Tracing method for lamination alignment deviation of multi-layer composite circuit board

Abstract

The invention discloses a multilayer composite circuit board lamination alignment deviation traceability method, which comprises the following steps of: acquiring process parameters such as temperature, pressure, tension and alignment images in real time through a distributed multi-mode sensor array, fusing time sequence process data, and extracting interlayer alignment deviation characteristics by combining image processing and a characteristic matching algorithm; modeling a propagation path of each error source based on a causal chain knowledge base, dynamically analyzing an error contribution degree by using a causal inference engine, and outputting an error traceability and sorting result with confidence; according to the type of the dominant error source, a corresponding parameter adjusting instruction is automatically generated and executed, closed-loop control is formed, rapid diagnosis and active compensation of interlayer alignment deviation are achieved, alignment precision and process reliability in the batch manufacturing process are improved, and the method has the capacity of dynamically adapting to working condition changes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of monitoring the deviation of the lamination alignment of a multi-layer composite circuit board, and particularly relates to a method for tracing the deviation of the lamination alignment of a multi-layer composite circuit board. BACKGROUND

[0002] In the field of manufacturing multi-layer composite circuit boards, the detection and deviation tracing of interlayer alignment accuracy has become a core link to guarantee the quality of high-density and high-complexity PCB products. With the rapid development of 5G communication, automotive electronics and high-end servers, the manufacturing of multi-layer composite structure PCBs has higher requirements for the accurate alignment of each layer during the lamination process. Therefore, multi-modal sensors (such as temperature, pressure, tension, and visual images) are widely used in the industry for process monitoring, and image comparison or process parameter analysis is used to detect, count and correct the alignment deviation.

[0003] However, the existing technology still has significant limitations in identifying the source of deviation and analyzing complex error links:

[0004] (1) For the alignment deviation in the lamination process, only shallow and direct factors can be traced, and it is difficult to restore the complete causal chain of "parameter disturbance-physical property change-macroscopic deviation performance". In the case of complex deviation or multi-source collaborative failure (such as the simultaneous action of temperature and tension, material difference and uneven pressure), traditional technology is difficult to distinguish the deep contribution and correlation path of each potential error source;

[0005] (2) Most existing systems use empirical formulas or single-step statistical discrimination, lacking causal relationship modeling and dynamic deduction mechanisms. For example, the interlayer deformation caused by thermal expansion, material shrinkage and pressure coupling is difficult to effectively distinguish and diagnose through single feature detection or static classification methods;

[0006] (3) In terms of error compensation decision-making, it is usually based on historical deviation mean to adjust the heating rate or pressure curve, which is not adaptive to dynamic fluctuations of manufacturing environment and batch process parameters, and is difficult to achieve self-evolution and precise closed-loop control. SUMMARY

[0007] To solve the above technical problems, the present application provides a method for tracing the deviation of the lamination alignment of a multi-layer composite circuit board. The technical solution of the present application is as follows: a method for tracing the deviation of the lamination alignment of a multi-layer composite circuit board, comprising:

[0008] S1: During the lamination process of a multi-layer composite circuit board, a multi-modal sensor array is used to collect process parameters including temperature, pressure, tension and interlayer alignment images, and a multi-dimensional process data set containing time series is generated;

[0009] S2: image comparison processing is performed on the collected interlayer alignment image, an interlayer alignment deviation atlas is extracted, key alignment feature points in the deviation atlas are vectorized and coded, and structured deviation feature data is generated;

[0010] S3: based on a preset causal chain knowledge base, an alignment deviation propagation path model is constructed, the model includes a causal mapping relationship between potential error sources such as thermal expansion difference, material shrinkage rate, and uneven pressure distribution and deviation atlas features;

[0011] S4: input the structured deviation feature data and the multi-dimensional process data set into the causal reasoning engine, dynamically update the probability weight of each potential error source, and output an error source list with confidence sorting;

[0012] S5: according to the dominant error source type in the error source list, a corresponding compensation control strategy is selected, and a control instruction set for adjusting the pressing parameters is generated, the pressing parameters including the heating rate, the pressure loading curve and the holding time;

[0013] S6: based on the control instruction set, the parameter adjustment operation is performed on the pressing equipment to form a closed loop control circuit, so as to dynamically compensate the current alignment deviation and suppress the deviation accumulation in the subsequent pressing process;

[0014] S7: after the pressing is completed, the actual interlayer alignment accuracy data is collected, compared with the predicted deviation value, a deviation compensation effect evaluation report is generated, and used to update the error source weight model in the causal chain knowledge base;

[0015] S8: in the continuous multi-batch pressing process, the distribution change trend of the multi-dimensional process data is dynamically monitored, and when a working condition mutation is detected, a causal chain model version switching mechanism is triggered to adapt to the error propagation characteristics in the current manufacturing environment.

[0016] The multi-layer composite circuit board pressing alignment deviation tracing method provided by the application has the following beneficial effects:

[0017] (1) Compared with the traditional scheme based on only single process parameter or simple pattern matching, the present application synchronously fuses the pressing real-time sensing data (multi-point temperature, pressure, tension and high-resolution alignment image) in time, and describes the multi-level progressive relationship between error sources and deviation features through a complete causal chain network model. When analyzing the error sources, the accuracy of dominant error source identification is significantly improved compared with the existing threshold method, the detection rate of deep process abnormalities (such as uneven thermal expansion, material layering shrinkage rate difference, etc.) is greatly improved, and the advantages are prominent especially in the deviation scene caused by complex or composite factors;

[0018] (2) The application not only can output an error source list with confidence ranking, but also can automatically match the optimal compensation strategy based on the causal chain reasoning result, and generate device executable compression process parameter adjustment instructions (such as heating rate, pressure time curve, pressure holding window, etc.), to realize online dynamic regulation and control of the compression process;

[0019] (3) For batch fluctuations or process mutations in the production environment, the application introduces multi-batch distribution monitoring and causal model version switching mechanism, automatically identifies process drift and switches the most suitable causal chain model, ensures that the error traceability effect is not affected by the change of production conditions, and significantly outperforms the generality and robustness of the existing static algorithm;

[0020] (4) The application does not depend on physical targets, special marks or special chemical formulas, etc. Physical structure restrictions, and is mainly supported by multi-modal data acquisition, reasoning and closed-loop decision algorithm integrated by software and hardware, so it can be flexibly deployed in multi-layer composite circuit board manufacturing scenarios of different materials, layers and compression processes (especially suitable for high-density interconnection, rigid-flex combination and other high-end PCB production lines). The applicability is much better than existing solutions based on specific structures or materials. BRIEF DESCRIPTION OF DRAWINGS

[0021] Fig. 1 A flowchart of a multi-layer composite circuit board compression alignment deviation traceability method of the application;

[0022] Fig. 2 A sub-flowchart of a multi-layer composite circuit board compression alignment deviation traceability method of the application;

[0023] Fig. 3 Another sub-flowchart of a multi-layer composite circuit board compression alignment deviation traceability method of the application. DETAILED DESCRIPTION

[0024] In order to make the purpose and advantages of the application more clear and obvious, the application will be further described below in combination with examples; it should be understood that the specific examples described herein are only used to explain the application, and do not limit the application.

[0025] The preferred implementation methods of the application will be described below with reference to the drawings. Those skilled in the art should understand that these implementation methods are only used to explain the technical principles of the application, and do not limit the protection scope of the application.

[0026] As used herein, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms "comprises", "comprising", "includes" and / or "including", or the like, when used in this specification, specify the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

[0027] Referring to Figs. 1-3 As shown in the drawings, a multi-layer composite circuit board pressing alignment deviation traceability method comprises:

[0028] S1: During the pressing process of the multi-layer composite circuit board, process parameters including temperature, pressure, tension and interlayer alignment image are collected by a multi-modal sensor array to generate a multi-dimensional process data set containing time series;

[0029] S2: The collected interlayer alignment image is subjected to image comparison processing, an interlayer alignment deviation atlas is extracted, and key alignment feature points in the deviation atlas are vectorized and coded to generate structured deviation feature data;

[0030] S3: Based on a preset causal chain knowledge base, an alignment deviation propagation path model is constructed, which contains the causal mapping relationship between potential error sources such as thermal expansion difference, material shrinkage rate, uneven pressure distribution and deviation atlas features;

[0031] S4: The structured deviation feature data and the multi-dimensional process data set are input into a causal reasoning engine, the probability weight of each potential error source is dynamically updated, and an error source list with confidence ranking is output;

[0032] S5: According to the dominant error source type in the error source list, a corresponding compensation control strategy is selected, and a control instruction set for adjusting the pressing parameters is generated, the pressing parameters including the heating rate, the pressure loading curve and the holding time;

[0033] S6: Based on the control instruction set, parameter adjustment operations are performed on the pressing equipment to form a closed loop control circuit to dynamically compensate for the current alignment deviation and suppress the deviation accumulation in the subsequent pressing process;

[0034] S7: After the pressing is completed, actual interlayer alignment accuracy data is collected, compared with the predicted deviation value, a deviation compensation effect evaluation report is generated, and used to update the error source weight model in the causal chain knowledge base;

[0035] S8: In the process of continuous multi-batch pressing, the distribution trend of multi-dimensional process data is dynamically monitored, and the causal chain model version switching mechanism is triggered when the working condition mutation is detected to adapt to the error propagation characteristics in the current manufacturing environment.

[0036] The step S1: In the process of pressing multi-layer composite circuit board, the process parameters including temperature, pressure, tension and interlayer alignment image are collected by multi-modal sensor array to generate multi-dimensional process data set containing time series. Specifically, it includes:

[0037] S1.1: In the process of pressing multi-layer composite circuit board, based on the distributed sensor network architecture in the field of industrial control, temperature sensors, pressure sensors, tension sensors and high-resolution visual cameras are deployed at key nodes of pressing equipment to synchronously collect multi-modal process parameters in the pressing process;

[0038] In the process of pressing multi-layer composite circuit board, for the requirement of collecting multi-modal process parameters such as temperature, pressure, tension and interlayer alignment image, the key monitoring nodes of pressing equipment are determined as the deployment position of sensor components;

[0039] Using distributed sensor network architecture (parameters: node type = temperature node, pressure node, tension node, image node; network topology = star + ring hybrid), the physical connection and data synchronization identification of various sensors are realized;

[0040] Further, through the signal link equalization method (parameters: sampling rate = 1kHz, signal amplitude tolerance = ±5%), stable communication of multiple sensors in high-noise industrial environment is realized, and the time synchronization accuracy reaches 1 / 1000 second level;

[0041] Further, through the clock synchronization protocol PTP (Precision Time Protocol) to distribute unified time reference in the sensor network, the timestamp consistency of all data acquisition units including temperature, pressure, tension sensors and visual cameras is realized;

[0042] Further, the installation technology combining mechanical fixed support and vibration isolation structure (parameters: vibration isolation frequency below 2Hz, installation tolerance ±0.1mm) is adopted to ensure the relative positioning stability of each sensor under the conditions of pressing mechanical vibration and thermal expansion and contraction;

[0043] Through the combination of multi-channel data acquisition card (parameters: resolution = 16 bits, bandwidth = 100kHz) and edge computing module, the synchronous data stream of various sensors is combined into raw data packet with timestamp to realize the input preparation of subsequent temperature compensation, pressure filtering, tension normalization and image preprocessing;

[0044] Exemplary, in a certain HDI multilayer board pressing task, 8 temperature sensors (model: K-type thermocouple, accuracy ±0.5℃) are selected to be arranged at the center and four corner positions of the upper and lower pressing plates, 6 pressure sensors (model: strain pressure sheet, full scale 10MPa, accuracy ±0.25500N, resolution 0.1N) are arranged at the corresponding positions, and 2 high-resolution cameras (resolution 4096×2160 pixels, frame rate 60fps) are installed at the parallel positions on both sides of the equipment. The sensor nodes are connected through industrial Ethernet and gigabit optical fiber to form a dual-ring network topology, each node is equipped with a PTP hardware timestamp module to achieve cross-node data time synchronization error less than 500ns. After installation and trial operation verification, under full load pressing conditions, the collected multi-modal process raw data has no packet loss and no error packet within 1 hour of operation period, the time calibration consistency of temperature, pressure, tension and image data is better than 0.001s, fully meeting the accuracy requirements of subsequent data fusion and causal chain analysis;

[0045] S1.2: Perform temperature compensation algorithm and filtering processing on the original temperature data collected by the temperature sensor to eliminate the influence of environmental temperature drift and sensor noise, and obtain the real-time thermal field distribution characteristics of each layer material in the pressing process;

[0046] For the original temperature signal collected by the temperature sensor, a temperature compensation algorithm based on reference node (parameters: reference node temperature accuracy ±0.1℃, compensation coefficient update period 1s) is used to realize real-time correction of sensor zero drift and environmental temperature drift;

[0047] Further, Kalman filtering algorithm (parameters: state transition matrix is constructed according to heat conduction model, process noise variance 0.0025, observation noise variance 0.0004) is used to realize dynamic smoothing processing of the compensated temperature data, and temperature continuous sequence with enhanced time consistency is obtained;

[0048] Further, multi-scale wavelet denoising processing (wavelet basis=db4, decomposition level=3) is performed to effectively suppress high-frequency random noise and generate temperature feature components at different scales;

[0049] Further, according to the thickness layering parameters of the pressing plate and the thermal conduction finite element model, the spatial interpolation thermal field distribution of each measuring point is calculated, and the interpolation function of the thermal field is defined as: T(x,y,z,t)=∑wiT(xi,yi,zi,ti)w(xi-x,yi-y,zi-z) i ·T i , where w i is the spatial weight coefficient, and T i is the filtered temperature value of the i th sensor;

[0050] Further, the time series thermal field matrix is input into the feature extraction module to extract the maximum temperature gradient temperature rise rate and temperature uniformity index The multidimensional thermal field feature vectors included;

[0051] Through the above temperature compensation and filtering link processing method, the original temperature acquisition results of the previous step are transformed into real-time thermal field distribution data with high spatiotemporal accuracy, so as to provide high confidence input for subsequent causal chain modeling.

[0052] For example, during the HDI multilayer board lamination process, the sampling frequency of the eight K-type thermocouples is set to 10Hz. The offset compensation is calculated using the reference node temperature (located within the insulation layer of the equipment casing, where the ambient temperature is stable), and the resulting drift correction value ranges from ±0.25℃. The initial state covariance matrix in the Kalman filter... The updated output temperature curve retains its trend under low-frequency fluctuations and suppresses high-frequency spikes by more than 90%; after wavelet denoising, the SNR is improved from 35dB to 52dB; the thermal field interpolation calculation uses a finite element mesh divided into 5mm×5mm×layer thickness units, with a relative error of less than 2%; the temperature rise rate extraction shows that the maximum gradient occurs in the center region of the plate 50mm away from the heat source during the heating stage, and the uniformity index U is 0.93, which meets the input conditions required for process control;

[0053] S1.3: Perform sliding window mean filtering and outlier removal on the raw pressure signal collected by the pressure sensor to extract the pressure loading curve of each pressure plate area during the pressing process and generate a pressure distribution matrix with time series characteristics;

[0054] S1.4: Normalize the tension data collected by the tension sensor and map it to the material elastic modulus to obtain the real-time stress state parameters of each substrate layer during the pressing process, which serve as the input basis for material deformation modeling.

[0055] S1.5: Perform image preprocessing operations on the interlayer alignment images acquired by the vision camera, including grayscale conversion, contrast enhancement and edge detection, to extract the geometric contour features of the interlayer positioning holes and conductive lines, and generate a basic image dataset for image comparison.

[0056] S1.6: Perform multi-source data fusion processing on the temperature, pressure, tension and image data output by multimodal sensors based on timestamps to construct a multidimensional process dataset containing time series characteristics, providing structured input for subsequent deviation map extraction and causal chain modeling.

[0057] Step S2: Perform image comparison processing on the acquired inter-layer alignment images, extract the inter-layer alignment deviation map, and vectorize and encode the key alignment feature points in the deviation map to generate structured deviation feature data. Specifically, this includes:

[0058] S2.1: The collected interlayer alignment image is subjected to grayscale and histogram equalization processing to enhance image contrast and suppress noise interference, and standardized grayscale image data is obtained;

[0059] For the collected interlayer alignment image, a grayscale method based on brightness component extraction is used (parameters: color space conversion mode = RGB→Grayscale, weight coefficient = 0.299, 0.587, 0.114) to remove irrelevant color information in the original color image and retain brightness information reflecting the geometric profile;

[0060] Further, a frequency distribution histogram H(i) of the grayscale image is generated by a histogram calculation module (parameter: number of grayscale levels = 256), where i is the grayscale value index and H is the number of pixels of the grayscale value;

[0061] Further, a median filter algorithm (parameters: window size = 5x5 pixels, boundary processing mode = mirror padding) is used to suppress high-frequency noise while avoiding edge blurring, thereby ensuring the accuracy of subsequent feature extraction;

[0062] Further, a self-adaptive histogram equalization (CLAHE) method (parameters: local window size = 8x8 pixels, limited contrast threshold = 2.0) is used to optimize the local contrast of unevenly illuminated areas during the pressing process, and a globally and locally equalized standardized grayscale image is obtained;

[0063] Through the above chain processing method, the original color interlayer alignment image is converted into a standardized grayscale image data with global contrast equalization, local detail enhancement, and noise suppression, realizing a stable input condition for subsequent SIFT feature point detection;

[0064] For example, during the pressing process of an HDI multi-layer board, a camera captures a color image with a resolution of 4096x2160 pixels, the RGB-to-grayscale weight is set to 0.299, 0.587, and 0.114, and the grayscale histogram shows that the main grayscale is concentrated in the 100-160 interval. After equalization formula processing, the pixel grayscale range is distributed in the 10-240 interval, and the image contrast is improved by about 32%. After median filtering, the proportion of noisy pixels is reduced from 1.2% to 0.03% after removing the salt and pepper noise introduced during the pressing process. After CLAHE processing, the line boundary clarity of the local dark area is improved by 28%, and the signal-to-noise ratio of the standardized grayscale image is improved from 27dB to 41dB, providing excellent input data conditions for subsequent feature extraction output stability. The processing is completed under the condition that the algorithm operation complexity does not exceed 15ms / frame, meeting the real-time requirements of the pressing process;

[0065] S2.2: Key point detection and descriptor extraction are performed on the normalized grayscale image data based on the Scale-Invariant Feature Transform (SIFT) algorithm to obtain a set of image feature points with scale and rotation invariance, and a set of local feature description vectors are generated;

[0066] S2.3: Feature point matching is performed on the local feature description vector set of the adjacent layer images using the nearest neighbor matching strategy to identify the inter-layer correspondence relationship, and a set of preliminary alignment deviation vectors are generated;

[0067] S2.4: Based on the Random Sample Consensus (RANSAC) algorithm, outlier elimination and transformation matrix estimation are performed on the preliminary alignment deviation vector set to obtain the inter-layer rigid transformation parameters, and an optimized alignment deviation atlas is generated;

[0068] S2.5: Vectorization encoding operation is performed on the key alignment feature points in the optimized alignment deviation atlas, multi-dimensional feature tuples are constructed based on geometric coordinates and deviation direction information, and structured deviation feature data are generated for input to the subsequent causal reasoning module.

[0069] The step S3: Based on the preset causal chain knowledge base, an alignment deviation propagation path model is constructed, which includes the causal mapping relationship between potential error sources such as thermal expansion difference, material shrinkage rate, pressure distribution unevenness and deviation atlas features. As shown in Fig. 2 The specific steps include:

[0070] S3.1: The error source types in the preset causal chain knowledge base are classified and modeled, including thermal expansion difference, material shrinkage rate deviation, pressure distribution unevenness, inter-layer tension imbalance and convection solidification rate difference. Based on process experience and historical data, a propagation path atlas of error sources is constructed to form a structured causal chain network model;

[0071] S3.2: The deviation atlas features output by the image comparison module are semantically analyzed, and the spatial distribution features, offset direction consistency index and local deformation coefficient of the key alignment offset points are extracted to generate a structured deviation feature vector set as the input feature space of the causal chain model;

[0072] For the optimized alignment deviation atlas, a feature analysis method based on spatial geometry clustering is used (parameters: clustering algorithm type = DBSCAN, spatial radius ∈ = 2.5 pixels, minimum sample number = 5) to realize spatial region aggregation of key alignment offset points and eliminate isolated irrelevant points;

[0073] Further, a two-dimensional centroid calculation method (parameters: coordinate system reference = image center, weight factor = 1.0) is used to realize spatial distribution feature extraction of each clustering region and obtain the center coordinates and variance index of each distribution unit;

[0074] Further, the vector angle consistency calculation method (parameters: reference direction vector = (1, 0), unit system = radian) is adopted, and the offset direction consistency index of each clustering area is calculated according to the formula cos (θ i -θ ref ), wherein N is the number of offset vectors in the area, θ i is the direction angle of the i-th vector, and θ ref is the reference direction angle;

[0075] Further, the affine transformation residual fitting method (parameters: number of iterations of least squares method = 100, convergence threshold = 1e-6) is adopted to fit the deviation point set to an ideal non-deformation model, calculate the local geometric deformation coefficient K loc , and obtain the relative deformation ratio through the formula , wherein ΔL is the length change amount of the fitted model, and L ref is the reference length;

[0076] Through the joint mode of spatial clustering, offset direction consistency calculation and local deformation coefficient solving, the deviation atlas is converted into a structured deviation feature vector set containing spatial distribution, direction consistency and geometric deformation characteristics, realizing the construction of high-dimensional precise input features required by the causal chain model;

[0077] For example, in a multi-layer composite circuit board pressing process, the optimized alignment deviation atlas contains 132 offset feature points, which are aggregated into 7 distribution areas through DBSCAN clustering (∈ = 2.5 pixels, minimum sample number = 5). After the centroid coordinates and variance of each area are calculated, the offset of area 1 relative to the image center coordinates (2048, 1080) is (12.3, -5.7) pixels, and the variance is (3.1, 2.8) pixels2. The direction consistency index C1 of area 1 is calculated as 0.92, indicating that the offset vectors in this area are highly consistent. The local deformation coefficient K loc is calculated, the reference length L ref = 50.0 pixels, ΔL = 0.85 pixels is measured, and K loc = 0.017, indicating that the local stress deformation ratio is 1.7%. The final output structured deviation feature vector contains 21 feature elements, which are directly input for dynamic material response feature fusion in the subsequent S3.3 step, realizing the technical effect of accurately representing the causes of deviation in error causal chain analysis;

[0078] S3.3: Based on the material attribute parameters and pressing process parameters of the multi-layer composite circuit board, the thermal expansion coefficient matrix and solidification shrinkage prediction value of each layer material in the pressing process are calculated, and dynamic material response features are generated to quantify the theoretical contribution of each error source in the thermal-mechanical coupling environment;

[0079] Based on the preset material property database and the real-time collected lamination process parameters, the material thermal physical parameter analysis method (parameters: input material type = FR4, polyimide, thermal conductivity unit = W·m -1 ·K-1, density unit = kg·m -3 ) is used to realize the extraction of the heat capacity C, density p and initial linear expansion coefficient a0 of each layer of substrate;

[0080] Further, through the real-time thermal field calculation module (algorithm type = finite element transient heat conduction equation, time step = 0.1s, boundary condition = constant edge temperature), combined with the lamination time sequence temperature curve T(t), the transient temperature distribution of the multi-layer material under non-uniform heating conditions is solved, and the local temperature increment AT(t) of each element at each time is obtained.

[0081] Further, the thermal expansion coefficient dynamic correction algorithm (parameters: temperature threshold number = 3 segments, correction model = piecewise linear fitting) is used to calculate the effective thermal expansion coefficient a eff (t) of each time according to the formula a eff (t) = a0 + k x AT(t), where k is the temperature-sensitive correction coefficient and AT is the transient temperature rise value.

[0082] Further, combined with the pressure distribution matrix P(x, y, t) measured by the pressure sensor, the material viscoelastic curing model (model type = Prony series fitting, order = N = 5) is used to calculate the viscoelastic deformation curve of each layer of material in the curing stage, and the predicted value of the curing shrinkage rate S(t) is calculated according to the volume strain formula ;

[0083] Further, the obtained time sequence effective thermal expansion coefficient matrix a eff (t) and the curing shrinkage rate prediction sequence S(t) are constructed into a dynamic material response feature vector group according to the layer number index, and the same scale of different physical quantities is realized through the working condition feature normalization processing (parameters: normalization interval = [0, 1]);

[0084] Through the thermal-mechanical coupling multi-parameter fusion algorithm (method: weighted superposition method, weight calculation based on error source sensitivity analysis), the multi-dimensional feature results of the previous step are converted into dynamic material response indexes quantifying the theoretical contribution of each error source, and the accurate quantification of the potential thermal expansion difference and the influence intensity of the material shrinkage rate is realized.

[0085] For example, in a lamination process of a multi-layer composite circuit board using FR4 and polyimide mixed lamination, the initial linear expansion coefficient a0 of FR4 in the material database is 1.45 x 10 -5 K-1, and the initial linear expansion coefficient of polyimide is 2.0 x 10 -5K-1. The peak of the pressing process temperature curve is 185℃, the average heating rate is 2.5K / min, and the ΔT of the FR4 core layer at 300s is 42K through finite element transient thermal field calculation. Substitute into the correction formula and take k1.2×10 -7 K-2, get the moment α eff = 1.954×10 -5 K-1. Pressure distribution matrix analysis shows that the average loading pressure of the edge area is 0.85 times that of the center area, resulting in a solidification shrinkage rate prediction of 0.32% in the edge area and 0.27% in the center area. After normalization and weight fusion processing, the deviation contribution index of the thermal expansion difference is 0.64, and the material shrinkage rate difference is 0.28. This result will be directly input into the causal chain network model in step S3.4 to activate the relevant error propagation path. This process realizes the dynamic quantification of material thermal-mechanical response under real-time process monitoring conditions, effectively supporting the root cause analysis of pressing deviation;

[0086] S3.4: Input the structured deviation feature vector set and the dynamic material response feature into the causal chain network model, activate the error propagation path with the highest matching degree with the current deviation atlas through the forward propagation mechanism, and normalize and weight the activation probability of each node in the path to generate a preliminary error source activation atlas;

[0087] Input the structured deviation feature vector set output in step S3.2 and the dynamic material response feature output in step S3.3 into the input layer of the causal chain network model to realize accurate coupling mapping of feature dimensions;

[0088] The causal chain forward propagation algorithm based on node weight initialization (parameters: network type = directed acyclic graph DAG, node initial weight interval = [0.1, 0.9], activation function type = Sigmoid) is used to realize the associated activation of each input feature to the potential error source nodes in the causal chain network, and the initial activation value matrix of multiple paths is obtained;

[0089] Further, through the path matching degree calculation method (parameters: similarity measure = cosine similarity, threshold = 0.85), the activation mode of the feature input is compared with the error propagation path feature vector preset in the knowledge base one by one, and the matching degree index of each path is calculated where v i is the input activation mode vector element, and w i is the path feature vector element;

[0090] Further, filter a number of paths with the highest matching degree (number parameter K = 3) as the candidate error propagation path subset, and calculate the node activation probability of each path;

[0091] Further, the node normalization weighting algorithm (parameter: weighting model = Softmax normalization) is adopted to calculate the normalized activation probability of node n according to the formula where a n is the original activation value of node n, and the denominator is summed over all nodes in the path, ensuring that the path probability sums to 1;

[0092] The weighted node activation intensity matrix is obtained by multiplying the normalized probability by the overall path matching degree;

[0093] Through the above cosine similarity matching, path screening and Softmax normalization chain processing, the input feature driven activation pattern is converted into a preliminary error source activation atlas representing the activity level of different error sources, realizing the quantitative identification of potential bias propagation chains.

[0094] For example, in a FR4 / polyimide hybrid lamination batch of a PCB, the structured bias feature vector set contains 21 feature elements, and the dynamic material response feature vector group contains 12 physical quantity indicators. When the causal chain network is initialized, the node initial weight is set in the range of [0.2, 0.8], and the Sigmoid function is defined as The input feature activation vector is calculated for cosine similarity with the 5 typical propagation paths in the knowledge base, and the matching degree of path 3 is 0.91, which is the highest and exceeds the set threshold value 0.85, entering the candidate path set. This path contains a thermal expansion difference node, an interlayer tension imbalance node and a local deformation response node. After Softmax normalization, the activation probabilities of the nodes are 0.52, 0.31 and 0.17 respectively. After weighting combined with the overall path matching degree, the activation intensity of the thermal expansion difference node is 0.47, indicating that this error source has a significant contribution to the current lamination bias. The preliminary activation atlas output in this calculation will be updated in the subsequent S3.5 step, supporting the formation of the causal chain explanation model;

[0095] S3.5: Based on the Bayesian inference algorithm, the posterior probability of the preliminary error source activation atlas is updated, the temperature gradient, pressure distribution curve and tension feedback value in the multi-dimensional process data are fused, and the error source causal chain path with confidence weight is generated to form the causal chain explanation model of the current alignment bias.

[0096] The step S4: input the structured bias feature data and the multi-dimensional process data set into the causal reasoning engine, dynamically update the probability weight of each potential error source, and output the error source list with confidence ranking. As Fig. 3 shown, specifically including:

[0097] S4.1: Perform data alignment processing on the structured bias feature data and the multi-dimensional process data set, align the time series data of image features and process parameters based on timestamps, and obtain a synchronized multi-modal input data set;

[0098] For the synchronization needs of structured bias feature data and multi-dimensional process data set, a timestamp index matching method is used (parameters: timestamp accuracy = millisecond level, matching tolerance interval = ± 5ms), to realize the initial pairing of image feature time labels and process parameter time labels;

[0099] Further, through linear interpolation method (parameters: interpolation order = 1, boundary processing mode = extrapolation), the process parameters of missing time points are numerically completed, and a continuous time series process data vector set is generated to eliminate the time misalignment caused by inconsistent sampling frequency;

[0100] Further, a sliding window resampling algorithm (window length = 100ms, step = 50ms) is used to synchronize the paired time series data, ensuring that each modal feature is aligned in the same time window;

[0101] Further, based on Z-score normalization processing (parameters: mean and variance are calculated based on the past 10 seconds of data), each modal feature is dimensionless, forming a dimensionless synchronization feature matrix with a mean of 0 and a variance of 1;

[0102] Further, according to the modal priority ordering rule (priority: visual features > temperature features > pressure features > tension features), the normalized feature matrix is column rearranged, and the synchronized multi-modal input data set that meets the input requirements of the causal reasoning engine is output;

[0103] Through multi-stage time registration and feature normalization processing, the structured bias feature data and multi-dimensional process data set of the previous step are converted into multi-modal fusion input with consistent time series and uniform scale, realizing high-precision modeling of feature correlation in the causal reasoning stage;

[0104] For example, in a FR4 / polyimide hybrid lamination batch, the image feature timestamp sequence is {1000, 1050, 1100, 1150} milliseconds, and the multi-dimensional process data sampling timestamp sequence is {1002, 1048, 1105, 1149} milliseconds. Set the matching tolerance interval to ± 5ms, and the matching result is four complete pairing items. The pressure feature is missing at 1050ms and 1150ms, and the linear interpolation formula The completion value is calculated, where t1=1048 milliseconds and t2=1105 milliseconds. The sliding window resampling obtains a sampling sequence with a uniform step of 50 milliseconds, and Z-score normalization makes the mean of each feature column of temperature, pressure, tension and image 0 and the variance 1. The final output of the synchronized multi-modal input data set is a 4x33 matrix, which is input to the causal inference engine after column rearrangement, and the subsequent modeling accuracy is improved by about 12.5%, proving that the data alignment processing improves the stability and accuracy of error source identification;

[0105] S4.2: Based on the pre-defined error propagation path model in the causal chain knowledge base, a multi-source error source candidate set is constructed, including thermal expansion difference, material shrinkage, pressure distribution unevenness, tension fluctuation and interlayer adhesion performance change;

[0106] Based on the synchronized multi-modal input data set, the causal chain knowledge base retrieval method is used (parameters: retrieval mode=full path matching, node similarity threshold=0.85) to obtain the historical causal path set matched with the current pressing batch characteristics;

[0107] Further, through the error source classification mapping algorithm (parameters: class label set={thermal expansion difference, material shrinkage, pressure distribution unevenness, tension fluctuation, interlayer adhesion performance change}), the error types corresponding to the terminal nodes in the retrieved path set are aggregated and mapped, and a preliminary error source category index table is generated;

[0108] Further, the path node weight fusion method (parameters: fusion model=weighted summation method, weight source=node sensitivity coefficient) is used to calculate the comprehensive weight value W of each error source category in all matching paths c , and the calculation formula is:

[0109]

[0110] where w i is the weight value of the i-th node in the path, σ i is the node sensitivity coefficient, and N(c) is the number of nodes involved in category c;

[0111] Further, the category confidence evaluation method (parameters: normalization range=[0, 1], normalization method=Min-Max normalization) is used to map the comprehensive weight value to the initial category confidence P c , which provides prior input for subsequent probability updating;

[0112] Through the above full-path matching retrieval, category mapping, weight fusion and confidence evaluation processing mode, the synchronized multi-modal input data of the previous step is converted into a multi-source error source candidate set covering all possible error categories, realizing the comprehensive coverage of the candidate error sources by the causal reasoning engine in the modeling stage.

[0113] Exemplarily, in a batch of FR4 / polyimide mixed lamination, the synchronized multi-modal input data set contains an image feature matrix of 4x21 dimensions, temperature features of 4x6 dimensions, pressure features of 4x4 dimensions, and tension features of 4x2 dimensions. Five typical paths are predefined in the causal chain knowledge base, and the node similarity threshold is set to 0.85. The search results hit path 3, path 4 and path 5. The terminal nodes of path 3 correspond to thermal expansion difference, path 4 corresponds to material shrinkage, and path 5 corresponds to uneven pressure distribution. After aggregation and mapping, the error source category index table is generated: {thermal expansion difference: path 3; material shrinkage: path 4; uneven pressure distribution: path 5}. When calculating the comprehensive weight of the node, it is assumed that path 3 contains 3 nodes, and their weight values are 0.54, 0.48 and 0.37 respectively, and the sensitivity coefficients are 1.0, 0.9 and 0.85 respectively. Substituting into the formula the comprehensive weight value of thermal expansion difference is 0.475. After Min-Max normalization of the comprehensive weights of all categories, the initial confidences of thermal expansion difference, material shrinkage and uneven pressure distribution are 0.84, 0.65 and 0.56 respectively. The candidate set is input to the subsequent causal graph modeling link of S4.3, effectively improving the comprehensiveness and prior accuracy of error source identification;

[0114] S4.3: using a causal reasoning engine to perform causal graph modeling on the synchronized multi-modal input data set, calculating the conditional probability distribution between each error source and the bias feature, and generating an error source-feature association matrix;

[0115] After obtaining the multi-source error source candidate set and its initial category confidence, the candidate set is used as the modeling input to construct a causal graph topology with error sources as the dependent variable and structured bias features and multi-dimensional process features as the independent variables. A causal graph generation method based on structural equation model is used (parameters: graph type = directed acyclic graph DAG, maximum in-degree = 3, regularization coefficient = 0.01) to realize the establishment of initial directed edges between nodes and the quantitative coding of causal relationship assumptions;

[0116] Further, through the conditional independence test algorithm (parameters: test method = G 2 likelihood ratio test, significance level = 0.05), redundant edges between synchronized multi-modal features and candidate error source nodes are filtered out, causal associations supported by statistical significance are retained, and a preliminary sparse causal graph structure matrix is obtained;

[0117] Further, the parameters of conditional probability distribution are iteratively optimized by using the maximum likelihood estimation method (EM iteration number = 500, convergence threshold = 1e-5) to ensure the stability and generalization of the estimated value under limited samples;

[0118] Further, by constructing the error source-feature association matrix A (the matrix row represents the error source node, and the matrix column represents the input feature), the normalized mutual information indicator is used The matrix element value is filled, I(X;Y) is the mutual information, H is the Shannon entropy, the correlation strength is measured and standardized;

[0119] Through the causal graph modeling, conditional probability fitting and mutual information normalization processing, the statistical dependence relationship between the candidate error source and the multi-modal deviation feature is converted into a quantifiable association weight matrix, which realizes the input enhancement and explicit representation of the association relationship in the subsequent Bayesian update stage.

[0120] For example, in a FR4 / polyimide mixed lamination batch, the candidate error source categories include three nodes of thermal expansion difference, material shrinkage rate, and uneven pressure distribution, and the input feature set includes 21 deviation geometric features and 12 thermal-mechanical response features. When constructing the DAG, the maximum in-degree is set to 3, the regularization coefficient is 0.01, and the G 2 It is verified that 4 features with no significant relationship with the uneven pressure distribution node are removed at a significance level of 0.05. In the conditional probability fitting stage, the continuous features adopt Gaussian distribution, and the estimated conditional mean and standard deviation of the thermal expansion difference node and the feature "local deformation coefficient" are 0.017 and 0.004, respectively; the binary features adopt Bernoulli distribution, and the estimated probability of the material shrinkage rate node and the "high consistency of offset direction" feature is 0.84. After the normalized mutual information is calculated, the NMI value of the thermal expansion difference-local deformation coefficient is 0.76, the NMI value of the material shrinkage rate-offset direction consistency is 0.65, and the NMI value of the uneven pressure distribution-space distribution variance is 0.58. The error source-feature association matrix is output for the Bayesian update stage of S4.4, which effectively improves the convergence speed and stability of the probability update, and increases the matching degree of the final error source confidence evaluation and the actual deviation root cause by about 11.2%;

[0121] S4.4: Apply the Bayesian update algorithm to the error source-feature association matrix, and dynamically adjust the prior probability distribution of each error source combined with the historical process parameter variation trend of the current batch lamination process;

[0122] S4.5: Perform error source sorting based on the updated posterior probability distribution, output an error source list with confidence weight, which includes error source type, confidence score and corresponding deviation contribution estimate value.

[0123] The step S5: according to the dominant error source type in the error source list, selecting the corresponding compensation control strategy, and generating the control instruction set for adjusting the pressing parameters, including the heating rate, the pressure loading curve and the holding time. Specifically, it includes:

[0124] S5.1: Based on the confidence ranking of each error source in the error source list, identify the dominant error source type, which includes one or more combinations of thermal expansion difference, material shrinkage inconsistency, and uneven pressure distribution, to obtain the dominant error source classification label;

[0125] S5.2: According to the dominant error source classification label, retrieve the corresponding compensation control strategy template from the preset compensation strategy knowledge base, which includes the heating rate adjustment rule, the pressure loading curve optimization strategy and the holding time adjustment model, to obtain the matched strategy template data;

[0126] Based on the dominant error source classification label, a strategy template retrieval algorithm (parameters: label exact match mode=true, priority sorting rule=confidence descending order) is used to locate the strategy template index corresponding to the current dominant error source type in the preset compensation strategy knowledge base, to realize the preliminary screening of the strategy template;

[0127] Further, by a strategy adaptation degree calculation method (parameters: adaptation degree function type=weighted cosine similarity, feature weight source=device process parameter sensitivity coefficient library), the similarity calculation is performed between the preliminary screened multiple candidate strategy templates and the process state feature vector of the current pressing batch, and the adaptation degree score vector A={a1, a2,..., a n} is obtained;

[0128] Further, a threshold filtering algorithm (parameters: adaptation degree threshold=0.75) is used to filter out templates with adaptation degree scores below the threshold, and a high matching strategy template set that meets the requirements is retained, and the set elements are sorted in descending order according to the adaptation degree value, so as to preferentially select the optimal template when there are multiple candidate templates;

[0129] Further, by a strategy element decomposition and rule mapping algorithm (parameters: rule parsing mode=structured JSON parsing, key fields={heating rate adjustment rule, pressure loading curve optimization strategy, holding time adjustment model}), the internal structure of the optimal template is parsed, the general adjustment rule data of the three control dimensions of heating rate, pressure curve and holding time are extracted, and they are stored as structured strategy template data objects;

[0130] Through the above search, matching, filtering and analysis processing mode, the dominant error source classification label is matched to the optimal compensation control strategy template, and the matching strategy template data containing the temperature rise rate adjustment rule, the pressure loading curve optimization strategy and the pressure holding time adjustment model are output, realizing seamless connection with the next parameter mapping processing link;

[0131] For example, in a FR4 / polyimide mixed lamination batch, the dominant error source classification label is "thermal expansion difference". The compensation strategy knowledge base contains 12 strategy templates, each with a corresponding applicable process feature vector (dimension 9, including material thermal expansion coefficient, target temperature rise curve slope, equipment rated pressure range, etc.). The process state feature vector of the current batch and the template feature vector stored in the knowledge base are calculated by weighted cosine similarity to adapt the degree, and the weight vector comes from the equipment process parameter sensitivity coefficient library (for example, thermal expansion coefficient weight 0.32, temperature rise rate weight 0.28, pressure holding time weight 0.24, pressure loading curve weight 0.16). Taking a candidate template feature vector V1 and a process state feature vector Vt as an example, the adaptation degree calculation formula is:

[0132]

[0133] Where w is the feature weight vector, and "·" represents element-wise multiplication and inner product operation. The template adaptation value is 0.812, which is higher than the set threshold 0.75, so it is retained in the candidate set. Finally, in the candidate template, the template V5 with the highest adaptation degree contains the temperature rise rate adjustment rule: the target temperature rise curve is divided into three sections, the first section rate is 2.5℃ / min, the second section rate is 3.0℃ / min, and the third section rate is reduced to 1.5℃ / min; the pressure loading curve optimization strategy: the initial pressure stage is loaded at 15MPa for 60 seconds, the holding stage is at 18MPa, and the slow descent stage is at 10MPa; the pressure holding time adjustment model: 12% longer than the reference process. The matching strategy template data is output as a structured object to S5.3 step, realizing high-precision matching and parameter adaptation of the control instruction generation process;

[0134] S5.3: Perform parameter mapping processing on the matched strategy template data, map the general control parameters in the strategy template to the control interface parameter space of the current lamination equipment, to generate an executable lamination parameter adjustment instruction set, including the target temperature rise rate, the segmented pressure curve and the pressure holding time window;

[0135] After obtaining the matching strategy template data containing the temperature rise rate adjustment rule, the pressure loading curve optimization strategy and the pressure holding time adjustment model, the parameter key-value mapping algorithm is used (parameters: mapping mode = bidirectional one-to-one mapping, allowed deviation range = ±1%), to realize accurate matching of the general control parameter field in the template and the lamination equipment control interface field;

[0136] Further, by controlling parameter normalization processing method (parameter: dynamic normalization interval based on real-time calculation of device rated parameter range), control parameters of different physical quantities are mapped to dimensionless space of 0 to 1 to ensure multi-parameter consistency in subsequent instruction set generation process;

[0137] Further, based on parameter scale transformation and correction algorithm (parameter: transformation function type = segmented linear scaling, correction strategy = second-order polynomial compensation), the normalized parameters are reflected to the physical control quantity space that can be actually executed by the device, ensuring that the generated target heating rate, segmented pressure curve and pressure holding time window meet the dynamic response characteristics of the field device;

[0138] Further, by multi-dimensional parameter sequence interpolation and smoothing algorithm (interpolation method = cubic spline interpolation, smoothing window width = 5 points), the segmented pressure curve and the heating rate curve are time series refined and noise suppressed to improve the smoothness and continuity of instruction execution;

[0139] Further, a parameter correlation consistency detection mechanism (detection formula: where Pt i is the mapped target parameter value, Pr i is the maximum allowed parameter value of the device interface) is used to verify whether the generated control parameters exceed the device execution limit, and if an abnormality is detected, the scaling correction is performed to avoid hardware overload risk;

[0140] Through the chain processing of parameter key value mapping, normalization processing, scale reflection mapping, sequence smoothing and consistency detection, the matching strategy template data is successfully converted into executable compression parameter adjustment instruction set that meets the control interface parameter space of the compression device, realizing the precise preparation of the heating rate, segmented pressure curve and pressure holding time window;

[0141] For example, in a FR4 / polyimide mixed laminated compression batch, the general heating rate parameter in the matching strategy template data is {2.5, 3.0, 1.5} ℃ / min, the pressure curve is {15, 18, 10} MPa corresponding to the duration {60, 180, 90} seconds, and the pressure holding time extension ratio is 12%. In the parameter key value mapping stage, the template field "TempRatePhase1" corresponds to the device interface field "HTR_SEG1_RATE", and the mapping deviation meets ±1%. In the normalization stage, 2.5 ℃ / min is mapped to 0.416 (device heating range 0-6 ℃ / min), and pressure 18 MPa is mapped to 0.6 (device pressure range 0-30 MPa). In the scale reflection mapping stage, the normalized 0.416 is reflected by the reflection mapping and the second-order compensation function y = 6x + 0.02x 2Convert back to physical quantities, get 2.508 ℃ / min. Sequence smoothing stage, on the basis of piecewise pressure curve (15, 18, 10) MPa, interpolation to generate curve point set with second precision and carry out cubic spline smoothing, reduce peak mutation. In the consistency detection, the highest pressure segment calculation value is 18.05 MPa, which exceeds the rated upper limit of the equipment 18 MPa by 0.27%, is reduced to 18.0 MPa in proportion, and the overload risk is eliminated. The output control parameter instruction set includes the corrected multi-stage heating rate, the smoothed pressure time curve and the corrected pressure holding time, which meets the control instruction packaging requirements of subsequent S5.4, and realizes the performance indicators of temperature control accuracy ±0.1 ℃ / min and pressure control error ±0.05 MPa in the field verification;

[0142] S5.4: Based on the pressure adjustment parameter instruction set, execute control instruction formatting packaging processing to generate control instruction data packet conforming to the control protocol of the pressure device, the control instruction data packet contains instruction header, parameter field and check code, which is used to ensure the accurate transmission and execution of the control instruction;

[0143] S5.5: Send the formatted control instruction data packet to the main control unit of the pressure device to trigger the dynamic adjustment operation of the pressure parameters, realize the active compensation of the current alignment deviation and the deviation suppression in the subsequent pressure process.

[0144] The step S6: based on the control instruction set, execute parameter adjustment operation on the pressure device to form a closed loop control loop to dynamically compensate the current alignment deviation and suppress the deviation accumulation in the subsequent pressure process. Specifically, it includes:

[0145] S6.1: Based on the confidence degree of each error source in the error source list, extract the dominant error source type, which includes thermal expansion difference, material shrinkage abnormality or pressure distribution unevenness, to determine the corresponding process parameter adjustment strategy;

[0146] S6.2: According to the dominant error source type, call the preset control strategy knowledge base to match the corresponding compensation control algorithm model to generate the parameter adjustment rule set for the current pressure state;

[0147] Under the condition that the dominant error source type has been identified and the corresponding pressure process state data has been obtained, read the compensation control algorithm library index corresponding to the error source type in the preset control strategy knowledge base as the basic input for subsequent rule set generation;

[0148] The strategy retrieval and algorithm model mapping method (parameter: retrieval mode = error type index matching, enable fuzzy matching coefficient 0.85) is adopted to realize one-to-one or many-to-one mapping relationship between the current dominant error source type and the compensation control algorithm model in the knowledge base, and extract the candidate algorithm model set;

[0149] Further, the quantitative adaptation degree evaluation of the candidate algorithm model and the current batch of lamination state is realized by the lamination state feature vector similarity calculation method (parameters: algorithm = weighted Euclidean distance, weight source = process parameter sensitivity matrix), and an adaptation degree score vector S = {s1, s2,..., sN} is obtained. m};

[0150] Further, the algorithm model with a score lower than the set value is filtered out by the adaptation degree threshold filtering method (parameter: threshold 0.78), so as to retain the high matching degree compensation control algorithm set to ensure the pertinence and effectiveness of the subsequent executed parameter modification strategy;

[0151] Further, the core control variables and adjustment formulas defined in the algorithm model are extracted by the algorithm model rule analysis engine (analysis mode = structured analysis, key field = {heating rate calculation function, pressure curve generation function, holding time correction function}), for example, the heating rate target calculation formula R = f (α, ΔT, K), wherein α is the material thermal expansion coefficient, ΔT is the target temperature difference, and K is the device temperature control gain coefficient;

[0152] Further, the calculation sub-modules corresponding to each control variable are fused to construct a parameter adjustment rule set containing multi-physical quantity coupling relationship, covering the dimensions of heating rate calculation, segmented pressure curve optimization and holding time correction, so as to ensure the input integrity and accuracy of the subsequent fuzzy logic reasoning processing;

[0153] Through the above retrieval, matching, analysis and fusion processing mode, the compensation control algorithm model associated with the error source type is converted into the parameter adjustment rule set adapted to the current lamination state feature, realizing the technical effect of directly driving the control instruction set generation;

[0154] For example, in a FR4 / polyimide hybrid lamination batch, the dominant error source type is "uneven pressure distribution", and the process state feature vector is {185, 0.65, 0.12}, which respectively represents the current molding temperature (℃), the pressure uniformity coefficient and the tension fluctuation ratio. The knowledge base contains 5 compensation control algorithm models corresponding to "uneven pressure distribution", and the weighted Euclidean distance is used to calculate the adaptation degree of the current process state, and the weight vector is {0.4, 0.35, 0.25}. Taking the feature vector {180, 0.60, 0.15} of the candidate model M3 as an example, the adaptation degree D = 1-0.4×(185-180) 2 +0.35×(0.65-0.60) 2 +0.25×(0.12-0.15) 2= 0.82, which is higher than the threshold value 0.78, and thus enters the high matching degree set. After analyzing the M3 algorithm model, the temperature rising rate calculation function R = 2.8℃ / min is obtained, and the segmented pressure curve optimization strategy is that the first segment pressure is increased by 5%, and the holding pressure time is extended by 8%. Combined with the analysis values of other high matching models, the parameter adjustment rule set formed by rule fusion contains the precise adjustment relationship of the three control variables of temperature rising, pressure and holding pressure. The control correction parameters are calculated by the subsequent input S6.3 fuzzy logic reasoning engine to realize the control target of improving the pressure distribution uniformity to more than 0.85;

[0155] S6.3: Perform fuzzy logic reasoning processing on the parameter adjustment rule set, and combine the real-time data of temperature, pressure and tension in the current pressing stage to calculate the correction parameters of temperature rising rate, pressure loading curve and holding pressure time to generate a control instruction set;

[0156] S6.4: Transmit the control instruction set to the PLC controller of the pressing equipment through the industrial communication bus to perform dynamic adjustment operation on the pressing parameters to realize online compensation for the current interlayer alignment deviation;

[0157] S6.5: After the parameter adjustment is completed, the change trend of multi-dimensional process data is continuously monitored, and the adjusted pressing process data is fed back to the causal reasoning engine to update the error source weight model and optimize the control strategy of the subsequent batch.

[0158] The step S7: After the pressing is completed, the actual interlayer alignment accuracy data is collected, compared with the predicted deviation value, a deviation compensation effect evaluation report is generated, and is used to update the error source weight model in the causal chain knowledge base. Specifically, it includes:

[0159] S7.1: Based on the actual interlayer alignment accuracy data collected after the pressing is completed, the alignment coordinates output by the image acquisition system are subjected to coordinate system normalization processing to eliminate the influence of equipment installation error on the measurement results, and standardized interlayer alignment error measurement values are obtained;

[0160] S7.2: The standardized interlayer alignment error measurement values are compared with the predicted deviation values output by the error analysis module point by point, the deviation residual errors of each key alignment feature point are calculated to obtain a deviation prediction error distribution map;

[0161] S7.3: Based on the deviation prediction error distribution map, the overall prediction error index is calculated by using the residual sum of squares algorithm, and a deviation compensation effect evaluation report is generated in combination with a preset evaluation threshold value. The evaluation report includes the deviation compensation success rate, the maximum residual value and the compensation stability score;

[0162] Based on the obtained deviation prediction error distribution map, the residual sum of squares algorithm (parameters: normalized mode = normalized by the maximum deviation absolute value, weight distribution = feature point importance coefficient set) is used to realize the quantitative calculation of the overall prediction error;

[0163] Further, by the residual value e i of each key alignment feature point

[0164]

[0165] Where w i is the feature point importance weight;

[0166] Further, by setting the evaluation threshold T E (according to the historical batch error mean and process tolerance), the comprehensive error value is compared with T E to determine whether the compensation effect meets the process target;

[0167] Further, the deviation compensation success rate is calculated, where N is the total number of feature points, N succ is the number of feature points whose residual error is lower than the tolerance limit after compensation;

[0168] Further, the maximum value of the residual absolute value sequence is extracted as the maximum residual value R max , and the ratio of the standard deviation to the mean of the residual value after compensation is calculated to obtain the compensation stability score S stab ;

[0169] Through the above residual sum of squares calculation and statistical analysis processing method, the deviation prediction error distribution map of the previous step is converted into a deviation compensation effect evaluation report containing the deviation compensation success rate, the maximum residual value and the compensation stability score, realizing the quantitative verification and performance evaluation of the compensation strategy effectiveness;

[0170] For example, in a 12-layer FR4 lamination batch, the deviation prediction error distribution map contains 120 key feature point residual values, and the importance weight vector is set to 1.0 in the positioning hole area, 0.8 in the signal trace area, and 0.5 in the edge trace area. After residual sum of squares calculation, E = 3.25 μm 2 , the evaluation threshold T E is 4.0 μm 2 , and the compensation is determined to be up to standard. There are 110 residual errors lower than the 2 μm tolerance limit in the feature points, so the compensation success rate is The maximum residual value R max is 2.85 μm, the residual standard deviation is 0.42 μm, and the mean is 1.15 μm, so the compensation stability score Sstab = 0.365. The final evaluation report shows that the compensation strategy significantly reduces the overall alignment deviation and performs well in stability, providing a reliable basis for subsequent batch process parameter optimization;

[0171] S7.4: input the deviation prediction error distribution map and the corresponding process parameter time series into the error source weight update model, and use the Bayesian inference method to dynamically correct the propagation weights of each error source in the causal chain knowledge base to obtain an updated error source weight parameter set;

[0172] S7.5: update the error source model in the causal chain knowledge base based on the updated error source weight parameter set, and write the updated knowledge base model into the model cache area of the causal reasoning engine to improve the traceability accuracy and response speed of the alignment deviation in the subsequent batch pressing process.

[0173] The step S8: dynamically monitor the distribution change trend of multi-dimensional process data in continuous multi-batch pressing process, and trigger the causal chain model version switching mechanism when detecting process condition mutation, to adapt to the error propagation characteristics in the current manufacturing environment. Specifically, it includes:

[0174] S8.1: divide the multi-dimensional process data collected in the continuous multi-batch pressing process into a sliding window to construct a dynamic time series data set, which reflects the time evolution characteristics of the process state;

[0175] S8.2: based on the dynamic time series data set, use principal component analysis algorithm for feature dimension reduction processing, extract the principal component feature vector of the key process variable, to reduce the computational complexity of subsequent trend detection;

[0176] S8.3: perform sliding window variance calculation and K-L divergence analysis on the principal component feature vector to identify the statistical deviation degree between the current process data distribution and the historical baseline distribution, and generate a distribution deviation index;

[0177] S8.4: compare the distribution deviation index with the preset process condition mutation threshold to determine whether there is a significant process state change, and generate a process condition mutation trigger signal if there is;

[0178] S8.5: in response to the process condition mutation trigger signal, based on the matching degree calculation result of the current process environment characteristics and the pre-stored causal chain model version, select the appropriate causal chain knowledge base version, complete the model switching operation, to adapt to the dynamic change of the current error propagation path.

[0179] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0180] The above only describes the preferred embodiments of the present application and is not intended to limit the present application; the present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included in the protection scope of the present application.

Claims

1. A method for tracing the alignment deviation during lamination of multilayer composite circuit boards, characterized in that, The method comprises the following steps: S1: During the pressing process of the multi-layer composite circuit board, the multi-modal sensor array is used to collect process parameters including temperature, pressure, tension and interlayer alignment image, and a multi-dimensional process data set is generated; S2: The interlayer alignment image is subjected to image comparison processing, the interlayer alignment deviation atlas is extracted, the key alignment feature points in the deviation atlas are vectorized and coded, and a structured deviation feature data is generated; S3: Based on the preset causal chain knowledge base, an alignment deviation propagation path model is constructed; S4: The structured deviation feature data and the multi-dimensional process data set are input into the causal reasoning engine, the probability weight of each potential error source is dynamically updated, and an error source list with confidence ranking is output; S5: According to the dominant error source type in the error source list, a corresponding compensation control strategy is selected, and a control instruction set for adjusting the pressing parameters is generated; S6: Based on the control instruction set, parameter adjustment operation is performed on the pressing equipment to form a closed loop control circuit, so as to dynamically compensate the current alignment deviation and suppress the deviation accumulation in the subsequent pressing process; S7: After the pressing is completed, the actual interlayer alignment accuracy data is collected, compared with the predicted deviation value, a deviation compensation effect evaluation report is generated, and used to update the error source weight model in the causal chain knowledge base; S8: In the continuous multi-batch pressing process, the distribution change trend of the multi-dimensional process data is dynamically monitored, and the causal chain model version switching mechanism is triggered when the working condition is suddenly changed.

2. The method of claim 1, wherein the method further comprises: The step S1 specifically comprises: During the pressing process of the multi-layer composite circuit board, based on the distributed sensor network architecture in the industrial control field, temperature sensors, pressure sensors, tension sensors and high-resolution visual cameras are deployed at key nodes of the pressing equipment to synchronously collect multi-modal process parameters during the pressing process; The original temperature data collected by the temperature sensor is subjected to temperature compensation algorithm and filtering processing to obtain the real-time thermal field distribution characteristics of each layer material during the pressing process; The original pressure signal collected by the pressure sensor is subjected to sliding window mean filtering and outlier rejection processing to extract the pressure loading curve of each pressing plate area during the pressing process, and a pressure distribution matrix with time sequence characteristics is generated; The tension data collected by the tension sensor is subjected to normalization processing and material elastic modulus mapping to obtain the real-time stress state parameters of each layer substrate during the pressing process; The interlayer alignment image collected by the visual camera is subjected to image preprocessing operation to extract the geometric profile characteristics of the interlayer positioning hole and the conductive circuit, and a basic image data set for image comparison is generated; The temperature, pressure, tension and image data output by the multi-modal sensor are subjected to multi-source data fusion processing based on the time stamp to construct a multi-dimensional process data set containing time sequence characteristics.

3. The method of claim 2, wherein the method further comprises: The distributed sensor network adopts a hybrid topology of star and ring type, the sensor nodes are connected through industrial Ethernet and fiber dual ring network, and the PTP protocol is used to distribute a unified time reference.

4. The method of claim 1, wherein the method further comprises: The step S2 specifically comprises: The collected interlayer alignment image is subjected to grayscale and histogram equalization processing to obtain standardized grayscale image data; Key point detection and descriptor extraction are performed on the standardized gray image data to obtain a set of image feature points with scale and rotation invariance, and a local feature description vector set is generated; A nearest neighbor matching strategy is used to match the local feature description vector sets of adjacent layer images, identify the inter-layer correspondence, and generate a preliminary alignment deviation vector set; An outlier rejection and transformation matrix estimation are performed on the preliminary alignment deviation vector set to obtain inter-layer rigid transformation parameters and generate an optimized alignment deviation atlas; Vectorization encoding is performed on the key alignment feature points in the optimized alignment deviation atlas, multi-dimensional feature tuples are constructed based on geometric coordinates and deviation direction information, and structured deviation feature data is generated.

5. The method of claim 4, wherein the method further comprises: The structured deviation feature generation includes a chain processing mode based on grayscale, histogram equalization, median filtering, and adaptive histogram equalization, which converts the original color inter-layer alignment image into standardized gray image data with global contrast balance, local detail enhancement, and noise suppression, and uses the SIFT algorithm to extract key points, the RANSAC algorithm for feature point matching and outlier rejection, and multi-dimensional structured deviation feature vectors are formed through clustering, centroid, and direction consistency calculation.

6. The method of claim 1, wherein the method further comprises: The step S3 specifically includes: Error source types in the preset causal chain knowledge base are classified and modeled, a propagation path atlas of error sources is constructed based on process experience and historical data, and a structured causal chain network model is formed; The deviation atlas features output by the image comparison module are semantically analyzed, the spatial distribution features, deviation direction consistency index, and local deformation coefficient of the key alignment offset points are extracted, and a structured deviation feature vector set is generated; Based on the material attribute parameters and pressing process parameters of the multi-layer composite circuit board, the thermal expansion coefficient matrix and solidification shrinkage prediction value of each layer of material in the pressing process are calculated, and dynamic material response features are generated; The structured deviation feature vector set and the dynamic material response features are input into the causal chain network model, the error propagation path with the highest matching degree with the current deviation atlas is activated through the forward propagation mechanism, and the activation probability of each node in the path is normalized and weighted to generate a preliminary error source activation atlas; Based on the Bayesian inference algorithm, the preliminary error source activation atlas is updated with posterior probability, the temperature gradient, pressure distribution curve, and tension feedback value in the multi-dimensional process data are fused, an error source causal chain path with confidence weight is generated, and a causal chain explanation model for the current alignment deviation is formed.

7. The method of claim 1, wherein the method further comprises: The step S4 specifically includes: Data alignment processing is performed on the structured deviation feature data and multi-dimensional process data set to obtain a synchronized multi-modal input data set; Based on the pre-defined error propagation path model in the causal chain knowledge base, a multi-source error source candidate set is constructed; A causal graph modeling is performed on the synchronized multi-modal input data set using a causal reasoning engine, the conditional probability distribution between each error source and the deviation feature is calculated, and an error source-feature association matrix is generated; A Bayesian updating algorithm is applied to the error source-feature correlation matrix to dynamically adjust the prior probability distribution of each error source in combination with the historical process parameter variation trend of the current batch of pressing process; Error source ranking is performed based on the updated posterior probability distribution to output an error source list with confidence weights.

8. The method of claim 7, wherein the method further comprises: The data alignment process performs multi-modal synchronization based on a timestamp-indexed data alignment method, missing data is completed by linear interpolation, and normalized data is input into the causal inference unit.

9. The method of claim 1, wherein the method further comprises: The step S5 specifically includes: Based on the confidence ranking of each error source in the error source list, the dominant error source type is identified, and a dominant error source classification label is obtained; According to the dominant error source classification label, the corresponding compensation control strategy template is retrieved from the preset compensation strategy knowledge base to obtain matched strategy template data; Parameter mapping processing is performed on the matched strategy template data to map the general control parameters in the strategy template to the control interface parameter space of the current pressing equipment, generating a set of pressing parameter adjustment instructions; Based on the pressing parameter adjustment instruction set, control instruction formatting and packaging processing is performed to generate control instruction data packets conforming to the control protocol of the pressing equipment; The control instruction data packets are sent to the main control unit of the pressing equipment to trigger the dynamic adjustment operation of the pressing parameters.

10. The method of claim 9, wherein the method further comprises: The error source type includes one or more combinations of thermal expansion difference, inconsistent material shrinkage rate, and uneven pressure distribution.

Images