High multilayer board processing method for inter-board glue filling uniformity

Abstract

The invention provides an inter-board glue filling uniformity-oriented high multilayer board processing method, which comprises the following steps of: integrating a high-temperature stable annular piezoelectric transducer array to carry out multi-region sound wave signal acquisition, improving signal quality through wavelet denoising and normalization, extracting time sequence characteristics such as sound velocity, attenuation coefficient and nonlinear parameters, and carrying out multi-region sound wave signal acquisition; establishing a mapping relation between sound parameters and colloid modulus by means of a graph neural network, and realizing real-time generation of a pressure compensation amount matrix; the distributed robust prediction controller is adopted to perform optimal distribution on driving signals of the flexible piezoelectric execution unit, sub-millimeter pressure dynamic redistribution and uniform glue filling are realized, the method has the advantages of high pressure spatial resolution, real-time adaptive adjustment and strong anti-interference capability, and the automatic control precision and the product quality of a pressing process can be improved.

Description

Technical Field

[0001] This invention relates to the field of electronic manufacturing process control technology, and in particular to a method for processing high multilayer boards with uniform inter-board adhesive filling. Background Technology

[0002] Currently, high-multilayer printed circuit board (PCB) lamination technology is widely used in high-end electronics manufacturing. Its core objective is to leverage the synergistic effect of pressure and temperature to successfully press multiple layers of copper foil and prepreg (PP) into a single board material with excellent electrical properties and structural consistency. In existing technologies, the pressure distribution during lamination is primarily regulated through mechanical control, heating parameter optimization, mold rigidity compensation, and material thickness homogenization. Common industry measures include localized padding with a top rigid plate / buffer paper, adjusting vacuum low pressure and heating rate, selecting different grades of PP sheets to improve resin flow boundaries, optimizing venting groove structure, and suppressing copper foil waviness deformation. These measures alleviate, to some extent, the problem of poor resin filling caused by inconsistent overall pressure distribution. Currently, traditional methods for improving pressure uniformity rely on static measures such as increasing mold rigidity, optimizing copper foil / glass cloth laying design, and adjusting the formulation or thickness of PP prepreg. These methods essentially determine the subsequent pressure distribution characteristics before pressing and have limited responsiveness to dynamic disturbances during the pressing process (material viscoelasticity changes, local gas leakage, differences in curing rates, etc.). In addition, the internal material distribution of high-multilayer PCB stacks is extremely complex. The slight differences in the thickness of copper foil, glass cloth, core board, etc., lead to obvious non-uniformity of the pressure field. Conventional constant pressure pressing processes are difficult to adapt to timely compensation and fine control of local thin adhesive layers, which can easily result in some areas having excessively low adhesive thickness (insufficient adhesive), excessively high adhesive thickness (overflow adhesive), or porosity. This can lead to a series of quality problems such as fluctuations in the electrical performance of subsequent products, interlayer short circuits, and structural failures. Existing multi-area pressure feedback compensation technologies are constrained by various factors, including sensor deployment, high-temperature resistance of signal cables, high-temperature insulation reliability of dielectric solutions, and the complexity of dynamic calibration. Dielectric, thickness-based, and temperature-responsive sensing methods generally suffer from accuracy degradation under high-temperature environments, increased system complexity, and poor long-term operational stability. Furthermore, they often rely on changes in the intrinsic properties of materials, requiring separate calibration for different stack-up structures and material systems, resulting in a heavy burden on on-site maintenance. In addition, traditional methods based on mechanical parameter adjustments (such as replacing pressure plates, adjusting venting, or adjusting copper foil structure) have a lag effect and cannot achieve real-time adaptive correction during the pressing process. They struggle to dynamically compensate for actual pressure distribution deviations in different areas of the circuit board, and are clearly insufficient for the increasingly complex quality control requirements of high-multilayer PCBs. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a high-multilayer board processing method for achieving uniform glue filling between boards.

[0004] The technical solution of this invention is implemented as follows: a processing method for high-multilayer boards with uniform inter-board glue filling, comprising: S1: Collect raw data of transmitted sound waves from multiple regions during the pressing process to obtain the spatial distribution differences of sound propagation parameters at different locations; S2: Perform denoising and normalization preprocessing on the original transmitted acoustic wave data to eliminate background noise interference and improve the signal-to-noise ratio; S3: Based on the preprocessed acoustic wave data, extract the temporal features of acoustic parameters, including sound velocity, attenuation coefficient and nonlinear parameters; S4: Based on the correspondence between the temporal characteristics of acoustic parameters and the local dynamic shear modulus and loss modulus of colloids in offline calibration data, a graph neural network mapping model is constructed to generate the pressure compensation matrix. S5: Input the real-time extracted acoustic parameter time-series features into the graph neural network mapping model to obtain the current pressure compensation matrix; S6: Input the pressure compensation matrix into the distributed robust predictive controller, which generates the drive signal of the flexible piezoelectric actuator based on the thermodynamic state of the pressing process; S7: Based on the driving signal, control the flexible piezoelectric actuator to apply dynamic pressure distribution to optimize the uniformity of glue filling between boards; S8: Monitor changes in acoustic propagation parameters during the pressing process. If parameter fluctuations exceed the threshold, trigger adaptive adjustment of pressure compensation parameters.

[0005] The high-multilayer board processing method for improving the uniformity of inter-board adhesive filling provided by this invention has the following beneficial effects: (1) This invention introduces the acoustic radiation force and acoustic-flow coupling effect of ultrasonic waves propagating in a viscoelastic medium as the core sensing mechanism of the pressing process, and constructs a new closed-loop control paradigm based on the dynamic inversion of resin rheological state by in-situ acoustic parameters; by using a ring piezoelectric transducer array to excite and receive pulsed ultrasonic signals on the non-contact side, and combining the multi-dimensional feature extraction of time-domain waveform, phase shift and envelope attenuation rate, the spatial-temporal synchronous identification of local shear modulus and loss modulus of semi-cured resin is realized, which significantly improves the observability and analytical accuracy of colloidal flow behavior in the high viscosity state at the beginning of the process, and enables the system to intervene in advance at the most critical heating stage of resin flowability, greatly enhancing the spatiotemporal matching capability of pressure distribution; (2) This invention establishes an offline mapping model of “acoustic parameter time sequence characteristics → rheological parameters → optimal pressure compensation amount” based on graph neural network, and combines it with a distributed robust predictive controller to generate a dynamic pressure compensation matrix of the whole area in real time online, realizing adaptive pressure redistribution control at millimeter-level spatial resolution; with the micro piezoelectric film stack execution unit integrated on the microstructured silicone substrate, its single-point response time is ≤3.2ms and the stroke is 4.5μm. After flexible coupling through the micro-bump array, it can accurately transmit the control command to the surface of the mirror steel plate. Without increasing the total pressing cycle, it significantly reduces the standard deviation of the glue filling uniformity between boards and significantly reduces the local glue deficiency rate, which fully ensures the reliable filling quality of micro-holes and blind buried holes inside the high-multilayer PCB; this control architecture does not need to rely on complex PP sheet formula adjustment, venting groove design or buffer paper configuration and other auxiliary means, and also avoids the risk of insulation failure of dielectric sensor in high temperature environment and the maintenance cost caused by frequent calibration, which significantly improves the industrial applicability, long-term operation stability and repeatability of field deployment of the system; (3) The overall technical path of this invention deeply integrates the principles of ultrasonic physics, soft matter mechanics, and edge intelligent control. For the first time, it systematically embeds the acoustic microfluidic mechanism into the closed loop of high-multilayer PCB lamination process, forming a new intelligent manufacturing mode with autonomous perception, intelligent decision-making, and precise execution capabilities. The entire process does not require modification of vacuum low-pressure equipment or involve additional processes such as liquid glue pre-filling or browning film control. While maintaining the compatibility of the original production line, it realizes a fundamental shift from "experience-driven" to "data-driven" and has good interpretability and scalability. It is suitable for flexible production scenarios with various stacking types and board thickness combinations. The above innovations not only break through the technical bottleneck of the difficulty in balancing pressure uniformity and dynamic response speed in traditional lamination processes, but also provide a new paradigm of process control with high reliability, high yield, and low maintenance costs for the high-end electronic packaging field, which has significant engineering application value and industrialization prospects. Attached Figure Description

[0006] Figure 1 This is a flowchart of a high-multilayer board processing method for achieving uniform inter-board adhesive filling according to the present invention. Figure 2 This is a sub-flowchart of the high-multilayer board processing method for achieving uniform inter-board adhesive filling according to the present invention. Figure 3 This is another sub-flowchart of the high-multilayer board processing method of the present invention, which aims to achieve uniform glue filling between boards. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, the present invention provides a method for processing high-multilayer boards with uniform inter-board adhesive filling, specifically including: S1: Collect time-domain waveforms, phase shifts, and envelope attenuation rate data of transmitted acoustic waves from multiple regions during the pressing process to obtain the spatial distribution differences of acoustic propagation parameters at different locations; S2: Perform denoising and normalization preprocessing on the transmitted acoustic wave data to eliminate background noise interference and improve the signal-to-noise ratio; S3: Based on the preprocessed acoustic wave data, extract the temporal features of acoustic parameters, including sound velocity, attenuation coefficient and nonlinear parameters; S4: Based on the correspondence between the temporal characteristics of acoustic parameters and the local dynamic shear modulus and loss modulus of colloids in offline calibration data, a graph neural network mapping model is constructed to generate the pressure compensation matrix. S5: Input the real-time extracted acoustic parameter time-series features into the graph neural network mapping model to obtain the current pressure compensation matrix; S6: Input the pressure compensation matrix into the distributed robust predictive controller, which generates the drive signal of the flexible piezoelectric actuator based on the thermodynamic state of the pressing process; S7: Based on the driving signal, control the flexible piezoelectric actuator to apply dynamic pressure distribution to optimize the uniformity of glue filling between boards; S8: Monitor changes in acoustic propagation parameters during the pressing process. If parameter fluctuations exceed the threshold, trigger adaptive adjustment of pressure compensation parameters.

[0010] Step S1: Collect time-domain waveforms, phase shifts, and envelope attenuation rate data of transmitted acoustic waves from multiple regions during the pressing process to obtain the spatial distribution differences of acoustic propagation parameters at different locations. Specifically, this includes: S1.1: Based on lead zirconate titanate-lead magnesium niobate composite ceramic material, a ring piezoelectric transducer array is integrated on the non-contact side of the upper and lower molds of the pressing mold to form an acoustic wave transmitting and receiving unit with a center frequency of 1.2 MHz and a bandwidth of ±150 kHz. This unit maintains a stability of electromechanical coupling coefficient greater than 0.68 in a high temperature environment of 200 degrees Celsius and outputs a calibrated non-contact acoustic wave sensing array. Based on the non-contact side of the upper and lower molds of the pressing mold as the implementation object, the array integration method of lead zirconate titanate-lead magnesium niobate composite ceramic material (parameters: center frequency 1.2MHz, bandwidth ±150kHz) is adopted to realize the ring spatial layout of the piezoelectric transducer array, so as to form a high-frequency sound wave transmission and reception structure unit. Furthermore, through a vacuum-assisted curing process (parameter: vacuum degree) (0.095MPa, curing temperature range ≤200℃), the composite ceramic material is uniformly embedded into the array slots on the non-contact side of the mold to ensure that there are no air bubbles inside the material and to maintain a stable performance with an electromechanical coupling coefficient greater than 0.68; Furthermore, by utilizing the array unit impedance matching algorithm (parameters: matching frequency band 1.05–1.35MHz, impedance tolerance ≤0.5Ω), topological matching between the transducer unit and the drive circuit is achieved to reduce acoustic energy transmission loss and enhance signal transmission and reception sensitivity under high temperature conditions. Furthermore, by using the array unit phase synchronization calibration method (parameter: phase offset tolerance ≤ 2°), the phase response of each ring array unit is adjusted under the excitation of a multi-channel signal generator to achieve full array acoustic wave synchronous transmission state, thereby eliminating spatial beam distortion; By using a temperature-controlled acoustic characteristic testing and processing method, the transducer array integrated in the previous step is transformed into a calibrated non-contact acoustic wave sensing array, thereby achieving non-destructive and high-precision acquisition of acoustic propagation parameters in multiple regions during the pressing process. For example, in a high-multilayer PCB lamination process, eight ring array units are arranged on the non-contact side of the upper mold. The ring diameter is 120mm, and the center-to-center distance between units is 15mm. Each unit is composed of a lead zirconate titanate-lead magnesium niobate composite ceramic sheet with a diameter of 5mm and a thickness of 1mm. A vacuum-assisted curing process is used to embed the ceramic sheet into the slot, and the vacuum degree is set to... The pressure was 0.095 MPa, the curing temperature was 180℃, and the holding time was 120 minutes. Using an impedance matching algorithm, the impedance of the drive circuit was adjusted to a specified tolerance range at a center frequency of 1.2 MHz, with a matching frequency band of 1.05–1.35 MHz and an impedance tolerance not exceeding 0.5 Ω. A synchronization pulse was applied using a multi-channel signal generator to perform phase calibration of the array units, controlling the phase offset tolerance within 1.8°. The following electromechanical coupling coefficient calculation formula was used:

[0011] in, The resonant frequency, It is the anti-resonant frequency, as tested. =1.198MHz, =1.202MHz, calculated =0.687, meeting the stability requirement of greater than 0.68. The final output non-contact acoustic wave sensor array achieves accurate acquisition of multi-region sound velocity, attenuation, and nonlinear parameters during the pressing process test, significantly improving the input data quality of the pressure distribution optimization stage; S1.2: Perform pulse modulation ultrasonic signal configuration processing on the integrated piezoelectric transducer array, set the excitation parameters with a pulse width of 20 microseconds and a repetition frequency of 5 kHz to generate a high-frequency ultrasonic excitation signal suitable for semi-cured resin adhesive layer detection, and output the signal transmission unit with the configured parameters. The integrated annular piezoelectric transducer array is input with calibrated center frequency and bandwidth parameters. A pulse-modulated ultrasonic signal generation method (pulse width set to 20 microseconds, repetition frequency set to 5 kHz) is used to construct a high-frequency excitation signal for the semi-cured resin adhesive layer. Furthermore, by using a time-domain waveform synthesis algorithm (with the input being a set frequency range and pulse width parameters), the amplitude and phase of the pulse envelope are synchronously modulated, and the fundamental wave of the excitation signal with the target spectral structure and the necessary second harmonic components are obtained. Furthermore, a frequency domain amplitude normalization processing method (with parameters being the actual transmit power curve of the transducer array) is adopted to achieve consistency of the transmit signal amplitude across all array units and generate a normalized modulation signal matrix. Furthermore, a timing synchronization control algorithm (with a normalized modulation signal matrix and a transducer array driving topology as inputs) is used to achieve phase synchronization and delay compensation for multi-channel signals, and the phase error is limited to within... The driver instruction set within; By using signal power control (the parameters are the target sound pressure level and the transmission power curve), the driving command of the previous step is converted into a high-frequency ultrasonic excitation signal that is adapted to the detection of the semi-cured resin adhesive layer, and the output signal transmission unit with the parameters configured is used to achieve stable excitation of multi-region transmitted sound waves during the pressing process. For example, in a high-multilayer PCB lamination test scenario, the number of ring piezoelectric transducer array units is 24, the input center frequency is set to 1.2 MHz, the bandwidth is set to ±150 kHz, and the pulse width is precisely set to... microseconds, repetition frequency is Hertz. For the acoustic matching layer of the semi-cured resin adhesive layer, after amplitude normalization of the actual emission power curve, the emitted sound pressure level of each array unit remains at [value missing]. Megapascals. After adopting the timing synchronization control algorithm, the phase error of the drive signal was measured and limited to within [a certain value]. Within this range, at this point, the adhesive layer thickness is... In the micrometer region, the amplitude ratio of the fundamental frequency to the second harmonic component of the transmitted sound wave signal remains stable at... The output signal transmitting unit exhibits significantly improved acoustic excitation stability and amplitude-frequency consistency during continuous pressing cycles, effectively ensuring the accuracy of subsequent acoustic propagation parameter acquisition. S1.3: Based on the configured high-frequency ultrasonic excitation signal, the multi-region transmitted sound waves during the pressing process are synchronously acquired and processed to obtain the original dataset of time-domain waveform, phase shift and envelope attenuation rate, so as to record the spatial change of the acoustic propagation characteristics of the resin layer in the early stage of heating, and output the original data matrix of transmitted sound waves containing regional coordinate information. A multi-channel synchronous acquisition algorithm (parameters: sampling rate 25MHz, number of channels = number of ring array units) is executed on the configured high-frequency ultrasonic excitation signal to achieve parallel capture of transmitted sound waves in multiple regions during the pressing process; Furthermore, by using a phase synchronization calibration method (parameters: reference signal phase zero point, time reference from a high-precision crystal oscillator), the time axes of different acquisition channels are aligned, and a raw time-domain waveform data stream with consistent phase across the entire region is obtained; Furthermore, an envelope detection algorithm (parameters: second-order Butterworth low-pass filter, cutoff frequency = 150kHz) is used to extract the envelope of the time-domain waveform, thereby realizing the quantitative acquisition of the acoustic wave amplitude change in the semi-cured resin medium and generating an initial value matrix of the envelope attenuation rate. Furthermore, based on the waveform peak localization algorithm (parameter: peak detection threshold = mean + 3 times standard deviation), the time localization of the main wave peak of each channel is achieved, and the sound wave propagation delay is calculated using the following formula:

[0012] in, Let i be the propagation delay of the i-th row and j-th column channel. Peak time, For reference only; Furthermore, by using the regional coordinate mapping method (parameter: converting the polar coordinate position of the array unit to Cartesian coordinates), the detected time-domain waveform, phase shift, and envelope attenuation rate data are associated with the spatial position of the pressing area, generating a raw data matrix of transmitted acoustic waves containing regional coordinate information; Through the above synchronous acquisition and processing algorithm, the high-frequency excitation signal of the previous step is transformed into a full-area original data matrix that simultaneously includes time-domain waveform, phase shift and envelope attenuation rate, so as to accurately record the spatial changes of the acoustic propagation characteristics of the resin layer in the early stage of heating. For example, in a high-multilayer PCB lamination test, a ring piezoelectric transducer array containing 64 units was used, with a sampling rate of 25MHz and a sampling window length of 1ms. After applying a pulse-modulated excitation signal, the waveform data acquired synchronously from multiple channels were phase-synchronized, and the phase offset of all channels was less than 0.5°. The initial attenuation rate obtained after envelope detection varied from 0.012 to 0.034 in different regions. The peak localization algorithm detected an average propagation delay of 42.7μs. After polar coordinate mapping to a rectangular coordinate grid, the spatial resolution of the original data matrix reached 0.8mm, and the matrix contained 64×64 data units, each recording the complete time-domain waveform, phase offset, and attenuation rate value. The results showed that this step could significantly improve the spatial perception accuracy in the lamination pressure distribution optimization process during the initial heating stage, and the output data matrix was successfully used for subsequent acoustic parametric feature extraction model training. S1.4: Perform time-domain feature extraction processing on the acquired raw transmitted sound wave data matrix, and use the cross-correlation algorithm to calculate the sound velocity, attenuation coefficient and nonlinear parameter B / A value to quantify the dynamic changes of colloidal shear modulus and loss modulus in each region, and output the sound propagation parameter feature vector of each spatial coordinate. S1.5: Based on the extracted sound propagation parameter feature vector, perform spatial distribution difference calculation processing. By calculating the spatial derivative of the sound velocity gradient and attenuation coefficient and mapping it to a 0.8 mm resolution coordinate grid, the spatial distribution difference of sound propagation parameters at different locations is characterized, and the spatial distribution difference matrix of sound propagation parameters is output.

[0013] Step S2: The transmitted acoustic wave data undergoes denoising and normalization preprocessing to eliminate background noise interference and improve the signal-to-noise ratio. Specifically, this includes: S2.1: Apply the discrete wavelet transform algorithm to the time-domain waveform data of transmitted sound waves to decompose multi-scale signal components and noise components; The input is the original transmitted acoustic wave data matrix containing multi-region coordinate information output from step S1.3. The data types include time-domain waveform, phase shift and envelope attenuation rate, which are derived from the acoustic propagation record of the resin adhesive layer under high-frequency ultrasonic excitation. The discrete wavelet transform algorithm (wavelet basis: Daubechies-8, decomposition level: 5) is used to realize the multi-scale decomposition of the acoustic wave time domain signal, which separates the original waveform signal into approximation components and detail components to distinguish the principal components of the signal from the noise components. Furthermore, by using a multi-scale threshold segmentation method (parameters: layer-dependent threshold calculation, scaling factor β=0.65), the noise-sensitive intervals of the coefficients of each decomposition layer are identified, and the coefficient matrix partitioning results of each layer are obtained. Furthermore, the noise coefficients are determined using a frequency domain energy calculation algorithm (formula below), and a noise component identification matrix is ​​generated:

[0014] in, For the first Layer coefficient energy, For the number of coefficients, For the first Layer coefficient value; Furthermore, by using a coefficient recombination method (excluding noise coefficients according to the noise identification matrix), the effective signal coefficient set and the noise coefficient set are separated, and a coefficient matrix for subsequent denoising processing is generated. By using multi-scale decomposition and noise identification of discrete wavelet transform, the result of the previous step is transformed into a coefficient matrix containing multi-scale signal components and noise components, thereby achieving the input preparation effect of subsequent soft threshold denoising. For example, during the lamination process of a multilayer PCB, the sampling rate of the transmitted acoustic wave signal is 2MHz, and the signal length is 2048 points. A 5-layer decomposition using the Daubechies-8 wavelet basis is performed, resulting in an approximation coefficient length of 64, and detail coefficient lengths of 64, 128, 256, 512, and 1024, respectively. The energy of each layer coefficient is calculated, such as the energy of the first layer detail coefficient. The value is 0.0025, significantly lower than the effective signal layer energy threshold of 0.005, therefore it is identified as a noise layer. The energy value of the third layer detail coefficients is 0.018, which is determined to be an effective signal layer. The noise layer coefficients are removed according to the noise identification matrix to form the coefficient matrix used for subsequent denoising. This matrix, applied to the reconstructed time-domain waveform after soft thresholding, improves the signal-to-noise ratio to 15dB in the signal-to-noise ratio test, significantly improving the accuracy of the acoustic parameters required for dynamic pressure redistribution calculation. S2.2: Based on wavelet decomposition coefficients, calculate an adaptive threshold and perform soft thresholding to remove noise correlation coefficients; S2.3: Perform inverse discrete wavelet transform on the denoised wavelet coefficients to reconstruct the noise-suppressed time-domain waveform data; S2.4: Calculate the arithmetic mean and sample standard deviation of the denoised time-domain waveform data to determine the normalization parameters; S2.5: Based on the normalization parameter, z-score normalization is performed on the denoised time-domain waveform data to output standardized acoustic wave data.

[0015] like Figure 2 As shown, step S3 involves extracting the temporal features of acoustic parameters based on the preprocessed acoustic wave data, including sound velocity, attenuation coefficient, and nonlinear parameters. Specifically, this includes: S3.1: Perform cross-correlation operation on the preprocessed acoustic time-domain waveform data to obtain the acoustic propagation delay in each region, and then calculate the sound velocity v and output the spatial distribution data of the sound velocity. The normalized acoustic time-domain waveform data output in step S2 is subjected to a cross-correlation operation method (parameters: window length 1024 points, sliding step size 8 points) to realize the delay matching calculation between the acoustic signal and the reference excitation signal in each region; Furthermore, the peak position of the delay is calculated by normalized cross-correlation function, the static offset is removed, and the propagation delay result matrix under each spatial coordinate is obtained; Furthermore, based on the propagation distance parameter, the sound speed is calculated (the formula is the transmission path length divided by the propagation delay) to achieve the physical quantization of the sound speed v in each region, as shown in the following formula:

[0016] in, The effective path length of the sound wave in the pressing region. The propagation delay corresponding to the peak cross-correlation value; Furthermore, spatial index mapping processing is performed through the multi-region sound velocity v result matrix to associate the sound velocity values ​​with the two-dimensional coordinate system of the pressing region and generate sound velocity spatial distribution data; By using cross-correlation calculations and spatial index mapping, the propagation delay data is transformed into a sound velocity spatial distribution technical indicator, enabling precise quantification of the sound velocity of the resin adhesive layer in the initial stage of heating, and providing a basic physical quantity for subsequent calculation of attenuation coefficient and nonlinear parameters. For example, in the initial stage of heating during the lamination process of high-multilayer PCBs, a 20-microsecond pulse-modulated ultrasonic wave is used as the reference signal, with a sampling frequency of 10MHz, a cross-correlation calculation window of 1024 points, a step size of 8 points, and a path length of [missing information]. The value is 12.8 mm. The peak cross-correlation value in a certain region appears at sampling point number 640, corresponding to a propagation delay of 12.8 mm. =64μs, substitute it into the formula The sound velocity v = 200 m / s was obtained. The above calculation was repeated in different regions to obtain the complete sound velocity distribution matrix. The verification results show that the sound velocity data varies significantly in space, which improves the dynamic judgment ability of the rheological properties of the inter-plate filler and provides high-precision input for the subsequent sound velocity-based attenuation coefficient calculation step (S3.2). S3.2: Based on the calculated sound velocity v, path length compensation is performed on the preprocessed envelope attenuation rate data to eliminate the influence of sound wave propagation geometry factors, and the attenuation coefficient is calculated. Output spatial distribution data of attenuation coefficient; S3.3: Apply bandpass filtering and Hilbert transform to the preprocessed acoustic time-domain waveform data to extract the fundamental frequency component and the second harmonic component. Calculate the nonlinear parameter B / A based on the component amplitude ratio and output the spatial distribution data of the nonlinear parameter. The preprocessed acoustic time-domain waveform data is processed using a bandpass filtering algorithm with a passband range (0.9MHz–1.5MHz). The filter type is set to FIR linear phase structure with an order of 128 to suppress low-frequency and high-frequency noise components while retaining the effective signal components of the fundamental frequency and second harmonic, thus achieving frequency domain preparation. Furthermore, the bandpass filtered waveform data is processed by Hilbert transform to generate an analytical signal and obtain its instantaneous amplitude envelope, so as to separate the phase and amplitude information of the fundamental frequency component and the harmonic component. Furthermore, using the amplitude data of the fundamental frequency component and the second harmonic component, the nonlinear coefficient of each acquisition area is calculated according to the definition formula of the nonlinear parameter B / A, wherein the calculation process adopts the following relationship:

[0017] in, The density of the resin adhesive layer, For the speed of sound, For the fundamental frequency amplitude, The amplitude of the second harmonic; Furthermore, for each spatial coordinate point, the calculated B / A value is bound to the region index to construct a nonlinear coefficient spatial matrix, so as to fully describe the spatial distribution of colloidal nonlinear sound propagation; The above algorithm is used to transform the results of bandpass filtering and Hilbert transform into nonlinear parameter spatial distribution data, thereby realizing the quantitative characterization of the microscale acoustic nonlinearity of the resin adhesive layer and providing input for subsequent pressure compensation mapping. For example, during the lamination process of high-multilayer PCBs, the acoustic time-domain waveform, after noise reduction and normalization, is acquired. A bandpass FIR filter (order 128, Hanning window weighted) with a passband of 0.95MHz–1.45MHz is used to filter out out-of-band noise, and then a Hilbert transform is performed to generate an analytical signal envelope. Based on the envelope signal, the fundamental frequency and second harmonic amplitude are separated, and the fundamental frequency amplitude in a certain region is measured. MPa, second harmonic amplitude MPa, speed of sound m / s, density g / cm³. Substitute the parameter into the formula:

[0018] The calculated B / A value for this region is approximately After being incorporated into the spatial matrix, a nonlinear parameter distribution map with a resolution of 0.8 mm is generated. This distribution map, in subsequent mapping, shows that the colloidal nonlinear changes are concentrated in specific areas of dense copper foil, verifying the necessity and accuracy of dynamic pressure compensation. S3.4: The speed of sound v and the attenuation coefficient... The nonlinear parameter B / A is integrated and processed according to the coordinates of the pressing region and the time series to generate the acoustic parameter temporal feature matrix and output the three-dimensional feature tensor. Based on the sound velocity v and attenuation coefficient output in steps S3.1, S3.2, and S3.3 respectively The nonlinear parameter B / A spatial distribution data are used, and a unified time index association algorithm is adopted (parameter settings: time step is 1.5 seconds, regional coordinate resolution is 0.8 mm) to achieve coordinate and temporal alignment of different acoustic parameters in the same pressing area within the same time step. Furthermore, through the matrix synchronization mapping method (parameter settings: rows represent the coordinate index of the pressing area, columns represent the time index, and matrix elements represent the corresponding acoustic parameter values), the spatial-temporal dual-dimensional binding of sound velocity, attenuation coefficient, and B / A value is realized, and preliminary combined matrix data is generated; Furthermore, by using a multi-parameter normalization fusion algorithm (parameter setting: performing minimum-maximum normalization on the three acoustic parameters in each time step, and then merging rows and columns), the numerical scale of acoustic parameters with different dimensions is unified, and a normalized fusion matrix is ​​obtained; Furthermore, a three-dimensional tensor construction method is adopted (structure: the first dimension corresponds to the coordinate index of the pressing area, the second dimension corresponds to the time index, and the third dimension includes the sound velocity v and the attenuation coefficient). The three indices (B / A, nonlinear parameters) are used to realize the mapping of the fusion matrix to the three-dimensional feature tensor and form a complete acoustic parameter temporal feature matrix. By using matrix linear mapping and tensor construction, the multi-parameter spatial distribution data from the previous step is transformed into a three-dimensional feature tensor with a clear spatial-temporal-parameter correspondence, thereby achieving a unified and structured expression of the acoustic state during the pressing process. For example, for a 12-layer high-multilayer PCB sample measuring 250 mm × 180 mm, the regional coordinate grid was set to a resolution of 0.8 mm to obtain 225 × 150 spatial grid points. The time step was 1.5 seconds, and the sound velocity v and attenuation coefficient were recorded during the lamination cycle. and the nonlinear parameter B / A. For the three-parameter data of a certain spatial point, the sound velocity v = 2800m / s, the attenuation coefficient... = 0.62 dB / mm, B / A value = 5.4, which are mapped to dimensionless values ​​of 0.74, 0.51, and 0.86 respectively during normalization fusion. A three-dimensional tensor is used to map the three-parameter values ​​of this point at each time step into the tensor structure. For example, at t = 3.0 seconds and region index (125, 75), the three-dimensional tensor element is... In the full tensor structure, the temporal features of this region can be continuously used as node inputs for subsequent graph neural network models. After verification, it significantly improves the matching accuracy of highly sensitive regions when predicting pressure compensation. S3.5: Apply moving average filtering to the generated acoustic parameter time series feature matrix to suppress instantaneous noise fluctuations and output smooth and stable time series feature data for subsequent graph neural network mapping model input.

[0019] like Figure 3 As shown, step S4 involves constructing a graph neural network mapping model to generate a pressure compensation matrix based on the correspondence between the temporal characteristics of acoustic parameters and the local colloid dynamic shear modulus and loss modulus in the offline calibration data. Specifically, this includes: S4.1: Based on the spatial topological relationship of the temporal features of acoustic parameters in offline calibration data, graph structure modeling is performed on the feature data to define the relationships between nodes and edges; multi-layer graph convolution operation is performed to extract local spatial features; global representation capability is enhanced through feature aggregation; attention mechanism is introduced to optimize the weight allocation of key regions; and the graph neural network model structure parameters are output as the basic framework for model training. Based on the smooth and stable acoustic parameter time-series characteristic data output in step S3, a graph structure modeling method is adopted (input: sound velocity v, attenuation coefficient). And nonlinear parameters B / A, coordinates of the pressing area corresponding to the node, and edge relationships based on spatial proximity and sound propagation correlation matrix) to realize the spatial topological expression of feature data; Furthermore, local spatial features are extracted through multi-layer graph convolution operations (parameters: 3 convolution layers, 64 filters, ReLU activation function), and a convolutional feature tensor containing the acoustic parameter variation patterns within the region is obtained. Furthermore, the overall representation capability is enhanced by a global feature aggregation method (parameters: a hybrid strategy of mean pooling and max pooling, with weight coefficients of 0.6 and 0.4), and a global feature vector across regions is generated to improve the model's ability to capture the correlation of colloidal rheological states in multiple regions. Furthermore, the key region weight allocation is optimized through a spatiotemporal attention mechanism (parameters: time window length 5 seconds, initial value of spatial attention weight is set to 0.5, dynamic update strategy is based on gradient normalization, and weight range is limited to 0.1 to 0.9), and an attention-weighted feature matrix is ​​generated to highlight the regional features that have a significant impact on stress compensation. By serializing the output, the graph neural network model structure parameters, including convolutional layer parameters, aggregation strategy parameters, and attention weight allocation mechanism, are output as the basic framework for model training, and the high-precision establishment of the pressure compensation quantity mapping relationship is achieved. For example, in an offline calibration scenario for a high-multilayer PCB lamination process, the input data consists of the sound velocity, attenuation coefficient, and B / A time-series features of the nonlinear parameters of 64 lamination region nodes. Each node establishes spatial topological edges to its four adjacent nodes, with edge weights selected based on a cross-correlation coefficient greater than 0.85. A 3-layer graph convolutional network is used, with 64 kernels per layer and ReLU activation function, reducing the size of the convolutional feature matrix from 64×3 to 64×16. In the global feature aggregation stage, a hybrid strategy of mean pooling and max pooling is used, with a mean pooling weight of 0.6 and a max pooling weight of 0.4. During the fusion process, the mean vector and maximum vector of the convolutional features in each region are linearly combined according to their weights to obtain the global feature vector. In the attention mechanism stage, a spatiotemporal attention model is adopted with a time window of 5 seconds. The spatial attention is initially set to 0.5 and dynamically updated through gradient normalization. The weight range is limited to 0.1 to 0.9. Regions whose decay coefficient changes more than a set threshold in the past 5 seconds are assigned an additional weight increment of 0.15, ultimately resulting in an attention-weighted feature matrix. The output graph neural network structure parameters include the convolution kernel weight matrix, pooling policy parameters, and attention weight update rules. This structure shows a significant reduction in the prediction error of stress compensation on the validation set, and the model stability and adaptability both meet expectations. S4.2: Based on the graph neural network model structure parameters, perform standardized preprocessing on the offline calibration dataset to eliminate differences in feature dimensions; perform data augmentation operations to balance the distribution of dynamic shear modulus and loss modulus; divide the preprocessed data into training and validation sets; store the divided datasets; and output standardized training and validation sets for model training and validation. Based on the graph neural network model structure parameters, a standardization preprocessing operation is performed on the offline calibration dataset, using the Z-score normalization method (parameter: the mean of each feature). and standard deviation Based on this, the dimensional differences in sound velocity, attenuation coefficient, and nonlinear parameter B / A on the numerical scale are eliminated; Furthermore, data augmentation methods (parameters: random rotation ±15° spatial topology mapping, temporal reversal, additive Gaussian noise) are employed. =0.002), to achieve a balance between the distribution of dynamic shear modulus and loss modulus, so that the sample proportion of various colloidal mechanical states tends to be balanced during model training, and an enhanced multi-domain feature dataset is obtained. Furthermore, by using a dataset partitioning algorithm (parameters: 70% for training set and 30% for validation set, spatial partitioning based on the coordinate index of the compression region), the functional separation of the training and validation purposes of the enhanced standardized dataset is achieved, and training set feature matrices and validation set feature matrices are generated. Furthermore, through the storage management module (parameters: binary serialization format, compression ratio 0.85, storage medium is high-temperature industrial SSD), efficient storage and fast reading of the partitioned dataset are achieved, and the data call latency in the subsequent model training and validation stages is less than 5ms. Through the above standardization and data augmentation processing methods, the temporal features of acoustic parameters in the previous step are transformed into a standardized training set and validation set with uniform feature scale, balanced sample distribution, and format that can be directly input into the graph neural network, thereby improving the numerical stability and generalization performance of the model training stage. For example, the speed of sound acquired during the offline calibration phase attenuation coefficient and nonlinear parameters The time-series feature dataset uses the mean with standard deviation Calculation formula:

[0020] in, These are the original eigenvalues. The arithmetic mean of the feature across all samples. This represents the sample standard deviation for the corresponding feature. This standardization formula is applied to all feature vectors to unify the data scale to a standard normal distribution with a mean of zero and a variance of one. During data augmentation, a rotation matrix transformation is performed on the spatial topology, with the rotation angle determined by a random number generator within the interval... Internal selection ensures the spatial diversity of the enhanced feature matrix. A reversal operation is performed on the time series to enable the model to learn the mapping features of the reverse colloidal rheological state. When partitioning the dataset, the coordinates of the compression region are used as the index to ensure consistent spatial coverage between the training and validation sets. After storage, the training set contains 10,000 samples, and the validation set contains 4,500 samples. Reading performance remains stable in industrial field pressure compensation model deployment, and the output results meet the numerical stability and performance improvement goals of the model training phase. S4.3: Perform batch input processing on the standardized training set to adapt it for graph neural network training; perform forward propagation to generate predicted values ​​of dynamic shear modulus and loss modulus; calculate the mean squared error loss between the predicted values ​​and the calibrated actual values; use the stochastic gradient descent algorithm to perform backpropagation to optimize the network weights; iteratively update the model parameters until the loss converges to obtain the trained graph neural network model. Batch input processing is performed on the standardized training set (batch size parameter: 64, random sampling to ensure uniform sample distribution) to adapt the training data to the input structure of the graph neural network. Furthermore, dynamic shearing modulus is achieved through forward propagation computation (network structure parameters: 3 graph convolutional layers, 256 nodes per layer, ReLU activation function, 8 attention layer heads). With loss modulus The predicted values ​​are generated, and the predicted output matrix is ​​obtained; Furthermore, the difference between the predicted value and the actual value calibrated offline is quantified by calculating the mean squared error loss function. The formula is as follows:

[0021] in, This is the mean squared error loss value. For the sample size, To calibrate the actual modulus value, To predict the modulus value; Furthermore, the network weights are optimized based on the gradient of the loss function by using the stochastic gradient descent algorithm (learning rate parameter: 0.001, momentum coefficient: 0.9), and an iteratively updated weight matrix is ​​generated. Furthermore, loss convergence detection is performed (convergence threshold parameter: This allows for the determination of the training loop termination condition and the acquisition of the completed graph neural network mapping model. Through the above training chain, batch input data is mapped to the prediction results of dynamic shear modulus and loss modulus, and transformed into stable, convergent and highly adaptable model weights, achieving a high-precision mapping effect for the generation of the pressure compensation matrix. For example, the standardized feature matrix (size: 1024×3) of the sound velocity, attenuation coefficient, and nonlinear parameters contained in the offline calibration dataset is input into a graph neural network model with a batch size of 64. The model is a three-layer graph convolutional network with 256 nodes per layer, using ReLU activation function and a multi-head attention mechanism with 8 heads. During the forward propagation phase, the model outputs the dynamic shear modulus. With loss modulus Predicted value matrix (size: 1024×2). In the loss calculation stage, the mean squared error formula is used:

[0022] in =1024. During the stochastic gradient descent update phase, the learning rate is set to 0.001, the momentum coefficient to 0.9, and the weights are iteratively updated until the loss value stabilizes at 1024. Within, and the convergence threshold is The final output model shows a significant reduction in the error between the predicted modulus and the actual modulus on the validation set, and has the ability to generate a pressure compensation matrix in real time during the pressing process, thus enhancing the stability and robustness of the filler uniformity optimization. S4.4: Based on the trained graph neural network model, perform prediction processing on the validation set data; calculate the absolute errors between the predicted dynamic shear modulus and loss modulus and the actual values; evaluate the model accuracy metrics; determine whether the accuracy meets the preset threshold; output the model performance evaluation report to confirm the model effectiveness; The inputs are the trained graph neural network model, the feature matrix of the validation set obtained by partitioning in step S4.2, and the corresponding calibrated actual values. Using the model prediction processing algorithm (parameters: batch size 64, prediction step 1), the dynamic shear modulus and loss modulus of each node in the validation set feature matrix are calculated, and the prediction result matrix is output; Furthermore, through the absolute error calculation method (parameters: element-by-element processing, error unit MPa), the deviation between the predicted value and the calibrated actual value is quantified, and the absolute error matrix is obtained, where the dynamic shear modulus error matrix and the loss modulus error matrix are stored separately; Furthermore, through the model accuracy evaluation algorithm (parameters: double-index calculation of mean square error MSE and mean absolute error MAE), the overall prediction performance of the model is quantitatively evaluated, and an accuracy metric set is generated; Furthermore, through the threshold determination method (parameters: preset accuracy threshold, MSE ≤ 0.002 and MAE ≤ 0.05), the qualification of the accuracy metric set is determined, and a determination result flag is generated, whose value is qualified or unqualified; Through the performance evaluation report generation algorithm (parameters: output in JSON format, including model structure information, prediction performance metrics, accuracy determination results), the determination result of the previous step and the model prediction performance data are combined into performance evaluation report data to confirm the model effectiveness; Exemplarily, in the off-line calibration verification of a certain high-rise and multi-layer PCB lamination process, the input validation set contains the acoustic parameter characteristics of 128 regional nodes, the feature dimension is 9, where the actual value range of the dynamic shear modulus is between 0.82 MPa and 1.25 MPa, and the actual value range of the loss modulus is between 0.31 MPa and 0.57 MPa. Using the prediction processing with a batch size of 64, the dynamic shear modulus prediction value matrix and the loss modulus prediction value matrix are output. Through absolute error calculation, the dynamic shear modulus error range is from 0.004 MPa to 0.041 MPa, the loss modulus error range is from 0.005 MPa to 0.039 MPa, the MSE is calculated to be 0.0018 by the formula and the MAE is calculated to be 0.032, both of which are lower than the preset accuracy threshold, and the determination result is qualified. The performance evaluation report clarifies the stability and high-precision characteristics of the model on the full validation set, supporting its ability to generate real-time pressure compensation amounts during the deployment and operation stage.

[0023] S4.5: Serialize the validated graph neural network model to generate a deployable mapping model file; store the model file in the system memory; output the pressure compensation matrix to generate the model; use it to calculate the pressure compensation amount in real time during the online stage; ensure the model's real-time application capability during the pressing process.

[0024] Step S5: Input the real-time extracted acoustic parameter temporal features into the graph neural network mapping model to obtain the current pressure compensation matrix. Specifically, this includes: S5.1: Based on the acoustic parameter time-series feature vector output in step S3, obtain the normalized sound velocity, attenuation coefficient and nonlinear parameter standardized data, and perform feature dimension alignment operation on the data to generate a feature tensor that meets the input dimension requirements of the graph neural network. Based on the acoustic parameter time-series feature vector output in step S3, normalized sound velocity, attenuation coefficient and nonlinear parameter standardized data are obtained, and feature dimension alignment operation is performed on the data to generate feature tensors that meet the input dimension requirements of graph neural networks. A feature normalization algorithm (parameters: mean μ, standard deviation σ calculated based on S3 output data) is used to normalize the sound velocity v and attenuation coefficient. And the numerical range of the nonlinear parameter B / A is unified to avoid different physical quantities affecting the weight allocation in the graph neural network model; Furthermore, Z-score normalization is performed using the following formula:

[0025] in, To standardize the eigenvalues, These are the original eigenvalues. The sample mean. The standard deviation is the sample standard deviation. By using the feature vectorization processing method (parameter: feature type index table), the sound velocity, attenuation coefficient, and nonlinear parameters are sorted according to the coordinates of the compression region and formed into a continuously stored multidimensional array structure, and a dimensional structure mapping table is obtained. Furthermore, a dimension alignment algorithm is employed (parameter: the target input dimension is determined during the training phase of the graph neural network). The feature array is interpolated or truncated to make the feature vector lengths of each region consistent, thus eliminating dimensional differences caused by different sampling frequencies or region sizes. Furthermore, a multidimensional tensor construction algorithm (parameter: time priority, space priority) is used to sequentially stack the aligned feature arrays on the time axis, space axis, and feature axis to generate a three-dimensional tensor. Its structure satisfies ,in For time steps, This represents the number of nodes in the space. By using feature encoding consistency processing, the normalization and alignment results of the previous step are transformed into tensor form, realizing the structural consistency of input features in spatial topology and time series, and ensuring that stable mapping effect is obtained when entering graph neural network. For example, in a high-multilayer PCB lamination process, the sound velocity output by S3 ranges from 1500 to 1580 m / s, the attenuation coefficient ranges from 0.12 to 0.18 dB / mm, and the nonlinear parameter B / A ranges from 6 to 8. The mean value is calculated based on the sound velocity characteristics. Standard deviation Substituting into the formula, we obtain the normalized value of the speed of sound in a certain region. Regarding the mean attenuation coefficient Standard deviation After normalization, the feature value of a certain node is For nonlinear parameters, the mean Standard deviation The normalization result is The three types of features from each region at the same time step are arranged according to an index table to form a feature array of length 3M. However, for some regions, insufficient sampling points may prevent this from being achieved. When the input requirement is 64, linear interpolation is used to pad to the specified length. Under the stacking rule of time priority followed by space priority, the three types of features of M nodes at each time step are tensorized to construct a shape of... The input tensor is ultimately ensured to perfectly match the input interface of the graph neural network pre-training stage, thereby achieving stable hidden feature extraction and subsequent pressure compensation inference. S5.2: Input the feature tensor into the graph neural network mapping model constructed in step S4, perform node feature aggregation operation based on pre-trained weights, and use graph convolutional layers to perform nonlinear transformation on the spatial topological relationship between regions to generate the hidden layer state matrix. S5.3: Based on the hidden layer state matrix, perform regional importance weighting calculation through a spatiotemporal attention mechanism, apply dynamic weight coefficients to high-influence regions, and generate a preliminary pressure compensation vector; Based on the hidden layer state matrix generated in step S5.2, a spatiotemporal attention mechanism (parameters: 8 multi-heads, 5-frame time window, and 2-node spatial neighborhood radius) is used to realize the weighted calculation of regional importance. Furthermore, through the temporal attention allocation unit (algorithm: Scaled Dot-Product Attention, scaling factor) (Input dimension is 64) Calculate the similarity weight between the state vector at each time step and the global time reference vector, and generate a time weighted coefficient matrix to enhance the time period features where the dynamic shear modulus changes significantly. Furthermore, the spatial domain attention allocation unit (algorithm: Graph Attention Network, adjacency matrix based on the topology of the compression region, and softmax for weight normalization) calculates the feature coupling weights of each node's state vector and its neighboring node's state vectors, and generates a spatial weighted coefficient matrix to highlight the feature expression of the pressure distribution sensitive area. Furthermore, the time-weighted coefficient matrix and the spatial-weighted coefficient matrix are fused through element-wise multiplication to obtain the spatiotemporal integrated weighted coefficient matrix. The hidden layer state matrix is ​​then weighted and modulated according to the node index and time index to enhance the output amplitude of high-influence region features. The modulated spatiotemporal features are mapped to a preliminary pressure compensation vector through a fully connected transform layer (parameters: input dimension 64, output dimension 1, activation function ReLU), thus realizing the mapping from the hidden layer to the compensation domain. By using a spatiotemporal attention mechanism, the hidden layer state matrix from the previous step is transformed into a preliminary pressure compensation vector containing regional importance information, thereby achieving an optimized pressure distribution effect after applying dynamic weights to high-influence regions. For example, in the high-multilayer PCB lamination process, the hidden layer state matrix size is 128×64 (128 region nodes, 64-dimensional features), the time window length is set to 5 frames, and the total time step is 20 frames. In the temporal attention allocation unit, the scaling factor is set to... The process calculates the inner product of the state vector of each node within a 5-frame window and the global temporal reference vector, and generates weighting coefficients using softmax. For example, the temporal weight of a node in frame 7 is 0.182. In the spatial domain attention allocation unit, the neighborhood radius is 2 nodes. The feature similarity (using cosine similarity) between the state of a node and its neighboring nodes is calculated, normalized, and then a spatial weight is generated. For example, the spatial weight of node 15 is 0.239. The temporal weight and spatial weight are multiplied element-wise to obtain a comprehensive weight coefficient matrix. For example, the comprehensive weight of node 15 in frame 7 is 0.043. The hidden layer state matrix is ​​modulated according to this comprehensive weight and input to the fully connected layer to output a scalar compensation value. For example, the initial pressure compensation value of node 15 in frame 7 is 32.6N. After subsequent S5.4 sub-pixel level spatial interpolation processing, this initial pressure compensation vector can significantly improve the pressure compensation accuracy for high-risk areas of local glue deficiency and improve the overall glue filling uniformity. S5.4: Perform sub-pixel level spatial interpolation on the preliminary pressure compensation vector, and perform mesh mapping operation based on the physical distribution topology of the piezoelectric actuator to generate a pressure compensation matrix with a spatial resolution of 0.8 mm. S5.5: Compare the pressure compensation matrix with the process safety threshold for boundary verification, and perform dynamic range trimming based on thermodynamic state feedback to output a calibrated pressure compensation matrix that conforms to the performance capability of the pressing equipment.

[0026] Step S6: The pressure compensation matrix is ​​input into a distributed robust predictive controller, which generates a drive signal for the flexible piezoelectric actuator based on the thermodynamic state of the pressing process. Specifically, this includes: S6.1: Based on the thermodynamic model parameters and real-time state feedback data of the pressing process, a distributed robust predictive controller instance is constructed. This controller adopts the model predictive control algorithm and integrates the H∞ robust optimization mechanism to generate a control strategy output with the ability to resist parameter perturbations and external disturbances. S6.2: Input the pressure compensation matrix output from S5 into the distributed robust predictive controller as a reference trajectory signal to initialize the controller's internal state estimation module and update the target pressure distribution in the prediction time domain; The inputs are the calibration pressure compensation matrix that conforms to the performance capability of the pressing equipment, which is output from step S5, and the already constructed distributed robust predictive controller instance. A matrix data injection method (parameters: matrix dimension 0.8mm resolution grid, frame update period 1.5 seconds) is used to load the calibration pressure compensation matrix into the controller's input buffer and bind it as a reference trajectory signal. Furthermore, the initial values ​​of the internal state estimation module are set through the state estimation initialization algorithm (parameters: the set of internal state variables of the model predictive control, including the temperature field state vector, the pressure field state vector, and the inertial parameters of the execution unit), and the prior state estimation data of the whole domain is obtained. Furthermore, a prediction time-domain target generation method is adopted (parameters: prediction time-domain length = 5 steps, sampling interval = 0.5 seconds) to map the reference trajectory signal to the prediction time-domain pressure distribution target matrix and generate a three-dimensional pressure distribution tensor containing target values ​​in multiple regions; Furthermore, a thermodynamic state consistency test algorithm (parameter: state deviation tolerance ±2℃ or ±15N) is used to evaluate the deviation between the predicted target matrix and the current state feedback matrix and generate a deviation correction coefficient matrix. By injecting reference trajectory signals, initializing the state, and updating the predicted target, the pressure compensation matrix of the previous step is transformed into target pressure distribution data that the controller can use, thus realizing the preparation for future pressure optimization calculation based on the current thermodynamic state constraints. For example, in the high-multilayer PCB lamination process, the input pressure compensation matrix is ​​a 100×120 grid with a unit pressure range of 0–120N and a frame update cycle of 1.5 seconds. After the matrix is ​​loaded into the controller input buffer, the state estimation initialization algorithm uses an average temperature field of 180℃ and an average pressure field of 350psi as the initial state, and the inertia parameter is set to the execution unit response time of 3.2ms. The prediction time domain length is set to 5 steps, the sampling interval is 0.5 seconds, and the generated target tensor contains 5 time domain slices, each slice corresponding to the global target pressure distribution from 0.5 seconds to 2.5 seconds in the future. In the thermodynamic state consistency check, the deviation tolerance is set to ±2℃ and ±15N. The deviation correction coefficient matrix is ​​calculated, and the target value in the prediction tensor is corrected by matrix multiplication to significantly reduce the error between it and the current state feedback matrix within the limit. After this step, the predictive controller can perform optimization calculations with a reliable initial state and target trajectory in the rolling time domain, ensuring the accuracy and stability of subsequent drive signal generation. S6.3: Based on the real-time thermodynamic state data of the pressing process, including the temperature distribution matrix and the pressure distribution matrix, execute the rolling time-domain optimization algorithm of the distributed robust predictive controller to calculate the future control sequence, so as to minimize the pressure distribution error and meet the physical constraints of the flexible piezoelectric actuator. The inputs include the temperature distribution matrix and pressure distribution matrix obtained in real time during the pressing process, and the execution target is the rolling time-domain optimization module of the distributed robust predictive controller; A rolling time-domain optimization algorithm (parameters: prediction time domain length 12 steps, sampling period 1.5 seconds) is adopted to jointly construct a state prediction model by combining the current thermodynamic state data with the target pressure compensation reference trajectory; Furthermore, by using the finite-time state prediction equation and control sequence solver (constraints: flexible piezoelectric actuator output force range 0–120N, stroke 4.5μm), dynamic simulation of the pressure distribution in the future time domain is achieved, and the predicted pressure matrix is ​​obtained; Furthermore, by employing an error minimization objective function, the sum of squared differences between the predicted pressure matrix and the reference pressure matrix is ​​calculated to generate a quantitative index of pressure distribution error. The formula is:

[0027] in, Represents the reference pressure matrix. Represents the predicted stress matrix. Indicates the number of prediction steps; Furthermore, by introducing H∞ robust optimization constraints and using a weight matrix to adjust the sensitivity of state prediction and control inputs, an optimization objective function resistant to external disturbances and material parameter perturbations is generated. Furthermore, the future control sequence that satisfies the objective function minimization without violating physical constraints is iteratively calculated by the solver, generating a control sequence matrix; By using a rolling time-domain optimization process, the thermodynamic state mapping of the previous step is combined with the reference trajectory to transform it into executable future control sequence data, thereby achieving the expected technical effect of reducing the compression pressure distribution error and synergistically satisfying the action constraints of the execution unit. For example, in a high-multilayer PCB lamination device with an upper operating temperature limit of 200℃ and a lamination cycle of no more than 1600 seconds, the prediction time domain length is set to 12 steps, the sampling period is 1.5 seconds, and the output force of the flexible piezoelectric actuator is set to a dynamic adjustment range of 40N–100N. The real-time temperature distribution matrix is ​​120×120 nodes, and the pressure distribution matrix is ​​of the same size. The target pressure compensation matrix is ​​generated by step S5. The temperature matrix and pressure matrix are input into the state prediction module to predict the pressure distribution for the next 18 seconds. The pressure distribution error is calculated, and the objective function weights are set to 0.35 for the temperature distribution error and 0.65 for the pressure distribution error, using the H∞-constrained gain parameter. =0.8. In the formula

[0028] middle, Taking 12 significantly reduces the sum of squared errors. The future control sequence calculated by the rolling time-domain solver reduces the pressure error in high-influence areas to a physically controllable range, smooths the action of the execution unit, and ultimately achieves a significant improvement in the uniformity of inter-board glue filling without extending the cycle time. S6.4: Based on the control sequence obtained by optimization calculation, a preliminary drive signal matrix is ​​generated. This matrix is ​​converted into sub-region pressure compensation commands through a spatial mapping algorithm to realize the preliminary drive signal allocation to the flexible piezoelectric actuator array. S6.5: Perform signal conditioning and amplitude limiting on the initial drive signal matrix to adapt to the dynamic response characteristics of the flexible piezoelectric actuator, and output the final drive signal to the actuator to ensure accurate execution of dynamic pressure redistribution and optimization of glue filling uniformity; An adaptive signal conditioning algorithm (parameters: conditioning filter type is finite impulse response (FIR), filter order is 32) is used on the initial drive signal matrix to achieve precise limitation of signal bandwidth and amplitude smoothing, so as to match the electrical input characteristics of the flexible piezoelectric actuator. Furthermore, by using a dynamic gain adjustment method (parameters: gain adjustment period is 5ms, target gain curve is set according to the force output-voltage mapping model of the execution unit), the amplitude of the drive signal and the mechanical response of the execution unit are synchronized in real time, and the signal matrix after gain matching is obtained. Furthermore, by using an amplitude limiting processing method (parameters: maximum allowable amplitude is 24V peak voltage, minimum allowable amplitude is 1V), based on the safe operating area of ​​the execution unit under high frequency response, element-level amplitude clipping is performed to ensure that the signal amplitude of each drive channel does not exceed the safe boundary, and an amplitude-limited signal matrix is ​​generated. Furthermore, a phase offset compensation is performed on the amplitude-limited signal matrix using a phase correction algorithm (parameter: the reference phase sequence is generated by the thermodynamic state prediction module), thereby achieving timing consistency of multi-region execution units and generating a phase-corrected signal matrix. By using a bidirectional digital-to-analog conversion (parameters: 16-bit sampling resolution, 200kHz sampling frequency) process, the phase correction result of the previous step is converted into digital drive instructions for the execution unit, thereby achieving precise execution of dynamic pressure redistribution and optimization of filling uniformity. For example, in the lamination process of a multilayer PCB, the input condition is the initial drive signal matrix generated by S6.4, with an amplitude between 0 and 30V and a spectrum mainly concentrated between 0 and 50kHz. A 32nd-order FIR filter is used for signal conditioning, with the filter bandwidth set to 45kHz to ensure that high-frequency components of the signal do not cause power overload of the actuator. In dynamic gain adjustment, according to the force-voltage mapping of the actuator, the gain factor is updated every 5ms to ensure that the force output curve is consistent with the predicted pressure distribution. Amplitude limiting processing cuts the peak voltage to 24V, which is less than the actuator's rated value of 25V, to avoid dielectric fatigue caused by long-term high pressure. In phase correction, based on the reference phase sequence predicted by the thermodynamic state, ±2° phase compensation is performed on each drive channel to ensure the synchronization of pressure application in multiple regions. Finally, the signal output by the 16-bit 200kHz bidirectional digital-to-analog converter forms a dynamic pressure distribution with a spatial resolution of 0.8mm at the actuator end. Verification results show that the pressure uniformity remains stable throughout the entire lamination cycle, significantly improving the consistency of glue filling.

[0029] Step S7: Based on the driving signal, control the flexible piezoelectric actuator to apply dynamic pressure distribution to optimize the uniformity of adhesive filling between boards. Specifically, this includes: S7.1: The drive signal is parsed and processed to obtain the specific control command parameters for the pressure compensation amount of each region; S7.2: Based on the control command parameters, calculate the driving voltage sequence required for the micro piezoelectric thin film stack in the flexible piezoelectric actuator; S7.3: The driving voltage sequence is applied to the micro piezoelectric thin film stack to generate a dynamic pressure distribution with a spatial resolution of 0.8 mm; Based on the required driving voltage sequence of the micro piezoelectric thin film stack calculated in step S7.2, it is used as an input signal to be applied to the corresponding piezoelectric stack unit to ensure that each unit receives a voltage command that is precisely matched to the spatial coordinates; A synchronous drive signal allocation method (parameter: drive sequence synchronization deviation less than 0.5μs) is adopted to achieve concurrent loading of voltage sequence on all piezoelectric thin film stacks and maintain timing consistency to prevent time delay of spatial pressure distribution; Furthermore, using the calibration coefficient k (unit: N / V) of the voltage-to-force conversion function, the output force of each stack is calculated using the following formula:

[0030] in, To output force, For calibration coefficients, To apply voltage, ensure that the force value accuracy error does not exceed ±1N; Furthermore, the forces of each stack are distributed to their corresponding physical locations using a spatial mapping matrix (0.8 mm resolution), forming a continuous two-dimensional pressure field; A closed-loop displacement monitoring algorithm (parameter: feedback sampling period 1ms) is adopted to measure the stack travel in real time and compare it with the target travel of 4.5μm. The pressure output error is calculated and real-time compensation is performed to stabilize the dynamic pressure distribution. By applying the voltage to the chain, the driving sequence of the previous step is transformed into a dynamic pressure distribution with a spatial resolution of 0.8 mm, thereby achieving the goal of uniform filling of the inter-board area. For example, in the lamination process of a 16-layer PCB with dimensions of 145mm × 200mm, a micro piezoelectric film stack array is arranged with a total of 256 units, and the calibration coefficient is configured as follows: N / V, the drive voltage range for each unit is set at... arrive Between V. In actual execution, the voltage sequence synchronization deviation is controlled within Within μs, the output force value of each unit is within arrive Fluctuating within the range of N, the travel distance remains at 4.5 μm ± 0.02 μm. After spatial mapping matrix transformation, the pressure difference between adjacent nodes in the generated two-dimensional pressure field at a 0.8 mm resolution grid does not exceed [a certain value]. N showed a significant improvement in the uniformity of glue filling between boards, and subsequent testing confirmed that there were no local glue shortages or overflows. S7.4: The dynamic pressure distribution is flexibly coupled to the surface of the mirror steel plate through a micro-bump array to achieve uniform pressure transmission; S7.5: Verify the uniformity of pressure distribution coupled to the surface of the mirror steel plate to output the state of achievement of the target for glue filling uniformity optimization.

[0031] Step S8: Monitor changes in acoustic propagation parameters during the pressing process. If parameter fluctuations exceed a threshold, trigger adaptive adjustment of the pressure compensation parameters. Specifically, this includes: S8.1: Based on the temporal characteristic distribution of acoustic parameters in the offline calibration data, perform multi-region statistical fluctuation range analysis to generate acoustic propagation parameter fluctuation thresholds; where the inputs are the sound velocity v and attenuation coefficient obtained in the offline calibration stage. The time series dataset of nonlinear parameter B / A is used to calculate the confidence interval of parameter fluctuation in each region using the sliding standard deviation algorithm. The output is a preset acoustic propagation parameter fluctuation threshold matrix, which serves as the benchmark for subsequent fluctuation judgment. S8.2: Continuously monitor the time-domain waveform data of transmitted acoustic waves acquired in real time during the pressing process to obtain the temporal variation sequence of acoustic propagation parameters; the input is the multi-region transmitted acoustic wave time-domain waveform, phase shift, and envelope attenuation rate data stream output from step S1. High-pass filtering and sampling synchronization techniques are used to eliminate periodic interference. The output is the sound velocity v and the attenuation coefficient. The real-time parameter time series change sequence of the nonlinear parameter B / A is used as the raw input for fluctuation calculation; S8.3: Based on the time-series variation sequence of acoustic propagation parameters, perform sliding window standard deviation calculation to generate parameter fluctuation values; where the input is the time-series variation sequence of parameters output by S8.2, the local variance of the sequence is calculated using a time-domain statistical method with a fixed window length of 50ms, and the output is a matrix of acoustic propagation parameter fluctuation values ​​for each region, which serves as a quantitative indicator for threshold comparison; S8.4: Compare the parameter fluctuation values ​​with the acoustic propagation parameter fluctuation thresholds point by point to determine the fluctuation exceeding the limit state; where the input is the parameter fluctuation value matrix output by S8.3 and the acoustic propagation parameter fluctuation threshold matrix generated by S8.1, Boolean logic operation is used to perform element-level threshold determination, and the output is a binary fluctuation exceeding the limit flag matrix, which serves as the trigger condition for adaptive adjustment; S8.5: If there is an over-limit flag in the fluctuation over-limit flag matrix, generate an adaptive adjustment trigger signal for the pressure compensation parameter to start the parameter update of the distributed robust predictive controller; wherein, the input is the fluctuation over-limit flag matrix output by S8.4, the edge triggering mechanism is used to generate pulse adjustment commands, and the output is the adaptive adjustment trigger signal for the pressure compensation parameter, which serves as the dynamic input source for the distributed robust predictive controller in step S6. The input is the binary fluctuation over-limit flag matrix output in step S8.4. An edge-triggered mechanism is used to perform element-by-element scanning of the matrix to detect and locate state transition points in the matrix. Furthermore, by using XOR logic operations (parameters: the current sampled state matrix and the previous sampled state matrix), single-point determination of state transitions is achieved, and an array of state change position indices is obtained. Furthermore, a pulse signal generation algorithm (parameters: single-point judgment index array, pulse width 3ms, amplitude normalization to 1.0) is used to construct the pulse-type adjustment command and generate the command sequence matrix. Furthermore, through timing coding processing (parameters: pulse sequence matrix, coding period 50ms, synchronization signal phase offset 0.25π), the mapping of the instruction sequence matrix to the pressing process control time reference is realized, and a time synchronization pulse matrix is ​​generated; By processing signal buffering and synchronous release (parameters: buffer depth 5 frames, release trigger delay ≤ 2ms), the time synchronization pulse matrix is ​​converted into a pressure compensation parameter adaptive adjustment trigger signal, thereby realizing the dynamic input supply to the distributed robust predictive controller in step S6. For example, the input is a binary fluctuation over-limit flag matrix containing 128×128 elements, where each element represents whether the corresponding pressing area is in a fluctuation over-limit state during the current sampling period. Using an XOR logic operation to compare the current sample with the previous sample matrix, 78 elements are detected to have transitioned from state 0 to 1. Substituting the indices of these elements into the pulse signal generation formula:

[0032] in, For the generated pulse signal, and For matrix row and column indices, Set the pulse amplitude to 1.0. The Dirac delta function represents the time delay. Based on this formula, pulse signals with an amplitude of 1.0 and a width of 3ms are generated for 78 transition positions. These signals are then mapped to the pressing process time base using a 50ms period timing code, resulting in a time-synchronized pulse matrix containing 78 valid pulses. A buffer depth of 5 frames ensures a release trigger delay ≤2ms, achieving accurate output of the trigger signal for adaptive adjustment of pressure compensation parameters. This trigger signal, input to the distributed robust predictive controller in step S6, immediately triggers parameter updates, ensuring dynamic compensation response and optimized filling uniformity for multi-region pressure distribution during pressing.

[0033] 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.

[0034] 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 processing method for high-multilayer boards with uniform inter-board glue filling, characterized in that, Includes the following steps: S1: Collect raw data of transmitted sound waves from multiple regions during the pressing process to obtain the spatial distribution differences of sound propagation parameters at different locations; S2: Perform noise reduction and normalization preprocessing on the original transmitted sound wave data, and output the preprocessed transmitted sound wave data. S3: Based on the preprocessed transmitted acoustic wave data, extract the temporal features of acoustic parameters; S4: Based on the correspondence between the temporal characteristics of acoustic parameters and the local dynamic shear modulus and loss modulus of colloids in the offline calibration data, a graph neural network mapping model is constructed. S5: Input the real-time extracted acoustic parameter time-series features into the graph neural network mapping model to obtain the current pressure compensation matrix; S6: Input the current pressure compensation matrix into the distributed robust predictive controller, which generates the drive signal of the flexible piezoelectric actuator based on the thermodynamic state of the pressing process. S7: Based on the driving signal, control the flexible piezoelectric actuator to apply dynamic pressure distribution to optimize the uniformity of adhesive filling between boards; S8: Monitor changes in acoustic propagation parameters during the pressing process. If parameter fluctuations exceed the acoustic propagation parameter fluctuation threshold, trigger adaptive adjustment of the pressure compensation parameters.

2. The method for processing high-multilayer boards with uniform inter-board glue filling according to claim 1, characterized in that, The raw data of the transmitted sound wave includes the time-domain waveform data, phase offset data, and envelope attenuation rate data of the transmitted sound wave.

3. The method for processing high-multilayer boards with uniform inter-board adhesive filling according to claim 1, characterized in that, Step S3 specifically includes: Perform cross-correlation operation on the preprocessed transmitted sound wave data output in step S2 to obtain the sound wave propagation delay in each region, and then calculate the sound speed. Based on the speed of sound, the preprocessed envelope attenuation rate data is subjected to path length compensation processing to calculate the attenuation coefficient. Bandpass filtering and Hilbert transform are applied to the preprocessed acoustic time-domain waveform data to extract the fundamental frequency component and the second harmonic component, and nonlinear parameters are calculated based on the component amplitude ratio. The sound velocity, the attenuation coefficient, and the nonlinear parameter are integrated according to the coordinates of the compression region and the time series to generate a temporal feature matrix of acoustic parameters; The acoustic parameter time-series feature matrix is ​​subjected to moving average filtering to output smooth and stable time-series feature data.

4. The method for processing high-multilayer boards with uniform inter-board adhesive filling according to claim 3, characterized in that, The passband range of the bandpass filter is 0.9–1.5MHz, the filter type is FIR linear phase structure, and the order is 128.

5. The method for processing high-multilayer boards with uniform inter-board adhesive filling according to claim 1, characterized in that, Step S4 specifically includes: Based on the spatial topological relationship of the temporal features of acoustic parameters in offline calibration data, graph structure modeling is performed on the temporal feature data to output the structural parameters of the graph neural network model. Based on the structural parameters of the graph neural network model, the offline calibration dataset is standardized and preprocessed, and data augmentation is performed to balance the distribution of dynamic shear modulus and loss modulus. The augmented standardized dataset is then divided into a standardized training set and a validation set. Batch input processing is performed on the standardized training set, forward propagation is performed to generate predicted values ​​of dynamic shear modulus and loss modulus, the mean square error loss between the predicted values ​​of dynamic shear modulus and loss modulus and the actual values ​​is calculated, and backpropagation is performed to optimize the network weights. The model parameters are iteratively updated until the loss converges, and the trained graph neural network model is obtained. Based on the trained graph neural network model, the validation set data is processed for prediction, the absolute error between the predicted dynamic shear modulus and loss modulus and the actual values ​​is calculated, the model accuracy index is evaluated, it is determined whether the accuracy meets the preset threshold, and a model performance evaluation report is output. The validated graph neural network model is serialized to generate a mapping model file, which is then stored in the system memory.

6. The method for processing high-multilayer boards with uniform inter-board adhesive filling according to claim 5, characterized in that, The graph structure modeling process specifically involves: Based on the spatial topological relationship of the temporal features of acoustic parameters in offline calibration data, the coordinates of the compression area are defined as nodes, and a graph structure is constructed using the acoustic propagation correlation matrix as edge weights. Local spatial features are extracted through three layers of graph convolution operations (64 convolution kernels per layer). A hybrid strategy of mean pooling and max pooling (weight coefficients 0.6:0.4) is adopted to enhance the global representation capability. A spatiotemporal attention mechanism (weights dynamically adjusted within a range of 0.1–0.9) is introduced to optimize the weight allocation of key regions. Finally, the graph neural network model structure parameters are output.

7. The method for processing high-multilayer boards with uniform inter-board adhesive filling according to claim 5, characterized in that, The standardized training set input to the graph neural network model during the training phase is Z-score normalized. The training set and validation set are divided according to spatial partitioning at 70%:30%, and the batch size is 64.

8. The method for processing high-multilayer boards with uniform inter-board glue filling according to claim 1, characterized in that, Step S5 specifically includes: Based on the temporal feature vector of acoustic parameters, normalized sound velocity, attenuation coefficient and nonlinear parameter standardized data are obtained. Feature dimension alignment operation is performed on the normalized sound velocity, attenuation coefficient and nonlinear parameter standardized data to generate feature tensors that meet the input dimension requirements of graph neural networks. The feature tensor is input into the graph neural network mapping model constructed in step S4, node feature aggregation operation is performed based on pre-trained weights, and the graph convolutional layer is used to perform nonlinear transformation on the spatial topological relationship between regions to generate the hidden layer state matrix. Based on the hidden layer state matrix, a regional importance weighted calculation is performed through a spatiotemporal attention mechanism, and a dynamic weight coefficient is applied to high-influence regions to generate a preliminary pressure compensation vector. The initial pressure compensation vector is subjected to sub-pixel level spatial interpolation, and a mesh mapping operation is performed based on the physical distribution topology of the piezoelectric actuator to generate a pressure compensation matrix. The pressure compensation matrix is ​​compared with the process safety threshold for boundary verification, and dynamic range trimming is performed based on thermodynamic state feedback to output the calibration pressure compensation matrix.

9. The method for processing high-multilayer boards with uniform inter-board glue filling according to claim 1, characterized in that, The distributed robust predictive controller is constructed based on the thermodynamic model parameters of the pressing process and real-time state feedback data.

10. The method for processing high-multilayer boards with uniform inter-board glue filling according to claim 1, characterized in that, The acoustic propagation parameter fluctuation threshold is generated by performing multi-region statistical fluctuation range analysis based on the temporal characteristic distribution of acoustic parameters in offline calibration data.

Images