation relationship regulation interface is integrated in the overall network modeling architecture.

8. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 7, characterized in that, The shared bottom feature tensor is generated by configuring a bottom weight sharing structure in a multi-branch neural network architecture, and inputting the characteristic subspace of each task branch according to a parameter-task mapping table. The bottom feature vector is dimensionally reduced by principal component analysis, and a shared feature tensor is output. 9.The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, wherein, The step S5 specifically includes: Based on the characteristic subspace division result, a set of process parameters with high physical correlation to the target impedance output is selected as the input factor of the task-specific decoupling layer, and a characteristic decomposition input matrix is constructed by parameter grouping mapping; A task-specific decoupling transformation operator is applied to the characteristic decomposition input matrix to realize explicit allocation of the process parameter set in each target impedance prediction branch, map the parameter mixed characteristics to the corresponding target branch according to the prior physical relationship, and form a controlled characteristic path matrix between multiple target tasks; A physical prior constraint function is introduced to impose soft constraints on the characteristic path matrix according to the parameter-impedance physical causal relationship in actual PCB design and manufacturing process; An interpretable regularization term is embedded in the above soft constraint path to sparsely modulate the parameter influence weight distribution using L1 regularization or physical prior guided norm penalty, so as to automatically exclude process parameters irrelevant to the target impedance output and condense the minimum necessary feature set; The strength parameter of the interpretable regularization term is dynamically adjusted, and the regularization influence is optimized in real time based on the loss function feedback during the training process to output the final task-specific decoupling structure for feature correlation constraint between multiple target tasks.

10. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 2, characterized in that, The multi-target performance monitoring report includes the target impedance prediction confidence interval of each transmission line type, the feature parameter contribution degree ranking, the physical consistency statistical distribution, and the process improvement knowledge points.

Images