A neural network-based PCB impedance characteristic prediction and evaluation method

Abstract

The application discloses a PCB impedance characteristic prediction and evaluation method based on a neural network, which comprises the following steps: collecting and standardizing multiple batches of PCB impedance sample data, normalizing and cleaning multi-dimensional features, dividing a feature subspace based on a parameter attribution algorithm, and constructing a multi-branch neural network with physical prior regularization to realize task-specific and collaborative decoupling. Through collaborative optimization of a multi-task loss function, the model accuracy and consistency are dynamically monitored, and multi-task prediction and feature interpretation are output. The scheme improves the prediction accuracy of the model under complex working conditions, and provides intelligent data basis and performance monitoring capability for PCB design and process optimization.

Description

Technical Field

[0001] This invention relates to the field of printed circuit board impedance characteristic modeling and multi-objective neural network prediction technology, and in particular to a method for predicting and evaluating PCB impedance characteristics based on neural networks. Background Technology

[0002] With the continuous advancement of printed circuit board (PCB) design and manufacturing technologies, and the ever-increasing demand for high-speed signals and highly reliable interconnects in electronic products, the accurate prediction, control, and collaborative optimization of PCB impedance characteristics have become a key technical challenge in the industry. For impedance modeling and performance prediction of various transmission lines (such as microstrip lines, striplines, and differential lines) under multiple process and environmental conditions, data-driven methods such as neural networks are gradually becoming mainstream technologies. Currently, in high-dimensional parameter spaces, the demand for multi-objective, multi-task impedance characteristic prediction is becoming increasingly significant, involving complex relationships between structural parameters, process parameters, environmental conditions, and other types of data. The industry commonly employs 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 improve the generalization and automatic modeling capabilities of the models.

[0003] In existing technologies, multi-objective PCB impedance modeling is typically based on multilayer fully connected neural network (MLP) architectures, shared low-level feature extraction with multi-branch output architectures, or general methods integrating multi-task learning (MTL). These approaches often rely on normalized feature processing and parameter attribution ranking to enhance automatic feature extraction and key parameter aggregation. Some literature or patents have also introduced regularization loss terms to improve the overall prediction performance of multi-objective tasks. In recent years, for practical applications with high-dimensional parameters and multiple process labels, technologies such as task collaboration under multi-branch architectures, hierarchical feature sharing, and parameter sensitivity analysis have continued to develop, gradually evolving towards model interpretability, task collaborative optimization, flexible decoupling, and industry feedback.

[0004] However, the aforementioned traditional multi-objective neural network modeling schemes still have the following prominent problems and limitations in the field of high-dimensional PCB impedance prediction:

[0005] (1) When multi-objective / multi-task joint modeling, the features of tasks overlap significantly and the parameters are coupled significantly, which leads to mutual interference of gradients between branch outputs. Ultimately, this causes a decrease in the prediction accuracy of one or more target impedances, an imbalance in the prediction distribution, and some branches may also have physically unreasonable outputs.

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

[0007] (3) Lacking systematic physical constraints and interpretable mechanisms, parameter paths are often mixed across target branches, and the decision-making logic inside the model is seriously "black boxed", which is not conducive to high-reliability application and process closed-loop feedback in the PCB design field. Some models are unable to provide parameter / process improvement suggestions.

[0008] (4) Insufficient consideration of inter-task coordination and consistency. Existing loss functions and training processes are difficult to effectively measure and adjust the physical consistency relationship between multi-target prediction results, which may lead to uncontrolled physical redundancy of output results or task conflicts.

[0009] (5) Most publicly available technical solutions fail to support adaptive dynamic optimization of structural parameters such as feature subspace and regular weights during training, making it difficult to ensure long-term model performance stability and synchronous accuracy improvement under actual parameter perturbation, distribution drift and the introduction of new processes. Summary of the Invention

[0010] In order to solve the above-mentioned technical problems, the present invention provides a method for predicting and evaluating PCB impedance characteristics based on neural networks.

[0011] The technical solution of this invention is implemented as follows: a PCB impedance characteristic prediction and evaluation method based on neural networks, comprising:

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

[0013] S2: Perform normalization, outlier removal, and operating condition labeling on the collected high-dimensional PCB impedance historical data samples to generate data preprocessing results that meet the requirements of multi-task collaborative modeling.

[0014] S3: Based on normalized high-dimensional PCB impedance historical data samples, the feature subspace is divided using a parameter attribution algorithm. The multi-dimensional features that dominate the impedance output of different transmission line types, frequency bands and environmental conditions are grouped to 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 the shared bottom-level feature extraction layer, and construct a multi-objective prediction branch with a task-specific decoupling layer for each type of transmission line and target impedance output.

[0016] S5: Apply physically prior-guided interpretable regularization terms to the task-specific decoupling layer. By adjusting the regularization strength, limit some process parameters to only affect the output of the specified target impedance, thereby achieving eigenvalue decomposition and correlation constraints for each target prediction branch.

[0017] S6: For the regularized multi-branch neural network modeling architecture, a multi-task loss function and a task-to-task collaborative consistency supervision mechanism are used for joint training to measure and adjust the physical consistency or known redundancy relationship between prediction results of different targets and dynamically optimize the training process.

[0018] S7: During joint training, the prediction accuracy of each target and the consistency index between tasks are monitored in real time based on the validation set. The decoupling ratio of the feature subspace and the strength of regularization constraints are dynamically adjusted to achieve accuracy coordination between tasks and adaptive optimization of model stability.

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

[0020] S9: Based on the target impedance prediction results, feature importance decomposition information and physical consistency analysis indicators of the inference output, evaluate the model performance and generate knowledge feedback and multi-objective performance monitoring reports for PCB design or manufacturing process improvement.

[0021] This application provides a PCB impedance characteristic prediction and evaluation method based on neural networks, with the following specific advantages:

[0022] (1) Existing PCB impedance multi-objective modeling often uses a complete set of parameters / shared feature input, which easily leads to gradient conflicts and irrelevant feature interference between different objective branches, resulting in decreased prediction accuracy and even model training instability. This invention uses a parameter attribution algorithm to perform fine-grained grouping of input features and designs a "shared-specific-decoupled" multi-branch neural network architecture based on the physical influence of each objective task. On the basis of shared bottom-level feature extraction, a task-specific decoupling layer is used to accurately allocate highly relevant parameters to each prediction branch. Combined with physical prior regularization constraints, irrelevant disturbances between tasks are effectively suppressed. Experiments show that this scheme can reduce the prediction error between multiple objectives by 15% to 35%, and significantly improve the overall stability of model training and inference;

[0023] (2) To address the "black box" nature of traditional multi-objective networks, this invention introduces prior physical knowledge into the feature decomposition and branch mapping stages, deeply integrating the physical coupling relationship between process parameters / structural parameters with artificial intelligence modeling. Through the adjustable design of the regularization term, some features are forced to act only on the impedance output of physically relevant targets, enabling the model to have parameter traceability and mechanism explanation capabilities. Finally, when outputting the prediction results for each target, the contribution and importance ranking of key parameters can be tracked simultaneously, significantly enhancing the transparency and trustworthiness of the model output for engineering users, and realizing "interpretable AI" in the field of impedance prediction;

[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] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0033] Please attach as follows Figure 1 As shown, this application provides a method for predicting and evaluating PCB impedance characteristics based on neural networks, specifically including:

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

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

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

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

[0038] S5: Apply physically prior-guided interpretable regularization terms to the task-specific decoupling layer. By adjusting the regularization strength, limit some process parameters to only affect the output of the specified target impedance, thereby achieving eigenvalue decomposition and correlation constraints for each target prediction branch.

[0039] S6: For the regularized multi-branch neural network modeling architecture, a multi-task loss function and a task-to-task collaborative consistency supervision mechanism are used for joint training to measure and adjust the physical consistency or known redundancy relationship between prediction results of different targets and dynamically optimize the training process.

[0040] S7: During joint training, the prediction accuracy of each target and the consistency index between tasks are monitored in real time based on the validation set. The decoupling ratio of the feature subspace and the strength of regularization constraints are dynamically adjusted to achieve accuracy coordination between tasks and adaptive optimization of model stability.

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

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

[0043] Step S1: Collect high-dimensional PCB impedance historical data samples containing different structural parameters, process parameters, and environmental parameters. For various transmission line types, operating frequency bands, and temperature conditions, record the corresponding task tag information for each sample. Specifically, this includes:

[0044] S1.1: Obtain historical PCB impedance measurement records as the input of the raw dataset. Based on a professional PCB impedance testing system, high-precision impedance parameter acquisition is performed on multiple batches of PCB samples with different structural parameters to obtain a complete data set including structural parameters such as transmission line type, width, thickness, and dielectric constant, as well as manufacturing process parameters and typical environmental parameters (such as temperature and humidity).

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

[0046] A professional PCB impedance automatic testing system is used to collect high-precision impedance data from a large number of batches of PCB physical samples, and to collect the structural parameters of each sample simultaneously, including transmission line type, width, thickness, number of dielectric layers and dielectric constant.

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

[0048] Industrial environmental monitoring and sampling equipment is used to measure and record environmental parameters corresponding to each collection process in real time, including environmental variables such as temperature, humidity, and air pressure. All parameters are collected into the same raw data structure through data acquisition cards or digital interfaces.

[0049] By using a data synchronization and archiving algorithm (parameters: measurement timestamp, sample number mapping, parameter acquisition frequency), impedance test results are automatically associated with corresponding structural, process, and environmental parameters, generating complete original datasets for various types of tasks.

[0050] By combining high-precision impedance measurement with synchronous collection of multi-dimensional parameters, complex and diverse sample physical information, manufacturing process information, and environmental operating condition parameters are integrated into a high-dimensional raw data set in a structured manner, achieving full data coverage of impedance characteristics under multiple transmission lines, multiple frequency bands, and multiple environmental / process batches.

[0051] For example, data was collected from a batch of multilayer boards produced by a PCB manufacturer in the second quarter of 2023. Each sample was subjected to 4-port impedance measurement using an Agilent E5071C vector network analyzer. Measurement parameters included 50Ω and 90Ω target lines for specific types of microstrip lines, striplines, and differential lines. The measurement width ranged from 0.2 to 1.2 mm, the thickness from 0.015 to 0.035 mm, and the dielectric constant from 4.2 to 4.8. For each sample, the etching batch number, lamination temperature (140–180℃), lamination pressure (10–15 MPa), and the ambient temperature (22–28℃) and humidity (30–60% RH) of the measurement station were recorded. All parameters were automatically numbered and entered into the database via the data acquisition terminal. This resulted in a high-precision impedance sample data set of 2000 sets with different structural, process, and environmental parameters, greatly enriching the foundational database for subsequent feature engineering and multi-task neural network collaborative modeling. Each data item possesses traceability, comprehensiveness, and multi-dimensional scalability.

[0052] S1.2: The acquired datasets of structural parameters, process parameters, and physical environment parameters are formatted using parameter standard coding rules to ensure that parameters of each dimension (such as transmission line width / thickness / material, interlayer dielectric type, etching process batch number, ambient temperature / humidity, etc.) are stored under a professional and unified naming system, laying the foundation for subsequent feature consistency processing and batch data processing to output structured data objects.

[0053] For the acquired datasets of structural parameters, process parameters, and physical environment parameters, standard parameter coding rules are used to standardize the format of the multi-source raw data, unifying the numerical representation of various physical quantities and process parameters across different acquisition batches, devices, and systems. A standard parameter mapping dictionary (parameters: transmission line type code, width specification value, thickness specification value, media type standard number, etching process batch number, environmental temperature and humidity code, etc.) is used to renumber all multi-dimensional data fields, including structural parameters (such as LineType, W, T, Er), process parameters (such as EtchBatch, LaminateTemp, LaminatePress), and environmental parameters (such as EnvTemp, EnvRH), according to international or industry standard naming systems, ensuring they conform to a unified data label template.

[0054] Furthermore, through structured data parsing and reconstruction algorithms (parameters: multi-level parameter field mapping table, abnormal field validation rules), non-standardized naming, inconsistent numerical types, and unit mismatches caused by heterogeneous acquisition systems, manual entry, or changes in acquisition batches are automatically identified in the aforementioned data fields, and automatically converted into a standardized format according to the standard parameter interface protocol. For example, for width (W) or thickness (T) fields, the unit (such as mm, mil, μm) is automatically validated, and unit conversion and decimal precision unification are performed according to the preset target unit, outputting a parameter list with unified dimensions.

[0055] By using a data field uniqueness verification and consistency identifier generation algorithm (parameters: sample number specification, field unique constraint), a unique structured primary key is generated for each set of parameters, enabling cross-platform and cross-batch data traceability capabilities, and configuring a unique coded index for each set of parameter data.

[0056] Furthermore, based on the parameter standardization processing results, a multi-dimensional parameter structured object-oriented output algorithm is adopted to encode all structural parameters, process parameters and environmental parameters into standardized data objects (such as JSON, tables or database entries), generating a structured parameter set for subsequent batch feature processing and efficient data interface docking.

[0057] Through the above parameter standard coding and normalization process, structural parameters, process parameters and physical environment parameters from multiple sources and batches are transformed into structured data objects that are completely consistent, dimensionally unified and named in a standardized manner. This lays a solid data foundation for feature consistency processing and subsequent automated batch data processing, and enables the structured, standardized management and efficient expansion capabilities of large-scale high-dimensional PCB impedance data.

[0058] For example, in one embodiment, for 2000 sets of multilayer PCB sample data collected in the second quarter of 2023, a parameter standard encoding dictionary was used to convert all sample structural parameters (such as transmission line type uniformly encoded using the "LineType" field, with values ​​using industry standard short codes, e.g., MBL representing microstrip line), width, and thickness parameters into decimal form in millimeters (mm) (retaining 3 decimal places), while dielectric constant parameters were stored as floating-point numbers with four decimal places. Process parameter fields, such as etching process batch number, lamination temperature, and lamination pressure, were all archived using the standard database master index numbering method. The process batch number used an 8-digit numeric code, temperature was uniformly expressed in degrees Celsius (°C), and pressure was uniformly expressed in megapascals (MPa), and each was individually bound to a sample number. Environmental parameters (such as temperature and humidity) were standardized to decimal form with three decimal places, named "EnvTemp" and "EnvRH" respectively. All parameters were 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 detected approximately 30 parameter items where unit confusion occurred (e.g., width was mistakenly entered as inches), and automatically corrected and standardized them using unit conversion rules. The final output structured parameter datasets were automatically numbered, thus supporting all batch data interfaces and task calls for subsequent feature engineering, neural network model task allocation, etc., achieving a data standardization rate of 100% and a parameter consistency rate of over 99.9%, providing solid data support and technical assurance for the feature consistency and controllability of multi-objective high-dimensional neural network collaborative modeling.

[0059] S1.3: Based on standardized structured data objects of structural and process parameters, combined with impedance values ​​and accuracy information output by impedance measurement equipment, data label mapping is performed to associate each data sample with its corresponding impedance measurement type (such as specific transmission line type, frequency band distribution, temperature / humidity test conditions), generating a labeled high-dimensional PCB impedance historical sample dataset 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 dataset, collect and summarize multi-label and multi-task attributes such as different transmission line types (e.g., microstrip lines, striplines, differential lines), multiple frequency bands (e.g., low frequency, medium frequency, high frequency) and environmental conditions (e.g., normal temperature, low temperature, high temperature, high humidity). Use the hierarchical classification method in the professional PCB design field to group the samples according to task labels to form a multi-task label distribution reference set for subsequent feature attribution analysis.

[0061] S1.5: For the high-dimensional PCB impedance historical sample dataset after multi-label grouping, a data consistency verification algorithm and missing value completion strategy are adopted to conduct integrity review 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 dataset with no missing items, accurate labels and standardized parameters, to establish a reliable data source for direct input and full-process chain transformation in downstream normalization and feature engineering stages.

[0062] Step S2: Normalization, outlier removal, and operational condition labeling are performed on the collected high-dimensional PCB impedance historical data samples to generate data preprocessing results that meet the requirements of multi-task collaborative modeling. Specifically, this includes:

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

[0064] For the standardized and encoded high-dimensional PCB impedance historical sample dataset, a standard normalization algorithm (parameters: full structural parameters, process parameters, and environmental parameters) is used to uniformly adjust the numerical range of each input feature dimension.

[0065] Furthermore, the normalization process involves using the min-max normalization method to transform each parameter to the [0,1] interval, then using Z-score standardization for the outlier parameters, and finally outputting the normalized feature matrix in a unified format for all input features.

[0066] Furthermore, using the min-max normalization method (Min-Max Scaling, standard interval [0,1]), the following transformations are performed on each structural parameter (such as LineType, W, T, Er, etc.), process parameter (such as EtchBatch, LaminateTemp, LaminatePress), and environmental parameter (such as EnvTemp, EnvRH):

[0067]

[0068] Where, x i Let x be the original value of the i-th feature, min(x) i ) and max(x i ) represent the minimum and maximum values ​​of the feature in the sample space, respectively, x' i These are the dimensionless eigenvalues ​​after normalization.

[0069] Furthermore, for feature parameters that still exhibit extreme outlier issues after normalization, the Z-score normalization algorithm (parameters: mean μ, standard deviation σ) is employed to map the sample features to a standard normal distribution.

[0070]

[0071] Where μ and σ are the characteristic F i The mean and standard deviation, x” i These are the standardized feature scores.

[0072] Furthermore, through the normalization processing module, normalization mapping tables of different dimensions of parameters under multiple batches and multiple task labels are integrated, and the processing results are output as a normalized feature matrix in a unified format for further feature analysis and modeling of the standardized data.

[0073] Through layer-by-layer processing of normalization and standardization algorithms, the structural parameters, process parameters, and environmental parameters are ensured to achieve absolute comparability and dimensionless interference in the numerical domain. This guarantees the numerical stability of downstream parameter attribution algorithms and feature engineering processing from the source, and improves the model's convergence efficiency and generalization ability in high-dimensional feature spaces.

[0074] For example, for 2000 sets of multilayer PCB impedance samples collected in the second quarter of 2023, the structural parameters (width W ranges from 0.2 to 1.2 mm, thickness T ranges from 0.015 to 0.035 mm, dielectric constant Er ranges from 4.2 to 4.8), process parameters (lamination temperature ranges from 140 to 180℃, pressure ranges from 10 to 15 MPa), and environmental parameters (measurement temperature ranges from 22 to 28℃, humidity ranges from 30 to 60% RH) were linearly transformed to the [0,1] interval using Min-Max normalization. Taking the width parameter as an example, the original minimum value was 0.2 mm and the maximum value was 1.2 mm. After normalization, 0.2 mm was mapped to 0, 1.2 mm was mapped to 1, and other values ​​were linearly mapped to the range. Further, considering the skewed distribution characteristics such as lamination temperature, the sample mean μ = 160℃ and the standard deviation σ = 14.14℃ were calculated. Z-score normalization was then used to convert all parameters into standard distribution data. Finally, the high-dimensional structured normalized feature matrix is ​​output as a standardized input, providing sufficient assurance for data consistency and convergence in the subsequent outlier detection and multi-task neural network modeling stages.

[0075] S2.2: Based on the normalization results, statistical outlier detection algorithms (such as IQR or Z-score thresholding) are used to perform outlier removal processing on each feature dimension to identify and exclude abnormal data points caused by equipment errors, abnormal operating conditions or data acquisition defects, and generate a high-confidence, high-dimensional PCB impedance historical data cleaning set to eliminate abnormal interference for subsequent operating condition labeling and feature division.

[0076] S2.3: Input the high-confidence, high-dimensional PCB impedance historical data cleaning set into the label automatic mapping module. Use the multi-label hierarchical mapping algorithm to standardize and normalize the transmission line type label, frequency band label and temperature condition label of each data sample to obtain a highly consistent information label matrix that is adapted to the multi-task structure, supporting efficient task division and downstream label dependency elimination in the multi-task modeling process.

[0077] S2.4: Based on the label matrix and the sample master index, execute the sample multi-dimensional grouping processing algorithm to automatically map the sample data with high confidence and completed label mapping to the multi-task collaborative modeling input structure according to multiple dimensions such as transmission line type, frequency band category, and temperature conditions, and generate a data input subset that is suitable for mining common parameters and individual features of multi-objective neural networks.

[0078] S2.5: Apply data consistency testing and sample balance discrimination methods to the aforementioned multi-task modeling input structure to 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 partitioning and task-specific network structure optimization, thereby achieving a closed-loop guarantee of the quality of data preprocessing results.

[0079] Step S3: Based on the normalized high-dimensional PCB impedance historical data samples, a feature subspace is partitioned using a parameter attribution algorithm. This grouping of multi-dimensional features that dominate impedance output under different transmission line types, frequency bands, and environmental conditions forms a parameter allocation scheme for task-specific modeling. Specifically, this includes:

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

[0081] The input is a normalized high-dimensional PCB impedance historical data sample, including standardized structural parameters, process parameters, and environmental parameter feature matrices.

[0082] A parameter category information screening method (parameters: list of structural parameters, list of process parameters, list of environmental parameters) is adopted to achieve preliminary category classification of input features.

[0083] Furthermore, a structural parameter mapping dictionary is established through a category mapping table, and normalized features are categorized into structural parameter subsets according to fields such as transmission line type, line width, line thickness, and dielectric constant, thereby achieving structural 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 normalized and categorized high-dimensional PCB impedance historical data samples, a parameter attribution algorithm (parameter category set: structural parameters, process parameters, environmental parameters) is used to calculate the characteristic attribution index of each input parameter on the target impedance output.

[0091] The SHAP (SHapley Additive exPlanations) method is used to calculate the contribution weight of each structural parameter, process parameter and environmental parameter in the normalized feature matrix for each impedance output label, thereby realizing the specific influence of a single parameter on the multi-target output.

[0092] By using feature importance ranking algorithms (such as feature importance statistics based on tree models), the ranking score of each parameter in the parameter category set for each impedance component (such as specific impedance targets under different transmission line types and different frequency bands) is calculated to supplement and corroborate the SHAP attribution results.

[0093] Furthermore, multiple correlation analysis algorithms (such as multi-objective correlation coefficient R and multivariate linear regression coefficient) are used to model and estimate the joint influence strength between each characteristic parameter and multiple impedance outputs, so as to achieve a 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-objective correlation analysis results, a structured feature attribution index matrix is ​​generated, where the matrix elements represent the influence intensity and direction of each parameter on each target impedance component.

[0095] By using the feature attribution index matrix, the attribution influence of each parameter in the parameter category set is vectorized, providing a numerical quantitative basis for the subsequent optimal partitioning of the multidimensional feature subspace, and realizing the identification of specific parameters for the target impedance output.

[0096] By using a parameter attribution algorithm chain, the structured features from the previous parameter category classification step are input into the system and transformed into specific attribution indices for each parameter on the target impedance component. This enables multi-objective fine measurement of feature attribution and lays a quantitative foundation for subsequent feature subspace grouping and decoupling modeling.

[0097] For example, for 2000 normalized PCB impedance samples, input structural parameters (e.g., W = 0.45, T = 0.86), process parameters (e.g., LaminateTemp = 0.62, EtchBatch = 0.35), and environmental parameters (e.g., EnvTemp = 0.55, EnvRH = 0.44), use the LightGBM regression model to perform multi-objective fitting on each impedance output label (e.g., microstrip line / high frequency / room temperature, etc.), and calculate the SAP values ​​for the structural, process, and environmental parameters under different objectives. For microstrip line high-frequency impedance prediction, the average SAP value for width W reaches 0.54, thickness T is 0.18, lamination temperature LaminateTemp is 0.12, and ambient temperature EnvTemp is 0.09, indicating that width is the dominant feature. Further, using feature importance ranking based on decision tree, W ranks 1st, T ranks 2nd, and LaminateTemp ranks 3rd. Pearson correlation analysis was applied between the parameters and the target impedance. The correlation coefficient R between the width W and the high-frequency microstrip line impedance was 0.81, and LaminateTemp was 0.28. The final attribution results were summarized into an attribution index matrix, such as [W-0.54, T-0.18, LaminateTemp-0.12, EnvTemp-0.09]. When applied to the prediction of low-frequency impedance of striplines, the contribution of process parameters was significantly improved, as shown in the attribution index matrix [W-0.27, T-0.31, LaminateTemp-0.33, EnvTemp-0.21]. The above attribution analysis achieves parameter specificity measurement, provides a quantitative basis for the subsequent feature subspace partitioning, and improves the input interpretability and attribution accuracy of the multi-objective prediction process.

[0098] S3.3: Using specific attribution indices, perform multi-dimensional feature subspace partitioning on the normalized parameter space. Group the parameter sets with the highest feature influence according to each target task (such as transmission line type, frequency band, environmental conditions) to form a task-specific feature subspace set.

[0099] Based on the normalized high-dimensional PCB impedance historical data samples, the input consists of a standardized feature matrix of structural parameters, process parameters, and environmental parameters, as well as a specific attribution index matrix between each parameter and each target impedance output generated by parameter attribution algorithms (such as SHAP value, feature importance ranking, correlation coefficient analysis, etc.).

[0100] A multi-objective specific attribution index screening method (parameters: attribution contribution threshold λ, attribution ranking TOP-N) is adopted to measure the impact of each feature in the normalized parameter space on each objective task. For each objective task (specifically including the impedance prediction component corresponding to transmission line type, frequency band category, and environmental condition label), the attribution strength of each feature parameter in the attribution index matrix is ​​sorted, and the TOP-N feature set with significant influence is screened. The value of N depends on the complexity of the objective output and the modeling requirements.

[0101] Furthermore, through feature influence aggregation analysis, the SHAP value, feature importance score, and multi-objective correlation parameters are integrated. A weighted normalization method is used to quantify and integrate various attribution indicators, forming a joint influence score S of each feature parameter on the output of each objective. i,j :

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

[0103] Among them, SHAP i,j Let be the average SHAP value of feature i for target j; Imp i,j To score the importance of feature i to target j based on a specific model (such as a tree model); Corr i,j Let αi be the correlation coefficient between feature i and target j; α1, α2, and α3 are empirically set weighting coefficients that satisfy α1 + α2 + α3 = 1. The above formula allows for flexible optimization of the attribution evaluation system based on the actual project situation.

[0104] Furthermore, for each target task, based on the joint score S among all normalized feature parameters... i,j Sort the parameters from highest to lowest, and select the set of parameters that are above a preset threshold λ, or extract the top N items as the set of specific features F for the target task. j This achieves the initial partitioning of the feature subspace. Each F... j The set refers to the set of key influencing parameters of the dominant target j (such as the impedance output of a specific transmission line type / frequency band / operating condition).

[0105] Furthermore, through feature intersection and mutual information analysis, the obtained set of specific features F for each target is analyzed. j Pairwise comparisons are used to identify completely specific subsets of parameters and sets of common parameters that highly overlap across multiple objectives. A task-specific parameter independence metric is employed to evaluate the parameter similarity matrix M between tasks. j,k Perform the calculation:

[0106]

[0107] Among them, M j,k The closer the value is to 0, the more independent the parameter subspaces of task j and task k are; the closer it is to 1, the more overlapping they are. This indicator provides data for subsequent analysis of parameter coordination and decoupling between tasks.

[0108] By employing attribution strength screening, joint score aggregation, intersection analysis, and subspace independence measurement, the normalized parameter space system is transformed into a task-specific feature subspace set {F1, F2, ... F}. t This provides quantitative and structured support for subsequent independent modeling of each objective task and configuration of task-specific network branches.

[0109] By using a multidimensional feature subspace partition driven by specific attribution metrics, the parameter category set and attribution matrix are transformed into a target specific feature set that can be used as input to a multi-task neural network architecture. This achieves optimal segmentation of the parameter space for multi-target tasks, effectively reduces feature interference between tasks, and improves the prediction accuracy and stability of the model in high-dimensional complex scenarios.

[0110] For example, for a microstrip line high-frequency impedance prediction task collected in the second quarter of 2023, in the input parameter attribution matrix, the structural parameter "linewidth W" has a SHAP value of 0.52, a feature importance score of 0.58, and a correlation coefficient of 0.79. After weighting using the joint scoring formula (α1 = 0.4, α2 = 0.4, α3 = 0.2), the S is obtained. W,HF-MTL =0.4*0.52 + 0.4*0.58 + 0.2*0.79 = 0.589. Among all parameters, S is obtained. i,HF-MTL Key parameters >0.2 include W (0.589), T (0.37), and LaminateTemp (0.21), forming the specific feature set F for high-frequency microstrip line tasks. HF-MTL ={W,T,LaminateTemp}. For the stripline low-frequency impedance prediction task, the top-scoring parameters are T (0.41), LaminateTemp (0.36), and EnvTemp (0.23). The parameter set F for both tasks... HF-MTL With F LF-SL Calculate the intersection, and the ratio of the intersection to the union is M. j,k A value of 0.5 indicates that the feature subspaces partially overlap in this scenario, but task-specific parameters exist. Using the above processing, the optimal set of specific feature parameters for each type of transmission line / frequency band / operating condition is finally output, achieving refined feature configuration for downstream multi-branch modeling. The significant improvement in prediction collaboration and accuracy between tasks is verified in actual model training.

[0111] S3.4: For parameter subsets with highly overlapping or coupled parameters among tasks, use collaborative feature analysis (such as multi-objective correlation coefficient and collaborative redundancy analysis) to determine the shared parameter subspace, and implement attribution decoupling labeling for redundant or mutually exclusive features to ensure that the parameter allocation scheme takes into account both specificity and synergy.

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

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

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

[0115] The input is a high-dimensional parameter grouping result output by the parameter attribution algorithm, including a set of specific parameter subspaces assigned to each transmission line type and target impedance output, as well as a related structured feature category mapping table.

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

[0117] Furthermore, through the feature subspace mapping algorithm (parameters: task label set, target impedance output type), the clustering and affiliation of each parameter subspace with specific transmission line types and impedance outputs are realized, dividing the parameter space into multiple data subsets oriented towards different tasks, and clarifying the input feature set corresponding to each target impedance branch.

[0118] Furthermore, a parameter-task allocation index matrix construction method (parameters: parameter allocation scheme, attribution score threshold) is adopted to automatically generate a 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, thereby achieving consistency and traceability of parameter mapping.

[0119] Furthermore, a feature structured labeling mechanism (parameters: parameter ID, attribution level label, role classification code) is applied to assign structured attribute labels to each parameter feature in the mapping table, indicating its specific index and participating branch in the network input structure, providing clear indication for configuring task-specific feature pathways at the network input end.

[0120] By using the parameter subspace mapping matrix and feature classification labels, a standardized transformation is achieved from high-dimensional parameter grouping results to the network input subspace feature mapping structure, providing a standardized feature entry point for subsequent multi-branch neural network modeling architecture and enabling unambiguous configuration of task-specific inputs.

[0121] For example, for 3240 historical impedance samples of multilayer PCBs collected in the first half of 2023, a feature attribution algorithm was used to extract the set of structural parameters {W(0.52), T(0.46), Er(0.17)} and the set of process parameters {LaminateTemp(0.21)} under the microstrip line-high frequency category, and to distinguish the specific parameter sets {T(0.41), LaminateTemp(0.36); EnvTemp(0.23)} under the stripline-low frequency category. Based on the parameter attribution matrix, the parameter categories were structured and encoded as F_{HF-MTL}={W,T,Er,LaminateTemp} and F_{LF-SL}={T,LaminateTemp,EnvTemp}. A parameter mapping table is used to generate: parameter W → microstrip line - high-frequency branch input, T → all branch inputs, LaminateTemp → high-frequency microstrip line and low-frequency stripline branch inputs, and EnvTemp → stripline - low-frequency branch inputs. A structured labeling mechanism is used to mark the task classification and input index of each parameter, such as [W,HF-MTL,idx=1], [T,ALL,idx=2]. This processing forms a multi-branch subspace feature mapping structure for the neural network input, enabling accurate clustering and unambiguous input of key parameters required by each prediction branch of the network. This improves the specificity and synergy of multi-branch modeling and supports the efficient implementation of downstream shared feature extraction and decoupled modeling stages.

[0122] S4.2: Using the parameter set of each feature subspace in the parameter allocation scheme as input, a unified feature extraction process is performed on the high-dimensional PCB parameter features through the deep neural network's underlying weight sharing mechanism to generate a shared underlying feature tensor, which serves as the common input basis for subsequent multi-objective decoupling modeling.

[0123] The input is the high-dimensional parameter grouping results output by the parameter allocation scheme. The results include the set of parameters in each feature subspace after being screened and structured based on task-specific attribution indicators. It covers information for different transmission line types and target impedance outputs. All parameters have been uniformly numbered and normalized.

[0124] A deep neural network weight sharing mechanism (parameters: feature subspace set, shared convolutional or fully connected module depth L, activation function type, normalization strategy) is adopted to achieve unified feature extraction of high-dimensional PCB parameter features.

[0125] Furthermore, through the feature vector concatenation algorithm (parameter: feature subspace input matrix), the structural parameters, process parameters, and environmental parameters subspaces assigned to each task are assembled in the order of the main index to form a multidimensional input tensor, which serves as the standard input data format for the underlying neural network.

[0126] Furthermore, a deep feature abstraction algorithm (such as a multi-layer nonlinear mapping network, parameters: number of layers L, number of nodes N, activation function ReLU or GELU) is employed to perform progressive feature mapping and nonlinear activation on the above input tensor, capturing the complex interaction relationships between high-dimensional parameters and outputting a globally shared low-level feature tensor T. shared .

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

[0128] Furthermore, through feature redundancy compression and principal component selection algorithms (parameters: principal component contribution rate threshold, compression ratio), T is... shared Dimensionality reduction optimization is performed, retaining the main feature vectors whose cumulative contribution rate is higher than a specified threshold, further improving the efficiency and generalization ability of subsequent multi-objective decoupling modeling.

[0129] Through the above algorithm chain, the parameter sets of each subspace in the parameter allocation scheme are taken as input and efficiently transformed into a unified and highly representative shared underlying feature tensor. This provides high-quality common input for the directional feature mapping of subsequent multi-objective decoupling branches, realizes feature collaboration and resource sharing between different task modeling, and reduces the risk of gradient conflict caused by feature redundancy.

[0130] For example, the input is a 2023 version PCB impedance feature dataset, consisting of an input matrix composed of structural parameters (e.g., W=0.55, T=0.76, Er=0.34), process parameters (e.g., EtchBatch=0.41, LaminateTemp=0.63), and environmental parameters (e.g., EnvTemp=0.48, EnvRH=0.33) selected by a parameter allocation scheme. Each data input is 20-dimensional. A deep neural network with a shared weight structure is used, configured with three hidden layers, each with 128, 64, and 32 nodes, respectively, and ReLU activation function. All input data batches first pass through a feature normalization layer (BatchNorm), with a batch training size of 64. After the data enters the three fully connected layers, it undergoes nonlinear transformation and multi-layer activation, outputting a final 32-dimensional shared feature tensor. This feature tensor is further analyzed by PCA, with a principal component contribution rate threshold set to 95%, and finally compressed to 16-dimensional principal component features. Experimental results show that this processing flow can effectively reduce feature redundancy while ensuring that the main physical and process information is not lost. The number of model convergence rounds is reduced by 30%, and the mean square error index of subsequent multi-objective neural network branch prediction is better than that of the baseline without shared feature extraction structure, and the prediction stability is significantly improved.

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

[0132] S4.4: The output of each task-specific decoupling layer adopts a multi-head parallel mapping structure to further refine the subspace coupling relationship between the transmission line type and impedance output branch and the input parameters, generating sub-target feature representations, and providing a structured basis for subsequent interpretable regularization and task-differentiated constraints.

[0133] S4.5: Integrating the task-specific decoupling output of the multi-objective prediction branch, the collaborative relationship control interface is integrated into the overall network modeling architecture, providing a refined structural entry point for subsequent application of physical prior constraints and supervision of inter-task collaboration consistency, and realizing the organized output of multi-objective feature collaborative reasoning.

[0134] Step S5: Apply a physically prior-guided interpretable regularization term to the task-specific decoupling layer. By adjusting the regularization strength, limit certain process parameters to only affect the specified target impedance output, thereby achieving eigenvalue decomposition and correlation constraints for each target prediction branch. Specifically, this includes:

[0135] S5.1: Based on the feature subspace partitioning results, a set of process parameters that are highly physically related to the target impedance output is selected as the input factors of the task-specific decoupling layer. The feature decomposition input matrix is ​​constructed 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 set of high-dimensional structural parameters, process parameters and environmental parameters corresponding to each target impedance output, as well as their corresponding attribution strength and physical correlation labels.

[0137] A physical correlation screening method (parameters: feature attribution strength threshold, physical prior knowledge base) is adopted to screen process parameters for each target impedance output. Process parameters with attribution strength higher than the threshold are given priority and a set of process parameters with significant physical effects on the target impedance is initially selected.

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

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

[0140] Furthermore, through a feature label enhancement mechanism (parameters: physical attribution label, influence weight score, parameter unique index), attribute labels such as physical correlation flag, attribution value, and network input index are added to each process parameter feature in the feature decomposition input matrix, thereby enhancing the interpretability and traceability of each field of the input matrix.

[0141] By filtering physical correlations in the feature subspace, grouping and mapping parameters, and constructing the feature decomposition input matrix, the set of process parameters to be input into the task-specific decoupling layer is transformed into a well-structured feature decomposition input matrix that meets the physical interpretation requirements. This provides standardized input for downstream decoupling mapping and interpretable regularization processing, achieving the physical interpretability design goal of the task-specific decoupling layer.

[0142] For example, for precision impedance test samples of microstrip lines and striplines collected from high-multilayer PCBs, feature attribution analysis revealed that the set of process parameters {EtchRate (attribution strength 0.41), LaminateTemp (0.38), EtchBatch (0.19)} is highly correlated with the high-frequency impedance of the microstrip line, while the main correlated parameters for the low-frequency impedance of the stripline are {LaminateTemp (0.36), EnvTemp (0.29)}. The physical correlation screening threshold was set to 0.15, and the above parameters were assigned to the decoupling layers of the high-frequency branch of the microstrip line and the low-frequency branch of the stripline, respectively. Using a parameter grouping mapping algorithm, EtchRate and LaminateTemp were input into the microstrip line-high-frequency branch, and LaminateTemp and EnvTemp were input into the stripline-low-frequency branch. A feature decomposition input matrix is ​​constructed, where the parameter fields of each sample include [EtchRate_idx1,LaminateTemp_idx2,EtchBatch_idx3] (microstrip branch) and [LaminateTemp_idx1,EnvTemp_idx2] (strip branch). Each field is appended with a physical attribution score and a physical label. After this processing, the feature decomposition input matrix to the decoupling layer possesses high physical interpretability and clear branch attribution, and the mapping relationship between upstream physical attribution and parameter encoding processes can be unambiguously traced. This matrix is ​​used for subsequent decoupling mapping and regularization constraints, which can significantly improve the physical consistency interpretation of parameter effects and network traceability. In the verification scenario, the correlation coefficient between decoupling parameter allocation and physical measurement results can be improved to over 0.93, meeting the high requirements of parameter attribution transparency in engineering applications.

[0143] S5.2: Apply task-specific decoupling transformation operator to the eigenvalue decomposition input matrix to achieve explicit allocation of the process parameter set in each target impedance prediction branch, and map the parameter mixture features to the corresponding target branches according to the prior physical relationship to form a controlled feature path matrix among multiple target tasks.

[0144] The input is the eigenvalue input matrix obtained by filtering based on the parameter attribution algorithm and physical correlation, including the subset of process parameters assigned to each target impedance prediction branch and their vectorized encoding.

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

[0146] Furthermore, by decoupling the multiplication operation of the mapping weight matrix, the process parameter feature inputs for each task branch are calculated:

[0147]

[0148] Among them, F (k) Let be the specific decoupling feature vector of the k-th target branch. X is the task-specific decoupling transformation weight matrix for the k-th branch. in The input matrix is ​​the standardized eigenvalue decomposition matrix.

[0149] Furthermore, by automatically adjusting based on physical priors and attribution weights The elements in the F-value are such that only parameters with physical correlation above a threshold are affected by F. (k) The output has a significant impact, and the path weights of other parameters tend to zero, thus achieving controlled feature decomposition.

[0150] Furthermore, a parameter distribution sparsity adjustment strategy is adopted, such as for Applying element sparsity regularization ensures that the cross-weights of parameters across target task branches are minimized, avoiding information interference caused by the mixing of parameters between different prediction branches of the model.

[0151] Furthermore, by integrating the mapping results of all target branches, a controlled feature path output matrix is ​​constructed for multi-target tasks:

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

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

[0154] By applying the task-specific decoupling transformation operator, the structured process parameter set generated in the previous step is mapped, allocated, and transformed into specific feature paths in a multi-objective network architecture. The feature path matrix with branch mapping weights is output, realizing the directional contribution of process parameters to the target impedance output and improving the interpretability and controllability of model feature decomposition.

[0155] For example, for the microstrip line-high frequency impedance prediction branch, process parameters EtchRate (physical attribution strength 0.41) and LaminateTemp (0.38) are assigned; for the stripline-low frequency branch, LaminateTemp (0.36) and EnvTemp (0.29) are selected. The eigenvalue decomposition input matrix X is set. inThe dimension is [batch, 5] (containing EtchRate, LaminateTemp, EtchBatch, EnvTemp, PressIndex), with branch weight matrix Acting on X respectively in The characteristic vector F of the high-frequency microstrip line branch for each batch is calculated. (HF-MTL) The low-frequency bandline branch feature vector F consists of EtchRate and LaminateTemp. (LF-SL) The model consists of LaminateTemp and EnvTemp, with other parameter paths having a weight of 0. L1 sparsity regularization is applied to further constrain the weight distribution, preventing the activation of parameter paths unrelated to the target. Model validation shows that after this mapping, the high-frequency impedance branch of the microstrip line significantly improves its response to changes in the EtchRate parameter (attribution correlation coefficient increases from 0.74 to 0.92), while the low-frequency branch of the stripline exhibits significant independent sensitivity to EnvTemp. The model branch responses show no obvious coupling, and the feature decoupling accuracy reaches 96%, meeting the requirements of engineering interpretability and physical controllability in high-dimensional multi-target impedance prediction scenarios.

[0156] S5.3: Introduces a physical prior constraint function, which applies soft constraints to the characteristic path matrix based on the physical causal relationship between parameters and impedance in actual PCB design and manufacturing process, so that some process parameters only activate the output branch of the specified target impedance, reducing interference from irrelevant features.

[0157] S5.4: Interpretable regularization terms are embedded into the soft-constraint pathways mentioned above. L1 regularization or physical prior guided norm penalty is used to implement sparsity modulation on the parameter influence weight distribution, so that the model can automatically filter out process parameters that are irrelevant 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, optimize the influence of regularization in real time based on the loss function feedback during training, achieve the optimal balance between feature interpretability and model prediction accuracy, and output the final task-specific decoupling structure for constraining feature correlation between multi-objective tasks.

[0159] Step S6: For the regularized multi-branch neural network modeling architecture, joint training is performed using a multi-task loss function and an inter-task collaborative consistency supervision mechanism to measure and adjust the physical consistency or known redundancy relationships between prediction results of different targets, and to dynamically optimize the training process. Specifically, this includes:

[0160] S6.1: Input multi-task feature samples into the regularized multi-branch neural network modeling architecture, and map the feature subspace input of each task to the corresponding prediction branch through the task label routing mechanism to achieve target distribution of multi-task features, laying 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 the impedance prediction and feature allocation schemes of each task, so as to obtain multi-target basic prediction capability and parameter constraint capability.

[0162] The input consists of high-dimensional feature subspace samples distributed to each target prediction branch through a task label routing mechanism, a regularized multi-branch neural network modeling architecture, and a parameter allocation scheme.

[0163] A multi-task loss function design method (parameters: number of prediction branches for each target, target type for each branch, and weighting coefficients of the loss function) is adopted to achieve a joint loss measure for the impedance output of all targets. The loss function structure combines mean squared error loss (MSE), L1 regularization loss, L2 regularization loss, and interpretable regularization term to form a total multi-target loss function, specifically expressed as follows:

[0164]

[0165] Where K is the total number of task branches, λ k The loss weight for the k-th objective is... Let L be the mean squared error loss of the k-th target branch. L1 For global L1 regularization terms, L L2 For global L2 regularization terms, L expl For interpretable regularization terms, μ, γ, and δ are the adjustment coefficients of the corresponding terms.

[0166] Furthermore, the basic prediction accuracy of each task branch is assessed using the mean squared error loss (MSE) calculation method (parameters: target branch prediction output, actual measured impedance label). The specific calculation formula is as follows:

[0167]

[0168] Where, N k Let k be the sample size of the k-th task branch. This represents the true impedance output of the i-th sample in the k-th task branch. This corresponds to the predicted value.

[0169] Furthermore, by using a global L1 / L2 regularization loss calculation method (parameters: model parameter vector, regularization coefficient), the global norm distribution of neural network parameters is constrained to mitigate overfitting.

[0170]

[0171] w j This represents all trainable weight parameters.

[0172] Furthermore, by employing an interpretable regularization term design method (parameters: feature influence weight distribution, physical prior sparsity target), the influence weights of the input parameters in each target prediction branch are sparsely modulated, causing the weights of process parameters independent of the target impedance to converge to zero. This term can be defined as:

[0173]

[0174] Among them, S (k) This is the set of parameter indices that are physically directly related to the output of the k-th impedance. The parameter weights are for the k-th branch.

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

[0176] By jointly optimizing and regularizing multiple loss functions as described above, the prediction error of the model for each objective task, the generalization ability of the global structure, and the interpretable constraints of physical correlation are incorporated into the same evaluation system, thereby achieving a synergistic improvement in the basic prediction ability and parameter constraint ability for multiple objectives.

[0177] For example, in a three-task parallel PCB impedance multi-objective prediction scenario, the input sample batch is 64, the MSE weighting coefficients for each branch are set to λ1 = 0.4, λ2 = 0.3, λ3 = 0.3, and the regularization coefficient is set to μ = 10. -4 (L1), γ = 10 -5 (L2), δ=10 -3 (Interpretable regularization). For the k-th branch sample in the batch, the actual impedance label is 45.2, and the predicted output is 44.9, then the corresponding MSE is (45.2-44.9). 2 / 64≈1.4×10 -2 After weight assignment, the number of L1 norm activation parameters was reduced to 38% of the initial number of parameters, and the influence of irrelevant parameters on the weights converged to 10. -4Order of magnitude. After 20 rounds of joint training, the MSE of the model's three-objective integrated prediction decreased from the initial 0.187 to 0.022, and the interpretability improvement metric (corresponding to the consistency rate between the process-dominant parameters and physical labels) increased from 0.71 to 0.93, meeting the requirements of high accuracy and feature interpretability for high-dimensional multi-objectives.

[0178] S6.3: Apply a weight adjustment mechanism to each objective loss component of the multi-task loss function. Adjust the task weights in a timely manner through a dynamic weight allocation algorithm to achieve a balance in the prediction accuracy of each objective impedance branch and alleviate gradient conflicts in the multi-objective optimization process. This enables the network training to maintain consistency and synergy in multi-objective prediction under high-dimensional parameters.

[0179] S6.4: After each training batch, collect the predicted output of all target impedance branches, and use the inter-task collaborative consistency supervision mechanism to calculate the inter-task physical consistency evaluation index and conflict loss term based on the known physical coupling relationship or redundancy constraints. Flexibly adjust the loss function to guide the model parameters to converge toward the optimal direction of physical consistency in real time.

[0180] S6.5: Based on the collaborative consistency evaluation results during the training process, execute the multi-task decoupling rate and interpretable regularization coefficient automatic adjustment algorithm, dynamically optimize the feature subspace segmentation ratio and regularization constraint strength, and realize adaptive tuning of model architecture and continuous improvement of prediction stability between collaborative tasks.

[0181] S6.6: After completing a full joint training, based on the model parameters and the output of each task branch, perform periodic backtracking optimization. By replaying the historical gradient trajectory and the loss evolution process between tasks, auxiliary stability tuning is implemented for sub-tasks that are difficult to converge. Finally, the weight parameters of the overall optimized multi-branch neural network modeling architecture are output.

[0182] Step S7: During joint training, based on real-time monitoring of the prediction accuracy of each target and the consistency index between tasks using the validation set, the feature subspace decoupling ratio and regularization constraint strength are dynamically adjusted to achieve inter-task accuracy coordination and adaptive tuning of model stability. Specifically, this includes:

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

[0184] S7.2: Apply the inter-task consistency judgment algorithm to the synchronously obtained target impedance prediction results, quantify the cooperative consistency index between the outputs of each target, 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 cooperative consistency criterion.

[0185] S7.3: Based on the target impedance prediction accuracy index obtained from real-time monitoring and the consistency criterion between tasks, the decoupling ratio of the feature subspace decoupling layer is dynamically optimized and adjusted. The decoupling parameter search algorithm is used to adjust the weight of highly correlated features and optimize the feature decomposition structure to improve the relative independence and accuracy coordination of each target prediction.

[0186] For the target impedance prediction accuracy index and inter-task consistency criterion obtained from the validation set, a dynamic optimization and adjustment algorithm for the feature subspace decoupling layer is used to achieve real-time adjustment of the weights assigned to highly correlated features. A performance-driven decoupling parameter search method (parameters: task accuracy index, consistency score, feature influence matrix) is adopted to perform flexible weight allocation for each feature subspace, automatically searching for the optimal allocation ratio of highly correlated features among various target tasks, and achieving accurate demarcation between task-specific features and shared features.

[0187] Furthermore, an adaptive gradient adjustment mechanism (parameters: loss function gradient, feature contribution) is used to progressively adjust the weights of each parameter within the feature subspace, responding in real-time to accuracy fluctuations and changes in collaborative consistency, ensuring optimal independence of features across tasks. The following feature allocation weight update formula is adopted:

[0188]

[0189] in, Assign new weights to the k-th feature subspace. Here, E represents the weight of the previous period, η is the step size coefficient, and E is the weight of the previous period. k Let C be the current accuracy of the k-th task, C be the consensus criterion, and α and β be the accuracy and consistency adjustment coefficients, respectively. For the loss function with respect to w k The gradient.

[0190] Furthermore, a momentum adaptive screening algorithm (parameters: momentum term, historical weight volatility) is used to smooth the weight adjustment trend, suppressing sharp parameter fluctuations caused by instantaneous noise and stabilizing the eigenvalue decomposition process. The following momentum modulation formula is adopted:

[0191]

[0192] in, The increment is the current feature weight, and μ is the momentum coefficient. Assign weights to the smoothed features.

[0193] Furthermore, based on the feature relevance threshold screening method, highly correlated features with large overlap between tasks are grouped, merged, and finely redistributed. Parameters above the relevance 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, minimizing the correlation and optimally suppressing information redundancy among the input feature subspaces of each target task, while retaining necessary shared features to support inter-task collaboration.

[0194] Through the above-mentioned chain-like parameter adjustment and structural optimization, the accuracy and consistency feedback obtained from real-time monitoring of the validation set is directly applied to the feature decoupling framework, realizing mutual enhancement and conflict suppression of the model's multi-objective outputs. Finally, the dynamically adjusted feature subspace allocation parameter set is output, providing an efficient and controllable feature structure configuration for downstream regularization constraint optimization and network-wide collaborative training.

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

[0196] S7.4: Based on the validation set performance, a regularization adaptive adjustment strategy is used to increase or decrease the constraint strength of the interpretable regularization term. The regularization weight adjustment algorithm is used to strengthen the discrimination effect of process parameters on the specified impedance target, reduce the impact of irrelevant perturbations between targets on the model training stability, and thus optimize the regularization weight distribution.

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

[0198] Step S8: Using the optimized multi-branch neural network modeling architecture, for the different structural parameters, process parameters and multi-condition data of the new input, the target impedance prediction results and their uncertainty ranges under each task are inferred and output respectively.

[0199] Specifically, it includes:

[0200] S8.1: Perform feature normalization and format adaptation processing on the newly input structural parameters, process parameters and multi-condition data. Use the same normalized parameters and structured data templates as in the model training stage to ensure that the sample to be predicted is strictly matched with the input feature space of the multi-branch neural network modeling architecture, and obtain the adaptive standard input feature matrix to provide structurally consistent data support for subsequent model inference.

[0201] S8.2: Based on the adaptive standard input feature matrix, the optimized multi-branch neural network modeling architecture is input. Through a unified low-level feature extraction layer, multi-level feature mapping and abstraction are performed on the new input data to obtain a globally shared deep feature representation, thereby providing a unified basic feature for the decoupling processing of downstream task branches.

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

[0203] Input the standardized and format-adapted structural parameters, process parameters, and multi-condition data. Obtain a globally shared deep feature tensor through the underlying unified feature extraction layer, which serves as the starting input for this step.

[0204] A task-specific decoupling mapping algorithm is adopted to input the globally shared deep feature representation into the decoupling layer corresponding to each task branch, thereby achieving efficient splitting of feature vectors in the parameter subspace.

[0205] Furthermore, by using linear projection or sparse coding methods based on parameter allocation mapping rules, feature subsets that are physically related to each target task (such as specified transmission line type, frequency band, and environmental conditions) are separated from the global feature tensor to form a three-dimensional target task-specific feature subspace.

[0206] Furthermore, by using feature redundancy suppression algorithms (such as maximum correlation minimum redundancy mRMR), redundancy discrimination and optimal subset selection are performed on the subspace features extracted from each task branch, thereby improving feature independence between tasks and reducing gradient interference caused by parameter coupling.

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

[0208] z k =P k ·f g

[0209] Among them, z k Let P be the specific feature subspace vector of the k-th target task. k Let f be the parameter projection matrix for the k-th task. g This represents a globally shared feature.

[0210] Through the above decomposition mapping, a highly sparse and optimally correlated multi-objective subspace feature vector is generated for each target task, which serves as the input for downstream impedance prediction inference, thereby achieving flexible control of feature modulation between tasks and minimization of cross-task interference.

[0211] By using multi-objective subspace feature decomposition mapping, global features are effectively transformed into data vectors that serve each objective task, thereby achieving parameter interference suppression and feature cross-denoising in high-dimensional multi-objective scenarios, significantly improving the accuracy and stability of multi-task impedance prediction.

[0212] For example, in a PCB multi-objective impedance prediction scenario, the input consists of standardized structural parameters (e.g., transmission line width 0.18mm, thickness 0.035mm, dielectric constant 4.2), process parameters (batch number A12, etching process PTH-3), and environmental conditions (temperature 85℃, humidity 60%). The system generates a global feature tensor of length 256 through a bottom-level feature extraction layer. The task branches include three main categories: microstrip lines, striplines, and differential lines. For the microstrip line branch, the projection matrix P1 is set (selecting 55 feature channels carrying coupled physical quantities, sparsity 0.8), the stripline branch projection matrix P2 is a 63-dimensional sparse feature selection, and the differential line P3 is a 92-dimensional feature mapping. mRMR is used for optimal feature subset selection, removing 18% of cross-task redundant features. 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 decreased to 3.6%, and the correlation of input features of each branch increased to 92%, laying the foundation for accurate inference of subsequent impedance components and ensuring physical decoupling and optimal parameter allocation among multi-target predictions.

[0213] S8.4: Apply the output layer mapping function to the feature vector of the multi-target subspace, and combine it with the interpretable regularization mechanism to perform impedance prediction inference for each task branch, and obtain a prediction result set containing the impedance component values ​​of each target and the corresponding physical consistency parameters, so as to realize the automatic inference of the 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, an uncertainty quantification algorithm (such as Bayesian inference or confidence interval estimation) is used to perform statistical confidence analysis on each target output, outputting the impedance prediction value and its uncertainty interval for each task, providing risk-aware data support for downstream result evaluation and process decision-making.

[0215] Step S9: Based on the target impedance prediction results, feature importance decomposition information, and physical consistency analysis indicators output by the inference, the model performance is evaluated, and knowledge feedback and multi-objective performance monitoring reports are generated for PCB design or manufacturing process improvement. Specifically, this includes:

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

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

[0218] S9.3: Perform physical consistency analysis on the impedance prediction results and feature importance decomposition information of each task objective, adopt inter-task correlation test and physical constraint verification mechanism, measure the physical redundancy consistency and reasonable range of key parameters among the prediction results, output physical consistency criteria for multi-objective collaborative modeling, and generate input consistency parameter set for performance monitoring reports.

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

[0220] S9.5: Integrates knowledge feedback text, multi-objective performance monitoring parameters, and structured evaluation data, and outputs multi-objective performance monitoring reports through an automated report generation component. The reports include the confidence interval of target impedance prediction for each transmission line type, the ranking of the contribution of characteristic parameters, the statistical distribution of physical consistency, and key points of process improvement knowledge, realizing a closed loop of full-process performance traceability and knowledge push for the high-dimensional multi-objective impedance prediction model.

[0221] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0222] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A 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, and recording corresponding task label information of the samples for multiple types of transmission lines, 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, the multi-branch neural network modeling architecture comprising a shared bottom feature extraction layer and independent multi-target prediction branches, each branch comprising a task-specific decoupling layer, and each branch corresponding to different transmission line types and target impedance outputs, and the multi-branch neural network modeling architecture construction process comprising: analyzing the high-dimensional parameter grouping result output by the parameter allocation scheme, determining the feature set corresponding to each transmission line type and target impedance output based on the parameter subspace obtained by the feature attribution algorithm, and forming a subspace feature mapping structure at the network input end; taking the parameter set of each feature subspace in the parameter allocation scheme as input, performing a unified feature extraction process on the high-dimensional PCB parameter features through a deep neural network bottom weight sharing mechanism, and generating a shared bottom feature tensor; based on the shared bottom feature tensor, constructing an independent multi-target prediction branch for each type of transmission line and target impedance output, and inserting a task-specific decoupling layer in the branch structure; adopting a multi-head parallel mapping structure for the output of each task-specific decoupling layer, further refining the subspace coupling relationship between the transmission line type and impedance output branch and the input parameter, and generating a sub-target feature representation; integrating the task-specific decoupling outputs of the multi-target prediction branches and integrating a cooperation relationship regulation interface in the overall network modeling architecture; S5, applying an interpretable regularization term of physical prior guidance in the task-specific decoupling layer, limiting part of the process parameters to only affect the specified target impedance output by regulating the regularization strength, and specifically comprising: based on the feature subspace division result, selecting a set of process parameters having a high physical correlation with the target impedance output as the input factor of the task-specific decoupling layer, and constructing a feature decomposition input matrix through parameter grouping mapping; applying a task-specific decoupling transformation operator to the feature decomposition input matrix to realize explicit allocation of the process parameter set in each target impedance prediction branch, map the parameter mixed features to the corresponding target branch according to the prior physical relationship, and form a controlled feature path matrix between multiple target tasks; introducing a physical prior constraint function, and applying a soft constraint to the feature path matrix according to the parameter-impedance physical causal relationship in the actual PCB design and manufacturing process; embedding an interpretable regularization term in the above soft constraint path, using an L1 regularization or a physical prior oriented norm penalty to implement sparsity modulation on the parameter influence weight distribution, and prompting the model to automatically exclude process parameters irrelevant to the target impedance output to condense a minimum necessary feature set; The strength parameter of the explainable regularization term is dynamically adjusted, the regularization influence is fine-tuned in real time based on loss function feedback during the training process, and the final task-specific decoupling structure for feature correlation constraint between multi-target tasks is output; S6, the regularized multi-branch neural network modeling architecture is trained jointly by 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; 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 proportion and the regularization constraint strength are dynamically adjusted; S8, using the optimized multi-branch neural network modeling architecture, input different structure parameters, process parameters and multi-condition data, and output 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, the model performance is evaluated, and knowledge feedback and multi-target performance monitoring report for PCB design or manufacturing process improvement are generated. 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 original data set input, collect high-precision impedance parameters of multiple batches of PCB samples with different structure parameters based on professional PCB impedance test system, and obtain complete data set of structure parameters, manufacturing process parameters and environmental parameters; The obtained structure parameters, process parameters and physical environment parameter data set is standardized in format by using parameter standard coding rule; 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, data label mapping is performed to associate each data sample with its corresponding impedance measurement type one by one, and a labeled high-dimensional PCB impedance historical sample data set with task labels is generated; For the labeled high-dimensional PCB impedance historical sample data set, collect and summarize different transmission line types, multi-frequency bands and environment working condition multi-label multi-task attributes, and use the hierarchical classification method in the field of PCB design to group the samples according to task labels to form a multi-task label distribution reference set; For the multi-label grouped high-dimensional PCB impedance historical sample data set, a data consistency checking algorithm and a missing value completion strategy are used to check the integrity of key structure 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 labels and standardized parameters.

4. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, characterized in that, The step S2 specifically comprises: Standardized normalization algorithm is applied to the collected high-dimensional PCB impedance historical data samples, and multi-dimensional input features of structure parameters, process parameters and environmental parameters are normalized to obtain data standardization results; Based on the data standardization result, an outlier rejection processing is performed on each feature dimension using a statistical outlier detection algorithm, abnormal data points caused by equipment errors, abnormal working conditions or data collection defects are identified and excluded, and a high-confidence high-dimensional PCB impedance historical data cleaning set is generated; The high-confidence high-dimensional PCB impedance historical data cleaning set is input into a label automatic mapping module, and a multi-label hierarchical mapping algorithm is used to standardize and normalize the transmission line type label, frequency band label and temperature working condition label of each data sample, so that a consistent and multi-task task structure adaptive information label matrix is obtained; Based on the label matrix and sample main index, a sample multi-dimensional grouping processing algorithm is executed, and the sample data with high confidence and label mapping completion is automatically mapped to a multi-task collaborative modeling input structure according to the transmission line type, frequency band category and temperature working condition multiple dimensions, and a data input subset is generated; The data consistency test and sample balance discrimination method is applied to the multi-task collaborative modeling input structure, 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: using the minimum-maximum normalization method to convert each parameter to the [0, 1] interval, and then using Z-score standardization for abnormal distribution parameters, and outputting a normalized feature matrix in a unified format.

6. The neural network-based PCB impedance characteristic prediction and evaluation method according to claim 1, characterized in that, The step S3 specifically includes: Based on the normalized high-dimensional PCB impedance historical data sample, a parameter preliminary screening classification operation is performed based on the process and structure parameter category information, and a preliminary parameter category set is obtained; According to the preliminary parameter category set, a 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; Using the specific attribution index, a multi-dimensional feature subspace division operation is performed 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, a collaborative feature analysis is used to determine a common parameter subspace, and redundant or mutually exclusive features are marked for attribution decoupling; The final grouped task-specific feature subspace set and 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 by, The generation of the shared bottom layer feature tensor specifically is: configuring a bottom layer weight sharing structure in a multi-branch neural network architecture, and inputting the feature subspace of each task branch according to the parameter-task mapping table, and the bottom layer feature vector is dimensionally reduced by principal component analysis to output a shared feature tensor.

8. 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 each transmission line type target impedance prediction confidence interval, feature parameter contribution degree ranking, physical consistency statistical distribution and process improvement knowledge points.

Images