A multi-layer board pressing parameter intelligent prediction and regulation method

Abstract

The application discloses a kind of multilayer board pressing parameter intelligent prediction and regulation method, including the standardization acquisition and synchronization of multi-source heterogeneous process data, the space-structure-process joint modeling of multilevel high-order tensor model and the local / global characteristic factor disintegration of multi-scale tensor decomposition algorithm, weighted fusion feature vector is input into physical constraint neural network model by adaptive weight fusion, generates pressing parameter prediction result, for pressing parameter prediction result and real-time quality feedback information, the dynamic monitoring of prediction accuracy is executed, and model self-evaluation and adaptive learning are triggered, online optimization feature fusion weight and prediction strategy, the present application realizes the explainability of pressing parameter prediction and process safety guarantee, and promotes pressing whole-process dynamic closed-loop optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent modeling and parameter prediction technology for multilayer board lamination processes, and particularly to a method for intelligent prediction and control of multilayer board lamination parameters. Background Technology

[0002] Currently, parameter modeling and intelligent prediction technologies for multilayer board lamination processes have become crucial for improving product quality, reducing production costs, and achieving intelligent equipment control, driven by the upgrading of the electronics manufacturing and composite materials industries. Mainstream technologies typically focus on rule-based modeling based on process experience, classical statistical modeling, or shallow machine learning algorithms, which can predict lamination parameters for conventional board types and operating conditions to a certain extent. However, as multilayer board designs evolve towards higher integration, heterogeneous materials, and complex microstructures, the number of characteristic variables involved in the lamination process increases dramatically. The parameter space exhibits characteristics such as high dimensionality, variability, and strong nonlinear interactions, posing higher demands for the collaborative modeling and accurate mapping of higher-order features.

[0003] Currently, common parameter prediction and intelligent control methods in the field of multilayer board lamination mainly include statistical analysis based on single-dimensional feature extraction, big data-driven regression modeling, and some data-driven prediction methods such as deep neural networks. For example, several publicly available solutions have achieved multi-sensor data acquisition and remote monitoring, transforming physical quantities such as board temperature, pressure, and humidity into inputs for parameter prediction. However, these solutions often disperse or process information such as structural properties, material properties, and process history in a low-dimensional manner, lacking the ability to integrate multi-source heterogeneous and large-scale high-dimensional features; and when faced with new board types, high-frequency process changes, or abnormal structural distributions, the generalization ability and robustness of the modeling are limited.

[0004] In the field of high-dimensional heterogeneous data analysis and fusion, theories such as tensor modeling, tensor decomposition, and multi-scale feature fusion have been gradually introduced in recent years. Methods like Tucker decomposition, CP decomposition, and Tensor Train are applied to scenarios such as structural health monitoring and optimization of complex production processes. These technologies can theoretically achieve highly joint modeling of multi-source, multi-dimensional data, separating local and global features, and providing valuable tools for dimensionality reduction and anomaly detection in complex parameter spaces. However, existing applications are mostly limited to structural health diagnosis, sensor data compression, or anomaly detection in local scenarios. They are insufficient for the deep fusion and physical interpretation of material structure, process dynamics, and multi-level spatial information involved in multilayer board lamination, and most fail to achieve end-to-end intelligent prediction or real-time adaptive optimization of process parameters. Therefore, there is an urgent need for a method for predicting and controlling multilayer board lamination parameters that can overcome these limitations and possess strong high-dimensional feature fusion and regional adaptive capabilities. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent prediction and control method for multilayer board lamination parameters. The technical solution of this invention is implemented as follows: An intelligent prediction and control method for multilayer board lamination parameters, comprising:

[0006] S1: Collect multi-source heterogeneous data from different material types, structural layouts and process stages during the multilayer board lamination production process, including process raw material parameters, historical quality data, structural layout information and real-time sensor measurements, to form a raw dataset covering the diversity of board types and the differences in multiple working conditions.

[0007] S2: Perform data synchronization and standardization processing on the original dataset for multi-source heterogeneous datasets, including time alignment, feature dimension normalization, and outlier removal, to improve the consistency and robustness of subsequent modeling of high-dimensional features.

[0008] S3: Based on the standardized multi-source heterogeneous dataset, a multi-level tensor model is constructed. The structural properties, process parameters and sensor measurements of the multi-layer board are expressed in a high-order tensor according to the spatial dimension, hierarchical dimension, time dimension and process stage, so as to realize unified data modeling for variable structures and multiple working conditions.

[0009] S4: Apply a multi-scale tensor decomposition algorithm, including but not limited to Tucker decomposition, CP decomposition or Tensor Train decomposition, to the multi-level tensor model to effectively decompose local feature factors and global feature factors in the higher-order tensor space, so as to separate abnormal distributions, capture the highly correlated features of key structures and processes, and obtain multi-scale feature factors.

[0010] S5: Based on the multi-scale feature factors obtained by tensor decomposition, calculate the regional attention mapping of the plate structure partition and apply regional adaptive weight allocation. Differentiated high-dimensional feature enhancement and fusion are performed on key areas such as new material layering, special structural nodes and outlier distribution to obtain a weighted fusion feature set, so as to fully express the feature information under specific plate types and special working conditions.

[0011] S6: Input the weighted fusion feature vector fused by adaptive weights into the physical constraint neural network. Combine the physical constraints of embedded domain experts and process safety boundaries, and realize the deep fusion of data-driven features and physical mechanism mapping through knowledge distillation mechanism to generate compression parameter prediction results with high interpretability and wide transferability.

[0012] S7: Based on the predicted results of the pressing parameters and the real-time quality feedback information, perform dynamic monitoring of the prediction accuracy. If a new structural distribution or a prediction deviation exceeding the threshold is detected under abnormal working conditions, the model self-evaluation and abnormal adaptive learning are triggered to optimize the feature fusion weights and prediction strategies online, so as to continuously improve the adaptability to complex high-dimensional scenarios.

[0013] S8: The optimized compression parameter prediction results are output as closed-loop control suggestions to the compression process control, driving the production equipment to realize personalized real-time compression process adjustment, and recording key parameters and model feedback logs to support subsequent performance traceability and continuous model iteration.

[0014] The intelligent prediction and control method for multilayer board lamination parameters provided by this invention has the following beneficial effects:

[0015] (1) Significantly improves high-dimensional feature fusion capability and prediction accuracy. By combining multi-layer plate structures, process parameters, and process dynamic sensing data in a multi-dimensional manner (spatial, hierarchical, temporal, and process stage) into high-order tensors, the high-dimensional variation features of complex structures, heterogeneous multi-materials, and multiple process stages are effectively captured and expressed. Multi-scale tensor decomposition (Tucker / CP / TT) performs structured decomposition and redundancy reduction of local anomalies, global features, and cross-level correlations. Compared with traditional planar feature mixing methods, it can improve the signal-to-noise ratio and physical discriminability of feature expression. Experiments show that under high complexity and multiple novel plate types, the mean square error (MSE) of parameter prediction can be reduced by more than 20%.

[0016] (2) Enhanced Key Region Feature Discrimination and Adaptation to Abnormal Operating Conditions. The proposed regional attention weighting and adaptive feature allocation mechanism dynamically increases the weight of key plate type partitions such as new material layers, microstructure nodes, and abnormal fluctuation regions during the model learning process, achieving "fusion of differentiated complex structures." This effectively solves the problems of sparse anomalies being easily masked and the prediction of special partition parameters being easily distorted in existing high-dimensional heterogeneous scenarios. When abnormal operating conditions or new structures are put into production for the first time, the model prediction deviation automatically alarms and triggers self-learning optimization, ensuring that the output results still have high robustness and generalization ability under unconventional conditions, and expanding the scope of application from conventional plate types to high-end customized complex plate types.

[0017] (3) Achieving interpretability and process safety assurance in intelligent prediction of pressing parameters. By embedding domain expert physical constraints and process safety boundary criteria, a knowledge distillation mechanism is used to integrate data-driven prediction with physical mechanism modeling. The output parameter results have physical interpretability and model transferability. This effectively avoids the problem of "distorted" prediction outside the physical boundary of pure data-driven models, improves the scientific nature and safety controllability of pressing process parameter optimization, and reduces downstream process failures and safety risks.

[0018] (4) Promote dynamic closed-loop optimization of the entire multilayer board lamination process. Real-time collection of lamination process and quality feedback, and online comparison of prediction results with actual quality closed-loop. For structural changes and abnormal batches, adaptive updates of feature fusion weights and prediction strategies enable continuous model evolution, effectively avoiding the problems of long-term static training and performance aging of traditional models. Field applications show that the model adaptive mechanism can shorten the convergence time of abnormal prediction correction in the early stage of production by 50%, significantly improving the level of intelligence and production yield under highly variable working conditions. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for intelligent prediction and control of multilayer board lamination parameters according to the present invention. Detailed Implementation

[0020] 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.

[0021] 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.

[0022] 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.

[0023] Please see Figure 1 As shown, a method for intelligent prediction and control of multilayer board lamination parameters includes:

[0024] S1: Collect multi-source heterogeneous data from different material types, structural layouts and process stages during the multilayer board lamination production process, including process raw material parameters, historical quality data, structural layout information and real-time sensor measurements, to form a raw dataset covering the diversity of board types and the differences in multiple working conditions.

[0025] S2: Perform data synchronization and standardization processing on the original dataset for multi-source heterogeneous datasets, including time alignment, feature dimension normalization, and outlier removal, to improve the consistency and robustness of subsequent modeling of high-dimensional features.

[0026] S3: Based on the standardized multi-source heterogeneous dataset, a multi-level tensor model is constructed. The structural properties, process parameters and sensor measurements of the multi-layer board are expressed in a high-order tensor according to the spatial dimension, hierarchical dimension, time dimension and process stage, so as to realize unified data modeling for variable structures and multiple working conditions.

[0027] S4: Apply a multi-scale tensor decomposition algorithm, including but not limited to Tucker decomposition, CP decomposition or Tensor Train decomposition, to the multi-level tensor model to effectively decompose local feature factors and global feature factors in the higher-order tensor space, so as to separate abnormal distributions, capture the highly correlated features of key structures and processes, and obtain multi-scale feature factors.

[0028] S5: Based on the multi-scale feature factors obtained by tensor decomposition, calculate the regional attention mapping of the plate structure partition and apply regional adaptive weight allocation. Differentiated high-dimensional feature enhancement and fusion are performed on key areas such as new material layering, special structural nodes and outlier distribution to obtain weighted fusion feature vectors, so as to fully express the feature information of specific plate types and special working conditions.

[0029] S6: Input the weighted fusion feature vector fused by adaptive weights into the physical constraint neural network. Combine the physical constraints of embedded domain experts and process safety boundaries, and realize the deep fusion of data-driven features and physical mechanism mapping through knowledge distillation mechanism to generate compression parameter prediction results with high interpretability and wide transferability.

[0030] S7: Based on the predicted results of the pressing parameters and the real-time quality feedback information, perform dynamic monitoring of the prediction accuracy. If a new structural distribution or a prediction deviation exceeding the threshold is detected under abnormal working conditions, the model self-evaluation and abnormal adaptive learning are triggered to optimize the feature fusion weights and prediction strategies online, so as to continuously improve the adaptability to complex high-dimensional scenarios.

[0031] S8: The optimized compression parameter prediction results are output as closed-loop control suggestions to the compression process control, driving the production equipment to realize personalized real-time compression process adjustment, and recording key parameters and model feedback logs to support subsequent performance traceability and continuous model iteration.

[0032] S1: Collects multi-source heterogeneous data from different material types, structural layouts, and process stages during the multilayer board lamination production process, including process raw material parameters, historical quality data, structural layout information, and real-time sensor measurements, to form a raw dataset covering the diversity of board types and the differences in multiple working conditions, including:

[0033] S1.1 sets up sensor acquisition nodes for each major process step in the multilayer board production site to acquire real-time sensor measurements such as temperature, pressure, humidity and interlayer thickness, so as to obtain process sensor data reflecting the dynamic changes in the pressing process and provide time series input for subsequent high-order tensor expression.

[0034] To address the process monitoring needs of multilayer board lamination production sites, high-sensitivity industrial sensors are configured for each lamination equipment station, work section, and key material conversion node to achieve real-time acquisition of multi-parameter process data.

[0035] A multi-point temperature sensor array (configured with PT100 high-precision temperature sensor, measurement range: -50℃~250℃, resolution ≤0.1℃) is distributed and installed in the heating plate, plate clamping area and ambient space of the pressing machine to achieve multi-point continuous monitoring of the temperature of each area during the pressing process.

[0036] Furthermore, by setting up high-precision pressure sensors (configuration: strain gauge pressure sensor, measurement range 0-10MPa, linearity not exceeding 0.2%FS) and deploying them at key nodes such as the main pressure roller, hydraulic cylinder and interlayer support plate, real-time pressure is sampled differentially and data redundancy is detected to capture pressure change curves in typical process stages such as dynamic pressure increase, pressure holding and pressure reduction during pressing.

[0037] Furthermore, by integrating a semiconductor capacitive humidity sensor (range 0~100%RH, error ≤1.5%RH) and deploying it in the airflow channel, plate space, and inlet / outlet of the pressing equipment, the system can monitor the humidity fluctuations in the pressing environment and perform temperature and humidity coupling calibration in conjunction with synchronous temperature output.

[0038] Furthermore, a laser displacement / ultrasonic thickness sensor (accuracy ≤1μm) is used to perform non-contact multi-point measurement of interlayer thickness and key bonding layer gap. By comparing time-series fluctuations through high-frequency acquisition, the dynamic relationship between interlayer thickness and pressure and temperature changes is reflected.

[0039] An industrial-grade edge data acquisition device (parameters: sampling frequency ≥10Hz, buffer depth ≥10,000 groups) is used to perform multi-channel synchronous acquisition and timestamp allocation of the raw signals collected by the aforementioned temperature, pressure, humidity, and thickness sensors in a time-division manner, forming a structured process sensing raw data stream.

[0040] The collected data is uploaded to the pressing data management platform in real time via industrial Ethernet or bus protocol, input to the time synchronization preprocessing module, and output a multi-source real-time process sensing data chain, providing a complete and accurate dynamic data foundation for subsequent time series input based on high-order tensor modeling.

[0041] Through the above-mentioned multi-parameter process sensing layout and high-frequency synchronous acquisition processing method, the various sensor data deployed in the previous step are efficiently transformed into multi-dimensional time series data of the entire pressing process, realizing full coverage of dynamic process characteristics and spatiotemporal continuity of data, thereby effectively supporting the technical goal of unified modeling of high-dimensional data of complex pressing structures.

[0042] For example, in a real multilayer board high-speed lamination production line, PT100 temperature sensors are placed at three points (left, center, and right) on the heating plates of each of the four lamination machines. Strain gauge pressure sensors are placed between the hydraulic cylinder of the main press and the support plates on both sides. Five capacitive humidity sensors are installed in the environment, and two laser thickness sensors are installed in the critical clamping area. All sensor data are collected in real time by an edge acquisition device at a sampling frequency of 20Hz and transmitted online using the Modbus TCP protocol. The system achieves a 1-second response time for the entire lamination process, and the raw data is used for high-order tensor time series modeling in the backend. Field verification data shows that, through the above networking and acquisition configuration, the system can capture 100% of the process dynamic information, with a sensor data acquisition loss rate of less than 0.03%, and a consistency error of less than 0.5% in temperature, pressure, and thickness measurements. The ability to capture fluctuations in the real process is significantly better than that of the traditional single-point / low-frequency acquisition mode. The final output dataset contains complete multi-dimensional time series data of temperature, pressure, humidity, and interlayer thickness. Each data point is accompanied by key metadata such as a unique sensor number, pressing batch, and timestamp, effectively supporting subsequent high-dimensional feature tensor quantization representation and intelligent prediction of process parameters for complex structures in multiple scenarios.

[0043] S1.2 Collects physicochemical performance parameters of various multilayer board materials, including basic information such as resin content, copper foil thickness, dielectric constant, glass fiber type and material batch, and summarizes them through a structured data interface as input for structural property modeling.

[0044] Using the on-site multilayer board raw material management database and supplier batch delivery records as input data sources, the system retrieves the registration files of various materials used in the multilayer boards to be laminated, covering categories such as resin prepreg, copper foil, prepreg, substrate, reinforcing materials, and functional fillers.

[0045] The automated material traceability code collection technology (parameter: RFID barcode / QR code data collection terminal) is adopted to realize the automatic association and identity registration of the production date, supplier batch number and physical label of each batch of materials, ensuring the unique traceability of the board type and raw materials.

[0046] Furthermore, based on the material list, the integrated physicochemical property testing module is used to conduct batch performance index testing and data recording for key raw materials (such as resin prepreg, copper foil, fiberglass cloth, epoxy resin, etc.), including but not limited to resin content (mass percentage), copper foil thickness (μm), dielectric constant (ε_r), glass fiber type (E-glass / S-glass, etc.), moisture content (%), molecular weight distribution, coefficient of thermal expansion, etc., to generate a preliminary database of physicochemical performance parameters.

[0047] The test data is structured and organized using automated data interfaces (such as MES system API and material LIMS system interface), compared with the warehousing standards, and parameter consistency verification and outlier screening are performed. Unqualified or defective batches are removed by the pass / fail judgment algorithm, and the parameters of the batches that pass are standardized.

[0048] Furthermore, according to the matching rules of material type and plate type, the physicochemical parameters of all materials that have passed the verification are structured and assigned feature labels. Multi-field nested indexes (such as material type_batch_attribute key) are used to establish a ternary structured data list of multi-layer plate structural attributes-material batch-physicochemical performance parameters.

[0049] Through the above-mentioned structuring of feature parameters and batch label binding, the physicochemical performance parameters of various multilayer board materials are efficiently summarized into structured tables or tensors, providing a solid data foundation for subsequent multidimensional structural attribute modeling and highly correlated feature modeling, and realizing high-quality and traceable input of multi-source material information.

[0050] For example, in the production process of high-end HDI multilayer boards, for a certain batch of mixed laminated boards, the prepreg used is batch number HJ20230601 from Company A, with a resin content of 37.4 wt%, E-glass fiberglass cloth, copper foil thickness of 18 μm / 12 μm for the upper and lower layers respectively, dielectric constant ε_r=4.5±0.05, moisture content of 0.07%, and the coefficient of thermal expansion of epoxy resin, the main material of the board, is 38 ppm / K. All the above parameters are automatically collected through the LIMS interface and automatically mapped with the material barcode RFID and the board production code.

[0051] All material parameters are standardized by Z-score and tested for batch consistency. Batch numbers with excessive moisture content and abnormal copper foil thickness are removed. Finally, a set of structured material physicochemical parameters is output, which includes five fields: material name, batch number, attribute label, parameter value and test timestamp. On average, each plate type is associated with 6-10 material data.

[0052] After the structured data mentioned above is aggregated through a unified standard interface, it is imported into the original dataset of multilayer board lamination along with the process sensing data, realizing a closed-loop data chain of "board type-batch-material-parameter", effectively supporting the high-dimensional feature input requirements for three-dimensional structural modeling and parameter prediction.

[0053] S1.3 For historical production batches, obtain the formulation parameters of each pressing process, including the pressurization curve, holding time, lamination path and cooling rate, and collect the corresponding quality test results, including warpage, interlayer peel strength and micropore filling rate, to establish a historical data mapping of materials-process-quality.

[0054] S1.4 For specific multilayer board products, extract detailed structural layout information, such as the number of layers, wiring density per layer, distribution of through holes / blind holes / micro holes, board area specifications, and stiffener design, and output it in a standardized manner through CAD structural files or data interfaces to ensure the complete correlation between structural parameters and process parameters.

[0055] S1.5 integrates raw process sensing data, material physicochemical parameters, historical process formulas and quality inspection results, and structural layout information. It uses unique batch identifiers for plate type and manages them by tagging, forming a multi-source heterogeneous raw dataset for multilayer plate lamination production. This provides multi-dimensional input conditions for subsequent unified preprocessing and modeling of high-dimensional features.

[0056] Step S2: Perform data synchronization and standardization processing on the original dataset, including time alignment, feature dimension normalization, and outlier removal, to improve the consistency and robustness of subsequent modeling of high-dimensional features, specifically including:

[0057] S2.1

[0058] Timestamp identification and reordering of multi-source heterogeneous raw datasets are performed to synchronize data on a unified timeline for different structural layouts, process stages and material sources, so as to obtain time-aligned datasets for the entire multilayer board lamination process.

[0059] S2.2 Based on the time-aligned dataset, a feature dimension normalization algorithm is applied to perform interval scaling or Z-score normalization on structural attribute parameters, process variables, and real-time sensor measurements to generate a standardized feature matrix with a consistent feature scale, thereby achieving a balance of influence weights among multi-dimensional features.

[0060] S2.3 applies industrial anomaly detection algorithms to the standardized feature matrix, including mechanisms based on IQR (interquartile range), Z-score threshold, and local outlier factor, to identify and remove outliers such as mutations and drifts that do not conform to physical laws, in order to obtain feature input samples with high reliability and high robustness.

[0061] S2.4 performs missing data completion processing on the feature input samples after outlier removal. By interpolation imputation, low-rank matrix reconstruction or autoencoder-based reconstruction algorithms, the missing intervals in historical quality data and structural parameter collection are filled to obtain a complete high-dimensional standardized dataset.

[0062] S2.5 regroups and verifies the consistency of the high-dimensional standardized dataset based on the original category labels of the multi-source heterogeneous data. Through statistical tests of distribution uniformity and elimination of contradictions in the label data, it finally outputs high-quality standardized training samples of multilayer board lamination that meet the consistency requirements of high-dimensional features, providing basic data support for high-order tensor modeling.

[0063] Step S3: Based on the standardized multi-source heterogeneous data, a multi-level tensor model is constructed. The structural properties, process parameters, and sensor measurements of the multi-layer board are expressed in a high-order tensor format according to spatial dimension, hierarchical dimension, time dimension, and process stage, thereby achieving unified data modeling for variable structures and multiple operating conditions. Specifically, this includes:

[0064] S3.1 classifies the standardized multi-source heterogeneous data by attribute classification. The original data is divided into three categories: structural attributes, process parameters and sensor measurements by template mapping algorithm, so as to construct the basic attribute dimensions of the multi-level tensor model.

[0065] S3.2 Utilizes structural layout rules to perform spatial hierarchical indexing and reorganization of structural attribute data. Through a multi-level nested spatial partitioning method, structural attributes are mapped to the spatial and hierarchical dimensions of tensors, thereby achieving unified processing of spatial information of different structural units.

[0066] Based on the categorized structural attribute data, the input includes parameters such as material type, spatial distribution coordinates, trace density, spatial location of vias / blind vias / microvias, board area, layer thickness, and reinforcing rib area for each layer of the multilayer printed circuit board.

[0067] A structural partitioning rule algorithm (parameters: spatial layout baseline, hierarchical segmentation strategy) is adopted to perform spatial partitioning on all structural attribute data, establish a multi-level spatial mapping index, and realize the preliminary division of the spatial geographical location, adjacency relationship and partition hierarchy of structural units.

[0068] Furthermore, by using a multi-level nested spatial partitioning method (parameters such as Quad-Tree spatial index model and recursive hierarchical partitioning depth), the structural attributes of each layer are progressively grouped by region, dividing the main layers such as top-middle-bottom into several spatial sub-blocks. At the same time, within each block, sub-layers are automatically further divided according to local features such as reinforcing ribs, special wiring, and micro-hole clusters.

[0069] A spatial hierarchical index reorganization algorithm (parameters: spatial ID mapping table, hierarchical depth label) is adopted to maintain the correspondence between the overall structural attribute data and the spatial / hierarchical index, generate the spatial dimension index table and the hierarchical dimension index table of structural attribute features and tensors, and form data blocks corresponding to the two main dimensions of space and hierarchy.

[0070] By using a normalized mapping function, the spatial partitioning and hierarchical indexing information mentioned above is encoded and embedded into the structural attribute tensor representation, forming a structural attribute tensor with multiple spatial dimensions and multiple levels of nested hierarchy, thereby realizing the unified positioning and indexing of various structural units in the tensor space.

[0071] By reorganizing spatial hierarchical indexes and using multi-level nested partitioning, structural attribute data is efficiently organized in the form of spatial-hierarchical dual-principal-dimensional structural blocks. This provides a complete and standardized structural information representation for subsequent multi-dimensional tensor combinations with process parameters and sensor measurements, thereby achieving spatially unified processing of multi-source structural attribute information.

[0072] For example, in a high-end PCB lamination production scenario involving a 12-layer asymmetric multilayer structure, a board size of 600mm × 450mm, and densely packed micro-blind vias in localized areas, the normalized input structural attribute data uses a two-dimensional X and Y coordinate system as the spatial layout reference. A Quad-Tree spatial index model is adopted, with a segmentation depth of 5 layers. Each main layer (e.g., layer 1, layer 6, and layer 12) is refined into 16 spatial sub-blocks. Boundary regions are further divided into 64 sub-layers based on micro-via density and reinforcing rib positions. A spatial ID mapping table is used to maintain the mapping relationship between spatial indexes and structural attributes, and the structural attribute data is reorganized into tensor blocks of (number of spatial partitions × layer depth × number of structural features). For example, the spatial ID of the micro-via cluster area in layer 6 is (3, 2, 58), corresponding to a trace density of 85mil, a local board thickness of 100μm, and a material type ID of 29. By reorganizing the multi-level spatial indexes described above, the unique location of all structural units of the complex multi-layer plate in the high-order tensor is achieved, and a spatially consistent index standard is provided for subsequent multi-dimensional feature fusion algorithms, thereby improving the accuracy and scalability of high-dimensional feature representation in complex structural scenarios.

[0073] S3.3 performs time alignment and stage labeling on the process parameter data according to the time sequence information of the process stage, and adopts a process segment hierarchical algorithm to encode the participation of process parameters in each process stage and the whole process, thereby mapping the process parameters to the time dimension and process stage dimension of the tensor.

[0074] For the processed process parameter data, the process stage timing parsing module is called (input: process parameter standardization matrix, with timestamp and batch unique identifier). Based on the preset pressing production process nodes (such as pressurization, constant temperature, heat preservation, cooling, etc.) and their stage start and end time windows, the timing block division of the process parameters is obtained.

[0075] A time alignment algorithm (parameters: time window width Δt, start and end times of each process segment T_start, T_end) is used. For each data point t in the original process parameter sequence, its corresponding process segment index K is calculated, and the allocation is performed according to a recursive judgment formula: if T k,start ≤t <T k,end Then K = k, thus labeling the continuous time-series process parameters into a stage grouping matrix.

[0076]

[0077] Among them, G i,k Let P be the set of values ​​for the i-th process parameter in the k-th process stage. i (t) represents the standardized value of process parameter i at time t.

[0078] Furthermore, a process segment hierarchical coding algorithm (parameters: process segment type table, parameter hierarchical threshold set) is introduced to hierarchically label the range of process parameter changes within each stage. For example, for the pressure parameters in the pressurization stage, multiple behaviors such as pressure increase, constant, and decrease are set according to the pressure gradient and control nodes to generate nested stage sub-labels.

[0079] By combining the aforementioned time-series grouping with hierarchical sub-labels through the feature mapping function, the process parameter tensor is given a dual index of time dimension and process stage dimension, realizing a high-order tensor expression of process parameters in the whole process scenario of "process segment-stage sub-level-time".

[0080] Combining spatial-hierarchical structure indexing, for process parameters of the same spatial location or hierarchical unit, according to the above segmented and layered mapping rules, they are associated with higher-order tensors using the following formula:

[0081] T s,l,k,t =f map (P i | (s,l) G i,k ,L k,x )

[0082] Where T s,l,k,t L represents the standardized process parameter tensor values ​​under space s, level l, process stage k, and time t. k,x The label for the xth sub-level within stage k.

[0083] Through the above process, the original temporal variation patterns, stage characteristics, and hierarchical behavior of process parameter data are fully mapped into the tensor data structure, realizing the dynamic association and hierarchical description of process parameters in high-dimensional tensors for multiple scenarios.

[0084] By using process stage temporal alignment and hierarchical coding algorithms, multidimensional process parameter data is transformed from the original temporal space into a tensor mapping matrix with a combination of time dimension and process stage dimension index. This enables a complete structured expression of high-dimensional dynamic process features under complex pressing processes, providing a consistent data foundation for process parameter scenarios for subsequent joint modeling of multi-source data and high-order feature fusion.

[0085] For example, in a multi-layer PCB mass production scenario with a complex lamination process, the time intervals for the pressurization stage (T1), isothermal stage (T2), heat preservation stage (T3), and cooling stage (T4) are set to [0,30] min, [30,90] min, [90,110] min, and [110,180] min, respectively. The total pressure P(t) data recorded on the single board is input after standardization in 1-minute units. A layering rule is used to classify the intervals with a pressure change rate >10 kPa / min within the pressurization stage as incremental sub-layers, and the constant section as constant pressure sub-layers. The actual pressurization process parameter tensor G_{P,T1} contains 30 data points in the incremental sub-layer; the isothermal parameter G_{T,T2} records 60 data points in the isothermal sub-layer. All the above segmentation and layering labels are mapped and stored to the time and stage dimensions of the tensor through a unique batch identifier, realizing the dynamic association between the single board's full-process parameters and stage status. The output results can accurately describe the process parameter variation curves of any spatial partition and hierarchical unit at each process stage and sub-layer, providing a complete time-stage mapping basis for subsequent abnormal area identification and parameter prediction.

[0086] S3.4 Based on the acquisition timestamps and spatial acquisition location information of real-time sensor measurements, a spatiotemporal synchronous embedding method is adopted to combine dynamic sensing data with structural attributes and process parameters according to the corresponding index dimensions to form a data block with a unified index at multiple spatiotemporal levels.

[0087] For the standardized dynamic process sensor measurements, structural attribute tensors, and process parameter tensors, the inputs include timestamp records and spatial acquisition coordinate information of various sensors (temperature, pressure, humidity, interlayer thickness, etc.), spatial / hierarchical index tables of structural attribute tensors, and process stage / time index tables of process parameter tensors.

[0088] A spatiotemporal synchronization embedding method (parameters: time window Δt, spatial acquisition ID, sensor node coordinates, structural spatial index table, hierarchical index table, and process stage label) is adopted to achieve precise alignment between the time series of dynamically sensed data and the spatial layout.

[0089] Furthermore, using a joint index mapping algorithm, based on each time t and acquisition position p, the spatial / hierarchical coordinates (s,l) in the structural attribute tensor and the process stage / time sub-index (k,t) in the process parameter tensor are searched respectively, thus connecting the sensor measurement values ​​V under the same spatial unit, the same process stage, and similar time windows. p,t With the corresponding structural attribute vector F s,l and process parameter vector P k,t Perform synchronous data collection.

[0090] Furthermore, a multi-dimensional data block generation function is used to fill multi-level, multi-temporal indexed data blocks by combining the synchronized structural attributes, process parameters, and sensor measurement triples according to the index (s,l,k,t).

[0091] B s,l,k,t =[F s,l ,P k,t V p,t ]

[0092] Among them B s,l,k,t F represents a unified data block under a certain spatial partition s, level l, process stage k, and time t. s,l P represents the structural attribute eigenvalue. k,t V represents the characteristic value of the process parameter. p,t These are the sensor measurements corresponding to the spatiotemporal location.

[0093] Furthermore, to address the spatial sparsity or temporal asynchrony issues present in actual sensor deployments, missing regions are filled in using nearest neighbor interpolation or masking methods based on spatial IDs.

[0094] Furthermore, through consistency verification and outlier synchronous removal mechanisms, data blocks in the combined block that have mismatched timestamps, spatial index conflicts, or abnormal sensor measurements are automatically detected and labeled as missing or invalid to ensure the data quality of subsequent tensor quantization construction.

[0095] Through the above multi-level processing, dynamic sensing data, structural attributes, and process parameters are integrated and combined under the same spatial-hierarchical-process stage-temporal multi-dimensional index, realizing a unified index of data blocks in multiple time and space levels, laying the foundation for the structured construction of high-order tensor models.

[0096] For example, in a high-end PCB lamination process scenario of 600mm×450mm 12 layers, 16 pressure sensors, 8 temperature sensors, and 4 thickness sensors are deployed in real time, with a sampling frequency of 1Hz. The spatial ID is determined by the (X,Y) partition of the board surface. The spatial partition numbers of the structural attribute tensor are 0-63, and the layer numbers are 1-12. The process stages of the process parameter tensor include four categories: pressurization, isothermal control, heat preservation, and cooling, with a time window Δt = 1min, and the process stage sub-interval numbers are 0-179.

[0097] At any given time t0 = 87 min, spatial partition s = 15, level l = 8, and process stage k = 3 (insulation), the system compares the collected P (s = 15, t = 87), F (s = 15, l = 8), and V (p = 15, t = 87) and assembles them into a unified data block B. 15,8,3,87 =

[0098] [F 15,8 ,P 3,87 V 15,87 If spatial partition 18 misses sensor data at t=87min, then weighted interpolation of measurements from partitions 17 and 19 is used to fill in the missing data. All data blocks with timestamps or spatial conflicts are automatically marked as invalid and exit tensor quantization.

[0099] Finally, the output is a multi-dimensional data block set consisting of spatial partitioning, hierarchical numbering, process stage, and time window dimension, which can be directly called by downstream high-order tensor quantization modeling functions to realize the information synchronization and efficient integration of multi-dimensional heterogeneous features in complex process scenarios, effectively supporting the integrated intelligent prediction modeling of structure and process in high-dimensional complex parameter space.

[0100] S3.5 takes multi-level tensor data blocks as input, uses high-order tensor construction functions, and performs structured operations according to spatial dimension, hierarchical dimension, time dimension and process stage dimension to realize high-order tensor quantization expression of the full parameter system of multi-level boards, generate standardized multi-level tensor models, and provide a unified and complete data foundation for subsequent multi-scale tensor decomposition and feature fusion.

[0101] S4: Apply a multi-scale tensor decomposition algorithm, including but not limited to Tucker decomposition, CP decomposition, or Tensor-Train decomposition, to the multi-level tensor model to effectively decompose local and global feature factors in the higher-order tensor space, so as to separate abnormal distributions and capture highly correlated features of key structures and processes. Specifically, this includes:

[0102] S4.1 applies the tensor dimension initialization method to the standardized multi-level tensor model. Based on the spatial dimension, hierarchical dimension, time dimension and process stage of the multi-layer board structural properties, process parameters and sensor measurement values, the decomposition benchmark dimension is set for the high-order tensor model to obtain the basic structural distribution of the multi-dimensional feature space.

[0103] For the input of the standardized multi-level tensor model, including multi-dimensional data blocks based on spatial partitioning, hierarchical indexing, process stages, and time window distribution, a tensor dimension initialization method is used (parameter: spatial dimension N). s Hierarchical Dimension N l Time dimension N t N-stage process dimension k This enables the setting of dimensional baselines and structural expression of multi-level tensor models.

[0104] Applying the tensor structure scan algorithm (parameter: set of dimension descriptors {D) i The tensor dimension maximization criterion discretizes each dimension of the input tensor, calibrates the specific dimension range and sequence length of all spatial partitions, hierarchical units, time windows and process stages, and obtains the matrix morphological constraints of the tensor model in multidimensional space.

[0105] Furthermore, through multi-source data association mapping analysis (parameters: partition index, stage mutual exclusion table, synchronization mask array), the cross-interaction areas of each main dimension of space / hierarchy / time / process are verified, generating an effective data distribution block index set that supports tensor decomposition, and clarifying the input structure boundary of subsequent decomposition algorithms.

[0106] A high-order tensor structure consistency check method is adopted (parameter: block sparsity threshold θ). sp Time integrity threshold θ t The algorithm identifies empty blocks, abnormal indexes, and asynchronous data regions within the tensor structure, dynamically removes abnormal data blocks, and performs dimensional filtering optimization to ensure that the tensor decomposition algorithm can obtain high-quality, physically consistent input tensors.

[0107] By setting the tensor baseline dimension and optimizing the consistency of higher-order dimensions, the original multi-source high-dimensional data is uniformly organized into decomposable high-order tensors under multiple principal dimensions of space, hierarchy, time, and process. This provides a rigorous data foundation for multi-scale tensor decomposition (Tucker, CP, Tensor Train, etc.) and enables a comprehensive revelation of the multi-dimensional feature space structure distribution.

[0108] For example, in a complex PCB lamination project (board size 600mm×450mm, 12 layers, dense area of ​​special micro-blind vias, 8 types of materials, production cycle 180min, sampling step size 1min, 4 process stages), the input tensor dimension initialization parameter N is... s =64 (spatial partitions), N l =12 (levels), N k =4 (process stage), N t =180 (time trajectory points). Using tensor structure scanning, a dimensional descriptor is established for the original multi-level tensor model: spatial partition numbers 0-63, level numbers 1-12, stage indices 1-4, and time windows 0-179. Through multi-source data association mapping, data within the audit space 15–18, levels 9–10, the insulation stage, and the 90–110 min interval are identified as densely populated anomaly areas with micro-blind holes, and are designated as blocks of special interest for decomposition. Using a block sparsity threshold of 0.1 and a time integrity threshold of 0.05, data blocks with high missing rates are removed and imputed, outputting a high-order complete tensor (64×12×4×180) of space × level × process stage × time window. Ultimately, under the initialization of tensor dimensions, all spatial partitions, hierarchical units, and process time periods of multi-source heterogeneous data are rigorously organized to form a unique and comparable tensor index system, providing standardized input for the subsequent Tucker / CP / TT decomposition output of high-confidence structure, ontology, time series, and other local-global feature factors.

[0109] S4.2 Based on the decomposition baseline dimension of the multi-level tensor model, multi-scale tensor decomposition algorithms such as Tucker, CP, or Tensor Train are applied to perform multi-level tensor space decomposition on the high-order tensor model, generating local feature factors and global feature factor combinations containing multiple layers of features including structural dimension, ontology dimension, and temporal dimension, thus initially realizing the subspace decomposition of high-dimensional complex data.

[0110] The input object is a standardized multi-level tensor model, which includes a multi-principal dimension structure such as initialized spatial partitions, level numbers, time windows, and process stages.

[0111] A multi-scale tensor decomposition algorithm (parameters: Tucker decomposition rank [space R_s, level R_l, time R_t, process R_k], CP decomposition rank R_c, maximum chain length of Tensor Train L_tt) is adopted to decompose the high-order tensor model into multi-level tensor space under each principal dimension.

[0112] Furthermore, through a dimension-wise decomposition mapping mechanism, the tensor is... Using space, hierarchy, process stage, and time as decomposition axes respectively, and employing master decomposition operations such as Tucker decomposition:

[0113]

[0114] in, For the core tensor, U (s) U (l) U (k) U (t) These are the feature basis matrices for space, hierarchy, process stage, and time, respectively.

[0115] Furthermore, using the CP decomposition method, the aforementioned higher-order tensor is expanded into a superposition of R sets of vector outer products, specifically:

[0116]

[0117] Where, λ r As weight, To decompose the feature vectors obtained in the spatial, hierarchical, process stage, and temporal directions, the outer products of each group form a set of local and global feature factors.

[0118] Furthermore, for ultra-high-dimensional complex data scenarios, the Tensor Train decomposition algorithm (parameters: decomposition chain length L_tt, truncation threshold ε) is adopted to recursively decompose large-scale high-order tensors into a series of three-dimensional tensor blocks, progressively decomposing the original data blocks into serialized weight representation chains.

[0119] Through a multi-level defined tensor decomposition process, local feature factors (such as structural dimension, hierarchical dimension, temporal dimension, and technological dimension) under each principal dimension are obtained. (etc.) and global feature factors (core tensor) CP weight λ r (TT sequence intermediate kernel block, etc.).

[0120] By decomposing and combining multi-scale feature factors in the output, the original complex data in the high-order tensor is transformed into feature parameters with multi-layered expressive capabilities such as structure, ontology, and temporal sequence, thus achieving preliminary decomposition and feature representation of the high-dimensional complex data subspace.

[0121] For example, in the complex lamination process of a 12-layer high-end PCB with dimensions of 600mm × 450mm, the input tensor The dimensions are 64 (spatial partitions) × 12 (levels) × 4 (process stages) × 180 (time order). Tucker decomposition is used, with decomposition ranks set as follows: spatial R_s = 10, level R_l = 6, process R_k = 3, and time R_t = 20. Alternating least squares (ALS) optimization is applied to the core tensor. With each basis matrix. After decomposition, we obtain The (10×6×3×20) core block and the U matrix respectively reflect the main change directions of each spatial partition, hierarchical unit, process stage, and time window. For the dense micro-blind hole region (such as spatial partition 18, level 9-10, heat preservation stage 130-170min, and time sub-segment), its prominent feature components on U^{(s)}{18}, U^{(l)}{9,10}, U^{(k)}{3}, and U^{(t)}{130-170} can be extracted. By benchmarking CP decomposition and setting R_c=30, 30 sets of CP feature vector combinations are obtained to capture the global and local implicit laws across space, level, and time. Tensor Train decomposition is adopted, with the chain length L_tt=4 and the decomposition error threshold ε=0.01, to obtain 4 reduced three-dimensional tensor kernel chains, compressing the original complex tensor to a highly interpretable low-order chain representation. In summary, under extremely high-dimensional, highly nonlinear, multi-level data scenarios, all types of tensor decomposition can effectively output local and global feature factors, completing the multi-scale subspace decomposition of complex structures and process scenarios. These feature factors provide high-quality, high-order structural information for subsequent anomaly distribution separation, feature fusion, and downstream intelligent parameter prediction, achieving systematic dimensionality reduction and accurate representation of high-dimensional features in complex PCB lamination.

[0122] S4.3 performs correlation calculations and feature redundancy removal on the local feature factors obtained from multi-level decomposition. Through feature entropy analysis and correlation weight evaluation, it focuses on core feature variables that are strongly correlated with process performance, reduces noise interference in high-dimensional information, and lays the foundation for more accurate anomaly distribution separation.

[0123] The input object is the local feature factor obtained by multi-level tensor decomposition, which includes the feature weight matrix of each main dimension of spatial partition, hierarchical number, process stage and time series.

[0124] The feature entropy analysis method (parameter: information entropy threshold θ_H) is used to calculate the entropy value of the component distribution of each local feature factor, thereby realizing the quantitative expression of feature information.

[0125] Furthermore, a correlation weighting evaluation algorithm (parameters: Pearson correlation coefficient threshold ρ, process quality label set C) is used to calculate the correlation between each local feature factor and process performance indicators (such as post-press warpage, interlayer peel strength, etc.), thereby ranking the correlation strength between feature factors and process performance indicators. The following formula is used:

[0126]

[0127] Where, r ij Let f be the Pearson correlation coefficient between the i-th local eigenfactor and the j-th process performance index. ikLet be the value of the i-th feature factor in the k-th sample. Let c be the mean of the i-th characteristic factor. jk Let j be the value of the process performance label in the k-th sample. Let be the average value of the j-th process index.

[0128] Furthermore, a feature redundancy elimination algorithm (parameter: correlation redundancy threshold ξ) is used to determine the correlation between pairs of local feature factors. If the correlation coefficient is higher than ξ, the feature with the highest correlation to the process performance index is retained and the redundant feature is deleted.

[0129] Furthermore, through feature contribution ranking and cluster analysis (parameter: feature contribution critical value α), feature factors that meet the requirements of high information entropy, high correlation and low redundancy are selected as the core feature variable set.

[0130] Through the above series of processes, the initial local feature factors obtained from the multi-level decomposition are continuously converged to the core feature expression that is strongly correlated with the performance of the pressing process and has little noise interference, laying a solid data foundation for subsequent anomaly distribution separation, feature region focusing and high-dimensional feature fusion.

[0131] For example, in a high-end PCB lamination process scenario with a 600mm×450mm, 12-layer structure and dense micro-blind vias, a preliminary set of 1200-dimensional local feature factors was obtained after multi-level tensor decomposition. Information entropy analysis was used, with a threshold θ_H = 1.5, to remove 230 low-variability features with information entropy less than 1.5. Combined with actual quality inspection labels (such as peel strength and warpage), the correlation between all remaining features and quality labels was calculated using Pearson correlation coefficients, with a correlation threshold ρ = 0.7. Only feature scores with a correlation coefficient higher than 0.7 with at least one performance indicator were retained, resulting in 310 highly correlated factors. For each of these highly correlated feature groups, Pearson coefficients were calculated pairwise, with a redundancy threshold ξ = 0.95. For feature groups with a correlation greater than 0.95, the feature with the highest correlation to quality was retained, and 195 redundant features were removed, forming a subset of 115 non-redundant highly correlated local features. Features were ranked by contribution and further categorized and aggregated based on a core contribution α = 0.09, resulting in 47 effective features for subsequent modeling. During the screening process, all statistical indicators were scored using the aforementioned formula, and a feature redundancy removal report was automatically generated, significantly improving the signal-to-noise ratio of target features in the high-dimensional local feature space. This enhanced the accuracy of subsequent abnormal distribution separation results and significantly strengthened the ability to detect outliers. Through feature entropy analysis, correlation weight evaluation, and a chain-like processing of redundancy removal algorithms, the massive local feature set output from the initial tensor decomposition was transformed into a high-confidence core bonding process-related feature set, achieving systematic noise reduction and enhanced discrimination capabilities for multi-dimensional correlated features in complex multilayer board structures.

[0132] S4.4 employs an anomaly distribution detection algorithm for global feature factor combinations. By utilizing statistical extreme value detection and feature distribution comparison, it identifies anomalous regions such as hidden asymmetric structures, novel material layers, and special nodes within high-order tensor models, and independently marks and tracks anomalous factors in the decomposition results.

[0133] S4.5 feeds back the abnormal region markers and core local feature factors to the tensor decomposition model, dynamically and adaptively adjusting the decomposition parameters and decomposition step size. Through the feedback optimization mechanism, the tensor decomposition process is iteratively upgraded, so that the local feature expression and the global structure representation are synchronously optimized, and the highly correlated features of key structures and processes are reliably captured.

[0134] S5: Based on the multi-scale feature factors obtained from tensor decomposition, calculate the region attention mapping of the plate structure partition, and apply region adaptive weight allocation. Perform differentiated high-dimensional feature enhancement fusion on key regions such as new material layering, special structural nodes, and outlier distribution to obtain a weighted fusion feature vector, thereby achieving full expression of feature information under specific plate types and special working conditions. Specifically, this includes:

[0135] S5.1 performs structural partitioning mapping on the multi-scale feature factors obtained by tensor decomposition, classifying the feature factors into structural partitioning feature matrices according to the spatial distribution, hierarchical division and process node information of the press plate type, so as to achieve the initial alignment of feature space and structural space.

[0136] S5.2 is based on the structural partition feature matrix and applies the region attention mechanism algorithm to calculate the region attention weight of the feature factors of each structural partition in order to measure the importance of the layering of new materials, special structural nodes and abnormal distribution regions in the overall feature expression, and realize the optimal selection of the weight of multi-scale feature factors.

[0137] S5.3 Based on the regional attention weight, the structural partition feature matrix is ​​processed by regional adaptive weight allocation, and the weights of multi-scale feature factors of new material layering, special nodes and abnormal areas are dynamically adjusted to form an adaptive weighted feature expression matrix, thereby improving the sensitivity of key regional feature expression.

[0138] S5.4 inputs the adaptive weighted feature expression matrix into the high-dimensional feature fusion algorithm. Through multi-dimensional correlation enhancement operations, it achieves sparse enhancement expression of structural interval and cross-regional interaction features, highlighting the advantageous feature factors under new plate types and abnormal working conditions.

[0139] S5.5 performs high-dimensional information redundancy removal and feature normalization on the feature fusion results after multi-level adaptive weighting and correlation enhancement, generating the optimal weighted fusion feature vector for specific plate types and special working conditions, providing input features with discriminative power and multi-scenario adaptability for the physical constraint neural network.

[0140] Step S6: The weighted fusion feature vector obtained through adaptive weight fusion is input into the physical constraint neural network. Combined with the physical constraints of embedded domain experts and process safety boundaries, a deep fusion of data-driven features and physical mechanism mapping is achieved through a knowledge distillation mechanism, generating highly interpretable and widely transferable compression parameter prediction results. Specifically, this includes:

[0141] S6.1 structurates the weighted fusion feature vector after the region adaptive weight enhancement fusion process and uses it as the standard input to the physically constrained neural network to ensure the complete transmission of feature information under the high-order tensor expression, laying the information foundation for subsequent model mapping.

[0142] S6.2 Based on the above structured high-dimensional feature factors, the physical constraint neural network backbone is invoked to perform feature embedding and nonlinear mapping on the weighted fused feature vector to obtain the preliminary compression parameter prediction vector, thereby realizing the end-to-end mapping of high-dimensional feature information to the parameter space.

[0143] S6.3 embeds domain expert physical constraints and process safety boundaries into the model structure for the preliminary compression parameter prediction vector. For example, it sets process physical constraint functions such as material compatibility, interlayer thermal expansion coefficient or pressure threshold, and performs boundary correction and legality screening on the prediction results to ensure the authenticity and feasibility of the output parameters.

[0144] In the model training and inference process, S6.4 adopts a knowledge distillation mechanism, using the physical mechanism model prediction distribution in the domain expert system as a soft label signal. By minimizing the distribution difference between the student network (i.e., the data-driven prediction model) and the teacher network (i.e., the physical mechanism model), physical knowledge is deeply empowered to the data-driven process, improving the interpretability and generalization ability of the compression parameter prediction model.

[0145] S6.5 integrates the prediction results after physical constraint optimization and knowledge distillation, and outputs compression parameter prediction results with high interpretability and wide transferability. These results serve as a reference for subsequent closed-loop process control and real-time parameter optimization. The model uncertainty index is also output along with the parameter results to facilitate dynamic adaptive processing and performance feedback monitoring in the backend system.

[0146] S7: Based on the predicted results of the pressing parameters and the real-time quality feedback information, dynamic monitoring of the prediction accuracy is performed. If a new structural distribution or a prediction deviation exceeding the threshold under abnormal working conditions is detected, model self-evaluation and anomaly adaptive learning are triggered to optimize the feature fusion weights and prediction strategies online, so as to continuously improve the adaptability to complex high-dimensional scenarios. Specifically, this includes:

[0147] S7.1 synchronizes the predicted results of pressing parameters with real-time quality feedback information, and based on the time consistency analysis algorithm, aligns the predicted output tensor with the corresponding quality control feedback tensor in terms of time and marks the process stage to form a highly matched prediction-feedback data pair as the basis for subsequent deviation judgment.

[0148] S7.2 uses a multidimensional error mapping algorithm to dynamically calculate the parameter prediction deviation on a plate-by-plate and structure-by-structure basis based on highly matched prediction-feedback data pairs, and outputs the prediction error tensor for each process stage, providing a refined indicator for online judgment of abnormal deviations.

[0149] S7.3 performs anomaly distribution identification on the prediction error tensor and adopts an adaptive threshold judgment algorithm. It sets a dynamic threshold for prediction deviation based on the historical distribution of plate type and current process parameters, automatically detects anomalies in prediction accuracy caused by new structural distributions or abnormal operating conditions, and extracts anomaly distribution labels as the entry point for model self-evaluation.

[0150] S7.4 triggers a model self-evaluation mechanism, combining anomaly distribution labels and historical model performance indicators. It employs a multi-scale feature importance backtracking algorithm to analyze the sensitivity and failure factors of the current feature fusion weights and process classification boundaries to predicted anomalies, and generates an optimization suggestion tensor for feature fusion weights and structure mapping strategies.

[0151] S7.5 performs online feature fusion weight adaptive adjustment based on the optimization proposal tensor and anomaly distribution labels. It combines incremental deep learning algorithms to optimize parameter prediction strategies (such as loss weighting, feature selection, and subspace activation) in real time, and outputs self-correcting high-dimensional feature fusion weights and dynamically corrected parameter prediction strategy vectors.

[0152] S7.6 integrates the self-corrected high-dimensional feature fusion weights and parameter prediction strategy vectors online, performs closed-loop accuracy tracking for the next round of compression parameter prediction, and continuously updates the model adaptive tuning log to form a dynamically evolving high-dimensional feature fusion compression parameter prediction module.

[0153] S8: The optimized compression parameter prediction results are output as closed-loop control suggestions to the compression process control, driving the production equipment to achieve personalized real-time compression process adjustment, and recording key parameters and model feedback logs to support subsequent performance traceability and continuous model iteration, specifically including:

[0154] S8.1 takes the optimized pressing parameter prediction results as input, and converts them into process control instructions that can be recognized by the production equipment. Through the process parameter conversion module, the key process variables such as temperature, pressure and heating curve in the prediction results are digitally packaged and converted to generate a set of control instructions that are adapted to the multilayer board pressing process control platform in real time, and efficiently map the pressing parameters into process adjustment actions.

[0155] Based on the generated process control instruction set, S8.2 implements parameter distribution and real-time process adjustment to the pressing process control system on the production line through industrial Ethernet or fieldbus communication protocols. It executes actions including but not limited to PID adaptive pressing adjustment, temperature programmable output, and dynamic pressure valve adjustment, so as to realize closed-loop high-precision personalized process control of the pressing equipment and ensure the accurate implementation of predicted parameters in the pressing process.

[0156] During the pressing process control and execution, S8.3 collects real-time process response data from the process control system, including actual temperature, pressure, equipment status codes, and key adjustment feedback parameters. Based on the pressing parameter adjustment log automatic recording module, it performs bidirectional correlation and traceability indexing on the collected equipment response data and process parameter commands to generate highly consistent pressing process key parameters and equipment response logs.

[0157] Based on key parameters of the pressing process and equipment response logs, S8.4 uses a model feedback data collection mechanism to statistically analyze the response effect of the pressing parameter prediction model in actual control scenarios, and calculates core indicators such as the pressing parameter prediction hit rate and process deviation to form a real-time performance evaluation dataset of the pressing prediction model in the production environment.

[0158] S8.5 inputs the real-time performance evaluation dataset of the above-mentioned pressing prediction model into the model continuous optimization module. Combined with the recorded key parameters of the pressing process and the equipment response logs, it dynamically adjusts the high-dimensional feature fusion weights, model parameters, and knowledge distillation strategies to form adaptive optimization configuration suggestions for the model. This provides strong data support for the continuous upgrading of the model and performance tracing in subsequent high-dimensional complex plate scenarios.

[0159] 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.

[0160] 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 multi-layer board pressing parameter intelligent prediction and regulation method, characterized in that, The method comprises the following steps: S1: Collecting multi-source heterogeneous data from different material types, structural layouts and process stages in the pressing production process of the multi-layer board to form an original data set covering the diversity of board types and the difference of multiple working conditions; S2: Performing data synchronization and standardization processing on the original data set to obtain a multi-source heterogeneous data set; S3: Based on the multi-source heterogeneous data set after the standardization processing, constructing a multi-level tensor model, and expressing the structural attributes, process parameters and sensor measurement values of the multi-layer board in a high-order tensor according to the spatial dimension, hierarchical dimension, time dimension and process stage; S4: Applying a multi-scale tensor decomposition algorithm to the multi-level tensor model to decompose the local feature factors and global feature factors in the high-order tensor space, and obtaining multi-scale feature factors, The step S4 specifically comprises: Applying a tensor dimension initialization method to the multi-level tensor model after the standardization processing, setting a decomposition reference dimension for the high-order tensor model according to the spatial dimension, hierarchical dimension, time dimension and process stage of the structural attributes, process parameters and sensor measurement values of the multi-layer board, and obtaining the basic structural distribution of the multi-dimensional feature space; Based on the decomposition reference dimension of the multi-level tensor model, performing multi-level tensor space decomposition on the high-order tensor model to generate a combination of local feature factors and global feature factors containing multi-level features of structural dimension, ontological dimension and time sequence dimension; Performing correlation calculation and feature redundancy elimination on the local feature factors, and performing feature entropy analysis and correlation weight evaluation; Applying an abnormal distribution detection algorithm to the combination of global feature factors, using statistical extreme value detection and feature distribution comparison, and independently marking and tracking abnormal factors in the decomposition result; Feeding the abnormal area mark and the core local feature factor back to the tensor decomposition model to dynamically and adaptively adjust the decomposition parameters and decomposition step length; S5: Based on the multi-scale feature factors, calculating the regional attention mapping of the board structure partition, and applying regional adaptive weight distribution to differentially strengthen and fuse the high-dimensional features of the key regions of material layering, special structure nodes and abnormal value distribution to obtain a weighted fusion feature vector; S6: Inputting the weighted fusion feature vector into a physical constraint neural network model, combining embedded field expert physical constraints and process safety boundaries, and generating a pressing parameter prediction result; S7: Performing dynamic monitoring of the prediction accuracy for the pressing parameter prediction result and real-time quality feedback information, and if it is detected that the prediction deviation exceeds the threshold under the structure distribution or abnormal working condition, triggering model self-evaluation and abnormal adaptive learning, and online optimizing the feature fusion weight and the prediction strategy.

2. The method according to claim 1, wherein, The step S7 further comprises: Step S8: Outputting the optimized pressing parameter prediction result as a closed-loop control suggestion to the pressing process regulation, and recording the key parameters and model feedback log.

3. The intelligent prediction and control method for multilayer board lamination parameters according to claim 1, characterized in that, The step S1 specifically comprises: Setting a sensor collection node for each main process link of the multi-layer board production site to obtain real-time sensor measurement values and obtain process sensor data reflecting the dynamic changes of the pressing process; Collecting physical and chemical performance parameters of various types of multi-layer board materials, including resin content, copper foil thickness, dielectric constant, glass fiber type and material batch basic information, through structured data interface to summarize as input of structure attribute modeling; For historical production batches, obtain the historical pressing process formula parameters and collect the corresponding quality detection results, including warping rate, interlayer peel strength and micro-hole filling rate, to establish a mapping of material-technology-quality historical data; For specific multi-layer board products, extract detailed structure arrangement information and output through CAD structure file or data interface standardization; Integrating original process sensing data, material physical and chemical parameters, historical process formula and quality detection results, and structure arrangement information, and based on the unique identification of the board type batch, the management is labeled to form a multi-source heterogeneous raw data set of multi-layer board pressing production.

4. The method of claim 3, wherein the method further comprises: The original process sensing data includes temperature, pressure, humidity and interlayer thickness.

5. The method of claim 3, wherein the method further comprises: determining the pressing parameters of the multi-layer board according to the pressing parameters of the multi-layer board in the database; and adjusting the pressing parameters of the multi-layer board according to the pressing parameters of the multi-layer board in the database. The historical pressing process formula parameters include pressure curve, holding time, lamination path and cooling rate.

6. The method of claim 1, wherein the method further comprises: The data synchronization and standardization processing of the raw data set includes time alignment, feature dimension normalization and outlier removal.

7. The intelligent prediction and control method for multilayer board lamination parameters according to claim 1, characterized in that, The step S3 specifically includes: Classifying the standardized multi-source heterogeneous data by attributes, mapping the original data into three categories of structure attributes, process parameters and sensing measurement values through template mapping algorithm to construct the basic attribute dimensions of multi-level tensor model; Using structure layout rules to reorganize the spatial level index of structure attribute data, mapping the structure attributes to the spatial and level dimensions of the tensor; Aligning the process parameter data in time and marking the stages according to the process stage time sequence information, and using process segment layering algorithm to encode the participation mode of process parameters in each process stage and the whole process, so as to map the process parameters to the time and process stage dimensions of the tensor; Based on the collection time stamp and spatial collection location information of real-time sensing measurement values, combining dynamic sensing data with structure attributes and process parameters according to the corresponding index dimensions to form a data block with unified index of multi-time and space and multi-level; Taking the multi-level tensor data block as input, using high-order tensor construction function to perform structured operation according to spatial dimension, level dimension, time dimension and process stage dimension to generate standardized multi-level tensor model.

8. The intelligent prediction and control method for multilayer board lamination parameters according to claim 1, characterized in that, The step S5 specifically includes: Performing structure partition mapping processing on the multi-scale feature factors obtained by tensor decomposition, classifying the feature factors into structure partition feature matrix according to the spatial distribution, level division and process node information of the pressed board; Based on the structure partition feature matrix, applying regional attention mechanism algorithm to calculate the regional attention weight of the feature factors of each structure partition, measuring the importance of material layering, special structure node and abnormal distribution area in the overall feature expression; According to the regional attention weight, implementing regional adaptive weight distribution processing on the structure partition feature matrix, dynamically adjusting the multi-scale feature factor weight of material layering, special node and abnormal area to form an adaptive weighted feature expression matrix. The adaptive weighted feature expression matrix is input into a high-dimensional feature fusion algorithm, and sparse strong expression of the structural partition interval and cross-region interaction features is performed through multi-dimensional correlation reinforcement operation; High-dimensional information redundancy removal and feature normalization processing are performed on the feature fusion result subjected to multi-stage adaptive weighting and correlation reinforcement, and an optimal weighted fusion feature vector for a specific plate type and special working condition is generated.

Images