High Multilayer Board Processing Method Based on Adaptive Control of Pressing Stress

Abstract

This invention provides a high-multilayer board processing method based on adaptive control of pressing stress. By embedding a micro-pulsed laser and an optical fiber coupled ultrasonic array within the upper and lower hot-pressing plates of the pressing mold, a full-board phonon detection network is established to excite and collect reflected sound waves from multiple interfaces in real time. Short-time Fourier transform is used to analyze the phonon group velocity, and combined with molecular dynamics simulation and experimental calibration, the local shear stress change rate is obtained. The stress spatiotemporal gradient is input into a spatiotemporal convolutional neural network to predict the stress diffusion path and peak value. If the stress exceeds a critical threshold, piezoelectric micro-region pressure compensation is intelligently triggered to achieve local pressure pulse control, effectively preventing interface instability. The process forms a closed-loop feedback correction. This invention significantly improves the real-time performance and accuracy of multi-material interface stress management.

Description

Technical Field

[0001] This invention relates to the field of adaptive control of lamination stress in the manufacturing process of high-multilayer printed circuit boards, and particularly to a processing method for high-multilayer boards based on adaptive control of lamination stress. Background Technology

[0002] High-multilayer printed circuit boards (PCBs) are widely used in the electronics and information industry. With the increasing demands for electrical performance and microstructure consistency in high-end fields such as 5G communication, data centers, and aerospace, achieving adaptive stress control during the lamination process in multi-material stacked structures has become a core technical challenge for the industry. Current high-multilayer PCB lamination processes typically employ a closed-loop thermo-mechanical parameter adjustment strategy. This involves overall temperature and pressure adjustment of the hot press plate, combined with offline process simulation models, to minimize quality defects such as delamination, bulging, and interlayer misalignment. However, with the increasing number of layers and the proportion of high-frequency and high-speed materials, the differences in interface material properties are growing. Traditional methods relying on macroscopic physical sensing quantities (such as overall temperature, displacement, or strain signals from ordinary force sensors) for process adjustment are no longer sufficient to achieve real-time capture and proactive response to micro-area, sub-second-level interface stress anomalies. Current mainstream technical approaches focus on two main aspects: one approach primarily utilizes thermo-mechanical-fluid coupled numerical simulation, establishing an experience database for different material combinations in the early stages of the process, periodically collecting temperature and macroscopic pressure data during production, and monitoring preset strain points, then correcting parameters through offline comparison; the other approach integrates multiple temperature or displacement sensors on the pressing mold, using feedback control algorithms such as PLC and PID to automatically adjust the heating rate and pressure loading speed of the pressing plate, achieving quasi-static closed-loop compensation. These approaches can meet the needs of some low-complexity, high-volume production high-multilayer PCB products, but they generally suffer from serious limitations such as dynamic response lag, low spatial resolution, critical stress instability and uncontrollability, and even sensing dead zones under multi-material interfaces, complex micro-stress fields, and variable-speed thermo-mechanical conditions. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a high-multilayer board processing method based on adaptive control of pressing stress.

[0004] The technical solution of this invention is implemented as follows: a high-multilayer board processing method based on adaptive control of pressing stress, comprising: S1: Embed a micro-pulse laser emitter and an optical fiber coupled ultrasonic receiver array in the upper and lower hot press plates of the pressing mold to form an interface phonon detection network covering the entire plate surface, so as to establish a millisecond-level sensing channel for multi-material interface stress. S2: During the pressing and heating stage, laser ultrasonic excitation pulses are synchronously emitted at a frequency of 50 Hz for each detection point on the pressing plate surface, and the reflected echo signals of the glass cloth / resin interface region are collected to obtain the raw data of phonon response in multiple regions. S3: Perform short-time Fourier transform processing on the collected reflected echo signals to extract the group velocity of the main peak of the transverse wave phonons in the 0.9–1.3 GHz frequency band, so as to generate the dynamic characteristic sequence of phonon states at each detection point. S4: Based on the mapping relationship function between transverse wave phonon group velocity and local shear stress, which was previously calibrated through molecular dynamics simulation and in-situ nanoindentation experiment, the local shear stress change rate dσ / dt at each detection point was calculated to form a multidimensional stress gradient quantification index. S5: Input the local shear stress change rate dσ / dt into a spatiotemporal convolutional neural network model trained based on 120,000 sets of multi-material combination compression historical data to obtain the prediction results of the stress spatial diffusion path and peak position within the next 200 ms, so as to realize the sub-periodic extrapolation of stress evolution trend. S6: Determine whether the predicted stress peak location exceeds the critical instability threshold. If it does, generate a pressure compensation command for the corresponding pressure plate micro-area to determine the triggering conditions and action area for active stress mitigation. S7: Activate the micro-area pressure actuator driven by the piezoelectric stack according to the micro-area pressure compensation command of the pressure plate, and implement local pressure pulse compensation with amplitude of 0.15–0.42 MPa and duration of 150–320 μs to complete the instantaneous control of the interface stress gradient; S8: Monitor the dynamics of the stress distribution after compensation. If abnormal fluctuations are detected, adjust the training parameters of the spatiotemporal convolutional neural network model to optimize the subsequent prediction accuracy and achieve closed-loop feedback correction.

[0005] The high-multilayer board processing method based on adaptive control of pressing stress provided by this invention has the following beneficial effects: (1) This invention significantly improves the control response speed and prediction accuracy of interface stress evolution during the lamination process of high-multilayer PCBs by constructing a priori sensing mechanism based on non-equilibrium phonon state reconstruction. It utilizes laser ultrasonic detection technology to capture the step-down phenomenon of shear wave phonon group velocity in the 0.9–1.3 GHz frequency band in real time. This physical response occurs approximately 3.2–5.7 seconds earlier than macroscopic deformation and exhibits a high local shear stress of up to 0.89 MPa. -1 ·ns -1The high sensitivity enables "predictive" identification of critical instability trends. By combining short-time Fourier transform with a calibrated mapping function between transverse wave phonon group velocity and local shear stress, the rate of change of shear stress in each region can be calculated within milliseconds. This provides a high spatiotemporal resolution pre-diagnostic basis for subsequent active intervention, effectively overcoming the fundamental defect of traditional thermo-mechanical coupling models that rely on posterior signals, resulting in delayed control actions.

[0006] (2) This invention achieves accurate prediction and rapid response to complex stress diffusion paths by introducing a micro-area pressure dynamic compensation closed loop driven by a spatiotemporal convolutional neural network. The spatiotemporal convolutional neural network model is trained based on 120,000 sets of historical data of multi-material combination pressing, and has a strong learning ability. It can deduce the stress field evolution trend within the next 200 ms from the spatial distribution of the current local shear stress change rate, identify potential peak positions, and guide the piezoelectric driven micro-area actuator to implement targeted pressure pulse compensation (amplitude 0.15–0.42 MPa, duration 150–320 μs). The response time of this actuator is ≤80μs. With the two-dimensional distributed pressure adjustment unit, sub-millimeter-level spatial accuracy of local stress suppression can be achieved without changing the overall process parameters. The entire control process does not require calling conventional feedback algorithms such as PID, nor does it rely on offline simulation or material database matching, which greatly improves the process robustness and cross-material system versatility.

[0007] (3) The integrated control architecture of "quantum scale sensing - intelligent prediction - micro-area execution" established in this invention is fully integrated into the standard lamination equipment without introducing any changes to the chemical formula, external compensation structures (such as shape memory alloys) or nanocomposite sensing layers. It is compatible with existing PCB stack-up structures and manufacturing processes, and has high engineering feasibility and industrial adaptability. Its innovative essence lies in transforming the intrinsic quantum scale vibration behavior of materials into a quantifiable process control signal source, realizing a paradigm shift from "observation-lag control" to "state change-predictive control". It not only significantly improves product yield and structural reliability, but also provides a brand-new physical layer sensing approach for the high-end electronic packaging field. It has good scalability and theoretical extension value, and is especially suitable for application scenarios with extremely stringent requirements for the integrity of the internal interface of PCBs, such as high-frequency high-speed communication and aerospace. Attached Figure Description

[0008] Figure 1 This is a flowchart of the high-multilayer board processing method based on adaptive control of pressing stress according to the present invention; Figure 2 This is a sub-flowchart of the high-multilayer board processing method based on adaptive control of pressing stress according to the present invention; Figure 3 This is another sub-flowchart of the high multilayer board processing method based on adaptive control of pressing stress according to the present invention. Detailed Implementation

[0009] 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.

[0010] 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.

[0011] like Figure 1 As shown, this invention provides a method for processing high-multilayer boards based on adaptive control of pressing stress, specifically including: S1: Embed a micro-pulse laser emitter and an optical fiber coupled ultrasonic receiver array in the upper and lower hot press plates of the pressing mold to form an interface phonon detection network covering the entire plate surface, so as to establish a millisecond-level sensing channel for multi-material interface stress. S2: During the pressing and heating stage, laser ultrasonic excitation pulses are synchronously emitted at a frequency of 50 Hz for each detection point on the pressing plate surface, and the reflected echo signals of the glass cloth / resin interface region are collected to obtain the raw data of phonon response in multiple regions. S3: Perform short-time Fourier transform processing on the collected reflected echo signals to extract the group velocity of the main peak of the transverse wave phonons in the 0.9–1.3 GHz frequency band, so as to generate the dynamic characteristic sequence of phonon states at each detection point. S4: Based on the mapping function between transverse phonon group velocity and local shear stress, which was pre-calibrated through molecular dynamics simulations and in-situ nanoindentation experiments, the rate of change of local shear stress at each detection point was calculated. In order to form a multidimensional stress gradient quantification index; S5: Local shear stress change rate Input a spatiotemporal convolutional neural network model trained on 120,000 sets of historical data on multi-material combination pressing to obtain the prediction results of the spatial diffusion path and peak position of stress within the next 200 ms, so as to realize the sub-periodic extrapolation of stress evolution trend. S6: Determine whether the predicted stress peak location exceeds the critical instability threshold. If it does, generate a pressure compensation command for the corresponding pressure plate micro-area to determine the triggering conditions and action area for active stress mitigation. S7: Activate the micro-area pressure actuator driven by the piezoelectric stack according to the micro-area pressure compensation command of the pressure plate, and implement local pressure pulse compensation with amplitude of 0.15–0.42 MPa and duration of 150–320 μs to complete the instantaneous control of the interface stress gradient; S8: Monitor the dynamics of the stress distribution after compensation. If abnormal fluctuations are detected, adjust the training parameters of the spatiotemporal convolutional neural network model to optimize the subsequent prediction accuracy and achieve closed-loop feedback correction.

[0012] Step S1: Embedding a micro-pulse laser emitter and an optical fiber-coupled ultrasonic receiver array within the upper and lower hot-press plates of the pressing mold forms an interface phonon detection network covering the entire plate surface, thereby establishing a millisecond-level sensing channel for multi-material interface stress. Specifically, this includes: S1.1: Based on the thermal field distribution characteristics and interlayer stress concentration area prediction results of the high-multilayer PCB board pressing mold, the internal flow channel structure of the upper and lower hot press plates is topologically optimized to generate a press plate substrate structure with embedded mounting slots, ensuring that subsequent sensor components can be embedded without damage and without affecting the uniformity of heat conduction. S1.2: Using precision micromachining technology, a micro-pulse laser emitter array with a wavelength of 266 nanometers is deployed at a fixed point in the mounting slot of the generated pressure plate substrate structure. Based on the optical path collimation principle, the emission angle of each emitter is adjusted to form a laser ultrasonic excitation source network covering the entire test area of ​​the plate surface, and outputs a high-energy laser pulse signal with picosecond pulse width. Within the pressure plate substrate structure with embedded mounting slots, precision micromachining technology (process parameters: laser micro-engraving depth 0.35 mm, slot flatness tolerance ≤ 2 micrometers) is used to achieve the fixed-point deployment function of the micro-pulse laser emitter array; Furthermore, by using an optical path collimation algorithm (parameter: collimation error ≤ 0.05 degrees, based on a dual-lens combination adjustment system), the emitted beams of each transmitter are highly consistent with the normal vector of the target interface region, ensuring that the laser ultrasonic excitation covers the entire test area of ​​the plate and forms a uniform energy distribution network; Furthermore, a pulse width compression modulation method (parameters: target pulse width 120 picoseconds, compression ratio 3:1, nonlinear optical medium material is barium β-borate crystal) is adopted to achieve high-energy laser pulse output with picosecond-level pulse width, so as to enhance the excitation efficiency of the interface phonon group; Furthermore, through power stabilization feedback control (parameters: output power stability ±0.8%, feedback period 500 microseconds), the laser pulse energy is evenly injected into different stiffness regions of the multi-material interface, and stable excitation source output power data is obtained; Furthermore, a thermal field disturbance compensation algorithm (parameter: temperature gradient threshold ≤ 0.2 K / mm) is used to correct the optical path offset of the laser array in real time, generating laser ultrasonic excitation source network configuration indicators that conform to the thermal conductivity uniformity of the entire plate; By using optical path collimation algorithm and power stabilization control, the array layout and pulse output characteristics generated in the previous step are transformed into laser ultrasonic excitation source network state data covering the entire plate surface, achieving millisecond-level triggering accuracy of multi-material interface stress. For example, with a mounting slot size of 40 mm × 15 mm and a depth of 0.35 mm, a mounting platform with a flatness tolerance of no more than 2 micrometers is formed by laser micro-engraving. An 8×8 matrix arrangement of micro-pulse laser emitters is then deployed at fixed points, with each emitter operating at a wavelength of 266 nm and a pulse width compressed to 120 picoseconds. A dual-lens combination adjustment system is used to control the collimation error of the emitted beam within 0.04 degrees, and power stabilization feedback control maintains the single-pulse output energy fluctuation range of 0.75 millijoules within ±0.8%. In the optical path design covering the entire plate surface, a thermal field disturbance compensation algorithm corrects the temperature gradient deviation to 0.15 K / mm, ensuring consistent excitation energy distribution across regions with different glass cloth stiffness. In this scenario, the optical path collimation angle and power equalization configuration parameters work together to form a high signal-to-noise ratio interface phonon excitation source network. The verification results show that in the phonon group velocity measurement link, the signal start-up time delay is significantly reduced, and the interface stress trigger response time is shortened to 1.2 milliseconds, supporting the advanced decision-making effect of the subsequent millisecond-level stress feedback control module. S1.3: For the deployed laser ultrasonic excitation source network, fiber-coupled ultrasonic receiving probes with a bandwidth of 0.8 to 3.5 GHz are coupled and installed around the corresponding excitation points. The connection between the probe and the signal transmission line is optimized through impedance matching circuit to build a high signal-to-noise ratio ultrasonic reflection echo acquisition channel and output the original analog electrical signal containing interface lattice vibration information. For the already deployed laser ultrasonic excitation source network, an optical fiber coupling installation method (parameters: bandwidth 0.8–3.5 GHz, fiber core diameter 125 μm, cladding diameter 250 μm) is adopted to achieve spatial coupling matching between the ultrasonic receiving probe and the corresponding excitation point. Furthermore, by adjusting the incident direction and distance of the probe relative to the excitation point using a coaxial fiber mechanical positioning frame (parameters: positioning accuracy ≤5 μm, shock resistance ≥80 Hz), the maximum energy capture of the reflected acoustic echo is achieved, and the original acoustic signal leader waveform is obtained. Furthermore, an impedance matching circuit (parameters: target impedance 50 Ω, matching tolerance ≤ 1 Ω) is used to achieve electrical matching between the probe output interface and the signal transmission line, and to generate transmission status data with low reflection coefficient. Furthermore, analog signal preprocessing is performed using a common-mode rejection differential amplifier module (parameters: common-mode rejection ratio ≥110 dB, amplitude-frequency response 0.8–3.5 GHz) to suppress environmental electromagnetic interference and obtain a high signal-to-noise ratio acoustic echo analog signal. Furthermore, a bandpass filter (parameters: passband range 0.8–3.5 GHz, insertion loss ≤0.5 dB) is used to perform frequency band limiting processing on the acoustic echo analog signal to achieve spectral separation of the target phonon mode and generate the original analog electrical signal containing interface lattice vibration information; By combining fiber optic coupling installation with impedance matching circuit processing, the probe deployment results of the previous step are transformed into raw ultrasonic reflection echo data with phase stability and optimized amplitude dynamic range, realizing a highly reliable input signal reference for the millisecond-level sensing channel of multi-material interface stress. For example, in a high-multilayer PCB lamination process, for a laminated structure with a resin Tg of 165 ℃ and a glass cloth stiffness of 24 GPa, a fiber-coupled ultrasonic receiver probe with a bandwidth of 0.8–3.5 GHz is arranged. The center distance between the probe and the excitation point is set to 2.5 mm, and the incident angle is set to vertical ±1°. An impedance matching circuit is used to precisely maintain the output impedance at 50 Ω, with a matching tolerance controlled at 0.5 Ω. A double-shielded coaxial cable is used for the transmission line to reduce external electromagnetic coupling. The common-mode rejection ratio of the differential amplifier is 115 dB, the amplitude-frequency response fully covers the target frequency band, the passband range of the bandpass filter is 0.85–3.45 GHz, and the insertion loss is maintained at 0.4 dB. Under this configuration, after filtering and amplification, the signal-to-noise ratio of the acoustic echo analog signal output by the receiver probe is significantly improved to the original level, the spectral energy concentration is enhanced, and the group velocity shift of the transverse wave phonon main peak is reliably captured within milliseconds. This configuration enables real-time, high-precision detection of interface stress state, providing a stable source of input data for subsequent stress inversion models; S1.4: Based on the preset spatiotemporal synchronization triggering protocol, the output original analog electrical signal is sampled and quantized by the high-speed analog-to-digital conversion module, and time alignment correction is performed in combination with the laser pulse emission time to generate a digital interface phonon response dataset with precise timestamps, and establish the data input benchmark for the millisecond-level sensing channel of multi-material interface stress. S1.5: Perform spatial interpolation and noise filtering on the generated digitized interface phonon response dataset to fill the spatial blind spots between detection points and suppress environmental thermal noise interference, so as to finally form a continuous and fully covered interface phonon detection network state map, and complete the closed-loop construction and readiness confirmation of the millisecond-level sensing channel for multi-material interface stress.

[0013] Step S2: During the pressing and heating stage, laser ultrasonic excitation pulses are synchronously emitted at a frequency of 50 Hz at each detection point on the pressing plate surface, and reflected echo signals from the glass cloth / resin interface region are collected to obtain raw data of multi-region phonon response. Specifically, this includes: S2.1: Monitor the real-time temperature status during the pressing and heating stage to generate a synchronous timing control signal that triggers the laser ultrasonic excitation pulse emission, ensuring that the excitation action is strictly locked in the critical temperature range where phonon state reconstruction occurs at the glass cloth and resin interface; For the real-time temperature status signal during the pressing and heating stage, a high-precision distributed thermocouple array measurement method (parameters: measuring point spacing of 15 mm, sampling frequency of 200 Hz) is adopted to realize multi-point synchronous acquisition of the temperature field of the entire plate surface; Furthermore, by using a three-dimensional thermal field finite difference prediction model (parameters: time step 10 ms, spatial grid resolution 5 mm), spatiotemporal interpolation of temperature field data is achieved, and a full-coverage temperature matrix containing the current temperature value of each detection point is obtained. Furthermore, by combining the temperature matrix with the critical temperature range data table for phonon state reconstruction at the glass cloth-resin interface (parameters: Tg range and corresponding temperature range for different resins), a temperature range determination algorithm is used to compare the temperature value with the critical phonon state reconstruction temperature range and generate a critical temperature range trigger flag matrix. Furthermore, by using a timing latch circuit based on a discrimination matrix, the mapping and conversion of the critical temperature zone trigger flag matrix to the synchronization timing control signal is realized, and a control signal stream containing transmission time window information is generated. Furthermore, a dynamic phase compensation algorithm (parameter: phase compensation amount ±3°) is used to perform delay correction processing on the control signal flow to ensure that the signal triggering time is consistent with the instantaneous state of phonon state reconstruction; The timing signal output module converts the phase-compensated control signal stream into a synchronous timing control signal for the laser ultrasonic excitation pulse emission, thereby ensuring that the excitation action is strictly locked within the critical temperature range where phonon state reconstruction occurs at the glass cloth and resin interface. For example, in the lamination process of a high-multilayer PCB with 18 layers and a resin Tg of 165 ℃, a distributed thermocouple measurement array with a spacing of 15 mm is used to collect a 200 Hz temperature signal. After interpolation and completion by a three-dimensional finite difference prediction model, a full-board temperature matrix is ​​obtained. According to the critical temperature zone data table, the critical temperature zone corresponding to this resin is 162–168 ℃. After matching by the temperature zone determination algorithm, a trigger flag matrix is ​​generated. The timing latch circuit maps this matrix into a control signal stream containing the target laser emission window. After correcting the 3° delay by the phase compensation algorithm, the final output is a control signal synchronized with the interface phonon state reconstruction time. In this scenario, the laser ultrasonic excitation pulse emission is strictly locked at the instant when the interface temperature reaches 163.5 ℃, ensuring that the accuracy of subsequent phonon group velocity measurement is significantly improved and the input data stability of stress change rate inversion is achieved. S2.2: Based on the synchronous timing control signal, a micro-pulse laser emitter is driven to emit a laser ultrasonic excitation pulse with a wavelength of 266 nanometers and a pulse width of 120 picoseconds into the glass cloth resin interface region to excite a high-frequency phonon group containing transverse wave modes at the interface. S2.3: Utilize an optical fiber coupled ultrasonic receiving array to capture the acoustic echo signal formed by the reflection of high-frequency phonon groups at the interface of multilayer materials, so as to complete the cross-physical field signal conversion from optical domain excitation to acoustic domain response and output the original time domain echo data stream. S2.4: Perform bandpass filtering and noise suppression on the original time-domain echo data stream to remove environmental thermal noise and circuit noise interference, thereby generating a high-purity interface reflection echo signal with a signal-to-noise ratio that meets the requirements of gigahertz frequency band analysis. S2.5: Perform full-plane spatial coordinate mapping and timestamp alignment operations based on high-purity interface reflection echo signals to construct a multi-region phonon response raw dataset containing spatial location information and temporal evolution characteristics, which serves as a standardized input object for subsequent short-time Fourier transform processing.

[0014] like Figure 2 As shown, step S3 involves performing a short-time Fourier transform on the acquired reflected echo signal to extract the group velocity of the main peak of the transverse wave phonons within the 0.9–1.3 GHz frequency band, thereby generating a dynamic characteristic sequence of phonon states at each detection point. Specifically, this includes: S3.1: The acquired reflected echo signal is windowed and truncated to generate a short signal segment with time localization characteristics, ensuring that subsequent spectrum analysis can capture the rapid changes in transient phonon group velocity. The original multi-region phonon response dataset constructed in step S2.5 is segmented using a time window function algorithm (parameter: window length). Based on the period range of the 0.9–1.3 GHz transverse wave phonon mode (set to 2.5–3 times the phonon period), time localization processing of data segments is achieved, and each segment contains sufficient waveform period to maintain spectral resolution. Furthermore, by using the Hanning window function for superposition processing (parameter: the sampling weight of the window shape gradually decays to 0 at the boundary), the boundary spectrum leakage suppression effect after time-domain truncation is achieved, and a short-time signal segment with smooth transition is obtained; Furthermore, using a frame shift setting algorithm (parameter: frame shift) By taking half the window length to ensure 50% overlap, partial sample sharing between consecutive segments is achieved, enhancing the temporal continuity of subsequent frequency domain analysis and generating a frame sequence index covering the entire plate. Furthermore, a high-precision zero-filling interpolation method (parameter: interpolation multiple is 4 times the number of sampling points) is used to perform sampling point expansion processing on short-time signal segments, thereby improving the frequency resolution of FFT operations and generating waveform data matrices that can be used for high-precision group velocity calculation; By using a multi-segment synchronous alignment processing method, the short-time signal segments from the previous step are reordered by timestamp and phase drift is removed, transforming them into a short-time signal input set with consistent timing, thereby achieving the expected technical effect of capturing rapid changes in transient phonon group velocity. For example, during the high-layer PCB lamination and heating stage, a window length is set from the high-purity interface reflected echo signal acquired by the fiber-coupled ultrasonic receiving array. The time interval is 1.2 μs, corresponding to a phonon mode period of approximately 0.5 μs. Time window truncation is performed, covering a period range 2.4 times the window length. The boundary weights of the Hanning window function are reduced to 0 to ensure a significant reduction in spectral leakage of the truncated waveform, and the frame shift... A 50% overlap structure between adjacent frames was formed by setting a time interval of 0.6 μs. Zero-padding interpolation was performed on each frame, expanding the number of sampling points from 256 to 1024, thus improving the FFT frequency resolution to approximately 0.98 MHz. Phase drift was removed by timestamp alignment of each frame's data, resulting in a stable short-time signal input set. This set can accurately capture dynamic changes in group velocity in the 0.9–1.3 GHz frequency band during subsequent short-time Fourier transforms. Verification results show that the peak value of group velocity fluctuations can be detected 0.4 ms in advance on the time axis, significantly improving the lead time for stress prediction. S3.2: Perform short-time Fourier transform processing based on the short-time signal segment to generate a time-frequency energy spectrum containing two-dimensional distribution information of time and frequency, realizing the first data form conversion from time domain waveform to frequency domain energy distribution; S3.3: Perform peak search and filtering on the time-frequency energy spectrum in the 0.9 to 1.3 GHz frequency band to extract the center frequency trajectory of the main peak of the transverse wave phonon, eliminate background noise interference, and lock the target vibration mode; For the generated time-frequency energy spectrum, a peak search algorithm (parameters: frequency window 0.9–1.3 GHz, step size 0.002 GHz) is used to locate the maximum energy point in each time slice of the target frequency band. Furthermore, by using an adaptive filtering method based on amplitude threshold (parameter: threshold is set to the mean of background noise plus twice the standard deviation), pseudo-peaks with energy below the set threshold and frequency deviation greater than 0.005 GHz are eliminated, and a preliminary purified peak candidate set is obtained. Furthermore, a continuous-time peak trajectory fitting algorithm is used (parameters: cubic spline interpolation, smoothing coefficient). =0.75), to achieve continuous reconstruction of the peak candidate set on the time axis and generate center frequency trajectory data of the target mode; Furthermore, a multi-mode separation filtering method is adopted (parameters: mode separation iteration number N=3, mode identification correlation coefficient threshold R=0.85) to reduce non-target mode components near the center frequency trajectory and output the center frequency trajectory with single shear wave phonon mode characteristics; By using peak search and filtering, the time-frequency energy spectrum generated in the previous step is transformed into a center frequency trajectory technical indicator that locks the target vibration mode and removes background noise, thereby achieving the expected technical effect of specific mode group velocity calculation. For example, in an experiment during the lamination process of a high-multilayer PCB, the acquired time-frequency energy spectrum contained multiple peaks in the 0.9–1.3 GHz frequency band. The mean background noise was 0.015 W, and the standard deviation was 0.004 W. The amplitude threshold was set to 0.015 + 2 × 0.004 = 0.023 W. The peak search algorithm located the maximum energy point with a frequency resolution of 2 MHz within each time slice, obtaining an initial peak candidate set of 1120 points. After applying the amplitude threshold to this candidate set, 854 points remained. Cubic spline interpolation was then used to continuously fit these points along the time axis, with a smoothing coefficient... =0.75, yielding a smooth center frequency trajectory curve. A multi-mode separation filtering method was employed, with iterative mode identification performed three times. A correlation coefficient threshold of 0.85 was set to remove frequency components with correlations below the threshold to the target shear wave phonon modes. The final locked center frequency trajectory remained stable at 1.072–1.081 GHz throughout the compression and heating process, suitable for subsequent group velocity calculations. This locking result significantly improves the accuracy and stability of group velocity inversion, ensuring the high purity of the stress gradient prediction input data. S3.4: Using the center frequency trajectory combined with the known glass cloth resin interface dispersion relation curve, group velocity calculation is performed to generate the instantaneous value of the transverse wave phonon group velocity at each detection point at the corresponding time, thus completing the second data form conversion from frequency characteristics to propagation velocity characteristics. The frequency trajectory of the main peak of the transverse wave phonon obtained by S3.3 processing is used to convert the frequency characteristics into propagation velocity characteristics by employing a group velocity calculation algorithm based on the dispersion relation of the glass cloth resin interface (parameters: frequency sampling point, elastic modulus of the interface material, density). Furthermore, through the dispersion relation function Achieving group speed Calculation, where Angular frequency, The wave number is used to obtain the initial propagation velocity value corresponding to each frequency sampling point; Furthermore, by using the dispersion curves measured and calibrated at the material interface, the initial group velocity value is corrected at each frequency sampling point using an interpolation algorithm, generating transverse wave phonon group velocity data with higher physical consistency. Furthermore, the corrected group velocity values ​​of each frequency sampling point are mapped to the corresponding time-series coordinates through a timestamp synchronization algorithm (parameters: echo acquisition time, window function center time), forming a time-series-serialized set of instantaneous values ​​of shear wave phonon group velocity. By using group velocity calculation and correction methods, the center frequency trajectory result of the previous step is transformed into the instantaneous value of the transverse wave phonon group velocity covering the entire plate surface, realizing the second data form conversion from frequency domain characteristics to propagation velocity characteristics. For example, in a high-multilayer PCB lamination process, for an interface region with a glass cloth stiffness of 28 GPa and a resin density of 1.85 g / cm³, the center frequency trajectory f(t) is extracted, ranging from 1.05 to 1.20 GHz, and the angular frequency is calculated. The wavenumber k is fitted based on the interface dispersion curve. (unit: m) -1 ), to obtain the initial group velocity value The group velocity was approximately 3200 m / s to 3180 m / s. After cubic spline interpolation correction at each frequency point using the measured dispersion relation curve, the group velocity values ​​were adjusted to 3225 m / s to 3198 m / s. By matching the echo acquisition time stamp with the window function center time, the group velocity values ​​were mapped to time coordinates t=0.15 ms to t=0.25 ms, forming a sequence of instantaneous shear wave phonon group velocities {3225, 3218, 3209, 3198} m / s. This sequence served as standardized input data in the subsequent S3.5 step, effectively characterizing the interfacial stress evolution trend. In actual verification, it significantly improved the stress inversion accuracy and effectively suppressed the risk of local stress hysteresis during the pressing process. S3.5: Based on the instantaneous values ​​of the transverse wave phonon group velocity, the sequence is recombined along the time axis to generate a dynamic feature sequence of phonon states covering all detection points on the plate surface, forming a standardized input dataset describing the trend of interface stress evolution.

[0015] like Figure 3 As shown, step S4 involves calculating the local shear stress change rate dσ / dt at each detection point based on the mapping function between transverse phonon group velocity and local shear stress, pre-calibrated through molecular dynamics simulations and in-situ nanoindentation experiments, to form a multidimensional stress gradient quantification index. Specifically, this includes: S4.1: Obtain the non-equilibrium phonon state reconstruction dataset of the glass cloth resin interface generated by molecular dynamics simulation and the micromechanical response dataset measured by in-situ nanoindentation experiment. Use the multiphysics coupling alignment algorithm to register the timestamps and spatial coordinates of the two datasets to generate a standardized joint calibration sample library containing transverse wave phonon group velocity values ​​and corresponding local shear stress values. Using the non-equilibrium phonon state reconstruction dataset of the glass cloth resin interface generated by molecular dynamics simulation as the input object, the data decoupling analysis method (parameter setting: molecular dynamics sampling interval 1 ps, data format is three-dimensional lattice vibration displacement matrix) is used to realize the correlation extraction between phonon state vibration mode and lattice pitch change, and obtain the corresponding mapping between transverse wave phonon group velocity characteristic curve and simulation time series. Furthermore, the micromechanical response dataset obtained by in-situ nanoindentation experiments (parameter settings: indentation load range 0.05–0.5 mN, displacement resolution 2 nm) was used with a force-displacement curve fitting algorithm (fitting accuracy better than 0.2%) to establish the correspondence between local shear stress and physical coordinates, and the experimental time sequence labeling and sampling frequency were standardized to a time reference consistent with the simulation data. Furthermore, a multiphysics coupling alignment algorithm (parameter settings: timestamp matching tolerance ≤ 0.5 ns, spatial coordinate matching tolerance ≤ 0.8 μm) is used to simultaneously align the molecular dynamics dataset and the nanoindentation experimental dataset in three dimensions, thereby achieving the registration of transverse wave phonon group velocity values ​​and corresponding local shear stress values ​​at the same spatial point and at the same time. Furthermore, a spatiotemporal dual-index remapping mechanism (spatial index based on probe matrix coordinates, temporal index based on laser pulse emission timestamps) is employed to merge data and generate a standardized joint calibration sample library, where each sample contains... numerical values ​​of transverse phonon group velocity Local shear stress values ​​and spatial three-dimensional coordinate labels; Through the above-mentioned multiphysics coupling alignment and remapping processing method, the independent data of molecular dynamics simulation and nanoindentation experiment are transformed into a unified transverse wave phonon group velocity-local shear stress joint calibration sample library, thereby improving the integrity and accuracy of the stress inversion model calibration sample. For example, the molecular dynamics simulation dataset acquired during the lamination process of high-multilayer PCBs contains non-equilibrium phonon state vibrational displacement matrices for 1024 interface nodes, with a sampling interval of 1 ps and a node spacing of 0.5 μm; the in-situ nanoindentation experimental dataset obtains force-displacement curves for 512 test points on the same board surface, with a load range of 0.1 mN to 0.3 mN and a displacement resolution of 2 nm. Vibrational mode decomposition is performed on the molecular dynamics data to extract the instantaneous values ​​of transverse wave phonon group velocities. Contact mechanics inversion was performed on the nanoindentation data to obtain the local shear stress. Values ​​were determined. Multiphysics coupling alignment was performed using time registration parameters with a matching tolerance of 0.4 ns and spatial registration parameters of 0.6 μm, ensuring complete matching of the two data types across 308 registration nodes. The resulting joint calibration sample library records the same spatial point at different timestamps during the compression heating process. and Numerical values ​​ensure that the stress inversion model constructed by subsequent Gaussian process regression has significantly improved prediction accuracy and robustness; S4.2: Based on the shear wave phonon group velocity values ​​in the standardized joint calibration sample library as independent variables and the local shear stress values ​​as dependent variables, a Gaussian process regression fitting operation is performed to construct the mapping relationship function between shear wave phonon group velocity and local shear stress, thereby outputting a stress inversion mathematical model with nonlinear compensation capability. S4.3: Receive the transverse wave phonon main peak group velocity time series data from the phonon state dynamic characteristic sequence of each detection point generated by the previous steps, call the stress inversion mathematical model to perform point-by-point mapping transformation on the transverse wave phonon main peak group velocity time series data, so as to calculate the instantaneous local shear stress time series data corresponding to each time point. The transverse wave phonon main peak group velocity time series data in the phonon state dynamic feature sequence covering the entire plate detection points are received as input objects, ensuring that the data format is consistent with the standardized dynamic features output by the preceding S3.5, and includes timestamp and spatial index information; A stress inversion mathematical model (with parameters derived from the regression fitting results of the S4.2 Gaussian process) is used to realize the nonlinear mapping and transformation function of the instantaneous velocity values ​​of the main peak group of transverse phonons, and to map the velocity signal point by point into the instantaneous values ​​of local shear stress. Furthermore, through point-to-point computation (time step) (Synchronized with spatial coordinates) to achieve parallel solution of velocity-directed stress across the entire plate surface and obtain instantaneous local shear stress time series data; Furthermore, the mapping calculation is performed using the following formula: in This represents the instantaneous value of the local shear stress. The instantaneous value of the transverse wave phonon group velocity. This is the nonlinear mapping function obtained by regression fitting using a Gaussian process; Furthermore, by combining the mean and variance functions of the Gaussian process regression model with the input velocity values, the corresponding stress values ​​and uncertainty ranges are calculated, and a complete stress time series matrix is ​​generated. By using stress inversion mapping, the group velocity time series results from the previous step are transformed into instantaneous local shear stress time series data, thus achieving the physical quantity conversion effect from shear wave phonon group velocity signal to stress quantification index. For example, during the lamination process of a multilayer PCB, the velocity sequence of the main peak group of transverse phonons at a certain detection point is [3050.2, 3048.7, 3045.1, 3040.3] m / s, with a time step of 1 ms. A stress inversion model is obtained using a molecular dynamics + nanoindentation calibration library. ,in =0.0021 MPa·s / m, =-5.8 MPa. (The speed value is missing from the original text.) Substituting these values ​​into the formula sequentially, we obtain the stress value sequence [0.69, 0.69, 0.68, 0.67] MPa. The model prediction residual is less than 0.01 MPa, and the stress fluctuation trend is consistent with the TCNN prediction result, which meets the accuracy requirements of the subsequent S4.4 difference calculation and realizes the high-fidelity stress quantification data output required for sub-period monitoring. S4.4: Apply the central difference differential algorithm to the calculated instantaneous local shear stress time series data to perform time series derivative calculations, so as to calculate the numerical sequence of local shear stress change rate, which characterizes the speed of stress dynamic evolution. S4.5: Integrate the numerical sequence of local shear stress change rate calculated from all detection points on the entire plate surface, and use spatial tensor reconstruction technology to reconstruct the one-dimensional change rate numerical sequence into a two-dimensional spatial distribution matrix to generate a multi-dimensional stress gradient quantization index to drive the subsequent prediction model. The numerical sequence of the local shear stress change rate at each detection point on the entire plate surface is received as the input data object. A tensor reconstruction algorithm based on spatial topological index mapping (parameters: set of probe point coordinates, set of rate of change values, geometric boundary of the plate) is adopted to establish a unique correspondence between each one-dimensional rate of change data and its spatial coordinates, and to form a preliminary spatial tensor representation. Furthermore, by using a two-dimensional grid interpolation algorithm (parameters: interpolation order is 3, boundary conditions are set to natural extension), the rate of change is smoothly extended in the continuous coordinate system of the plate surface, and an interpolation matrix with full domain coverage is obtained. Furthermore, a Gaussian blur filtering method is employed (parameter: kernel function standard deviation). =1.5 mm, convolution window size 5×5 grid points), to achieve the smoothing of local high gradient spikes in the interpolation matrix and generate a noise-suppressed spatial distribution matrix; Furthermore, through normalization mapping operations (parameters: minimum value mapped to 0, maximum value mapped to 1), the dimensionless processing of the change rate data of each grid point is realized, and a multidimensional stress gradient quantization index matrix that conforms to the input range of the spatiotemporal convolutional neural network is generated. The standardized value of the rate of change for each grid point is calculated using the following normalization formula: in, This is the normalized rate of change value. This is the original rate of change value. and These are the minimum and maximum values ​​of the rate of change in the matrix, respectively; By using spatial tensor recombination and normalization, the numerical sequence of local rate of change from the previous step is transformed into a two-dimensional spatial distribution matrix, thereby constructing a multidimensional stress gradient quantification index and providing a standardized input benchmark for subsequent prediction models. For example, 1024 detection points are arranged on a 750 mm × 600 mm high multilayer PCB board, and the local shear stress change rate ranges from... The range is between 0.58 MPa / ns and 1.02 MPa / ns. After inputting the set of probe point coordinates and the set of rate of change values ​​into a spatial topological index mapping algorithm, a preliminary tensor representation is generated. This representation is then expanded into an interpolation matrix with a resolution of 256×256 using a third-order interpolation algorithm. After Gaussian blur filtering, the peak gradient of the matrix is ​​reduced to 30% of its original value, and the noise distribution is significantly reduced. Normalization operations are set... = 0.58 =1.02, the normalized value of each grid point is calculated. The normalized matrix is ​​uniformly distributed between 0 and 1. When this normalized matrix is ​​input into the spatiotemporal convolutional neural network model, the deviation between the predicted stress diffusion path and the actual diffusion path is significantly reduced, and the sub-period prediction accuracy of the model is improved.

[0016] Step S5: The local shear stress change rate A spatiotemporal convolutional neural network model trained based on 120,000 sets of historical data on multi-material composite pressing is used to obtain predictions of the spatial diffusion path and peak position of stress within the next 200 ms, thereby achieving sub-periodic extrapolation of stress evolution trends. Specifically, this includes: S5.1: Perform tensor reconstruction on the input local shear stress change rate sequence to generate a four-dimensional stress gradient spatiotemporal tensor containing spatial topological index and time step dimension, which serves as the standardized input data basis for the spatiotemporal convolutional neural network model. The input multidimensional stress gradient quantization index matrix is ​​used as the processing object, which includes the spatial coordinate index of the detection points on the entire plate surface and the local shear stress change rate value at the corresponding time step. A spatial topology mapping algorithm (parameters: probe point index matrix, material stack structure diagram) is used to map each probe point in the two-dimensional stress gradient distribution matrix to the corresponding three-dimensional physical space coordinates. Furthermore, by extending the time step (parameters: sampling frequency, prediction window length), the stress change rate values ​​corresponding to each spatial coordinate are extended into a time series, forming three-dimensional tensor data with a time dimension index. Furthermore, by utilizing the dimension binding operation (parameters: spatial index dimension, time series dimension, and rate of change numerical dimension), the above three-dimensional tensor is bound to the original spatial topology index to generate complete structured data containing the spatial topology index and the time step dimension. Furthermore, by using a tensor format normalization algorithm (parameters: neural network input specification, batch size), the structured data is dimensionally arranged, numerically normalized, and format converted to obtain a four-dimensional stress gradient spatiotemporal tensor that meets the input requirements of a convolutional neural network. By using tensor reconstruction and format standardization, the multidimensional stress gradient quantization index from the previous step is transformed into a four-dimensional input data base that can be directly parsed by the convolutional neural network, thereby achieving a high-precision data-driven effect for stress evolution trend inference. For example, in the scenario of high-multilayer PCB lamination stress prediction, the input stress gradient distribution matrix has a size of 128×128, covering the two-dimensional coordinates and rate of change values ​​of all detection points on the board surface. A spatial topology mapping algorithm is used to map the detection points of the 128×128 matrix to the physical coordinate system of the board surface, forming a mapping set with a coordinate range of 0–350 mm. Time step expansion processing, based on a sampling frequency of 200 Hz and a prediction window length of 0.2 s, expands the rate of change value corresponding to each coordinate point to 40 time sampling points, generating a three-dimensional tensor of size 128×128×40. Dimension binding adds the rate of change value dimension to the three-dimensional tensor, forming a four-dimensional structure of 128×128×40×1. A tensor format normalization algorithm performs normalization in batches of 32, mapping the rate of change values ​​to the [-1,1] interval, and outputting a four-dimensional stress gradient spatiotemporal tensor that satisfies the input specification of a convolutional neural network. In this embodiment, the reconstructed tensor significantly improves feature extraction efficiency in the subsequent convolution processing stage. The prediction model's deduction of stress diffusion path and peak position within 200 ms shows a significant improvement in response speed and a significant enhancement in matching accuracy in experimental verification. S5.2: Perform multi-scale spatiotemporal convolution operation based on the four-dimensional stress gradient spatiotemporal tensor to extract the local directional feature vector of stress wave propagation at the glass cloth resin interface and the global coupled feature map to form a high-dimensional hidden layer state representation characterizing the stress evolution trend. Multi-scale three-dimensional convolution operation is used on the four-dimensional stress gradient spatiotemporal tensor (the convolution kernel size is set to 3, 5, and 5 according to the time dimension and spatial topology dimension, respectively) to extract the directional propagation features of the stress gradient tensor in a short time window and local spatial neighborhood. Furthermore, by performing parallel operations of convolution kernels at different scales (small-scale kernels capture transient local modes of stress propagation, while large-scale kernels capture global modes of cross-interface coupling), multi-scale stress waveform feature maps are generated, and a dataset containing local fluctuation direction vectors of the glass cloth / resin interface is obtained. Furthermore, channel fusion convolution (with a kernel size of 1×1×1) is performed on the multi-scale feature maps to achieve cross-scale feature aggregation of multi-scale stress wave features and generate a global coupled correlation feature mapping matrix. Furthermore, a deep residual convolutional structure (with a residual chain length of 4 layers and an activation function of ReLU) is adopted to achieve a nonlinear combination of local directional feature vectors and global coupled feature maps, thereby obtaining a high-dimensional hidden layer state tensor that describes the stress evolution trend. By using multi-scale spatiotemporal convolution processing, the four-dimensional stress gradient tensor of the previous step is transformed into a high-dimensional hidden layer state representation that includes local directional feature vectors and global coupled correlation mapping, thus realizing a sub-periodic prediction basis that combines global perception and local sensitivity to stress evolution trends. For example, a four-dimensional stress gradient spatiotemporal tensor generated during the lamination process of a multilayer PCB is input into a multi-scale convolutional module. The spatial topology index dimension is set to 64×64, the time step dimension is set to 20, and the tensor channel number is 1. In the local scale branch, a convolution kernel size of 3×3×3 is used to extract short-time and nearest-neighbor propagation direction vectors. In the global scale branch, a convolution kernel size of 5×5×5 is used to extract the overall diffusion pattern of cross-interface coupling. The feature distribution at different scales is stabilized by batch normalization, and then multi-scale information aggregation is achieved through channel fusion convolution. Finally, a high-dimensional hidden layer state tensor with a feature dimension of 128 is generated by fusion processing in the residual convolutional network. This hidden layer state tensor can significantly improve the accuracy of stress spatial diffusion path inference in the prediction task and maintain high stability of local stress peak position prediction. In the verification, the prediction bias is significantly reduced, meeting the sub-period prediction accuracy requirements. S5.3: High-dimensional hidden layer state representation is used to drive gated recurrent units to perform long-range dependent memory updates, so as to integrate the prior knowledge of 120,000 sets of multi-material combination pressing processes in history and generate stress dynamic evolution intermediate prediction sequences with noise resistance. It receives the high-dimensional hidden layer state representation generated by multi-scale spatiotemporal convolution operations as the input tensor, and adopts a gated recurrent unit network structure (parameters: hidden layer dimension 128, memory gate bias initial value 0.5) to realize long-range dependent memory update of stress evolution trend; Furthermore, the correlation weight between the current hidden state and the multidimensional stress gradient is calculated through an input gating mechanism, and the formula is used... in, For the input gate vector, For the Sigmoid function, Input the feature vector at the current time. and These are the input and hidden layer weight matrices, respectively. This represents the hidden state at the previous time step. This serves as a bias vector, enabling selective preservation of new input features; Furthermore, the decay coefficient of historical hidden states is calculated using the forget gate mechanism, and the formula is used. in, Forget gate vector, and Here is the forget gate weight matrix. To bias the forget gate, ensure that the model can eliminate historical patterns that are irrelevant to the current stress state; Furthermore, a candidate state generation mechanism is used to combine input features with the previous hidden state to generate memory update content, utilizing the formula... in, For candidate state vectors, It is the hyperbolic tangent function. and The candidate state weight matrix is... The candidate state is biased to supplement the nonlinear memory of stress dynamic evolution; Furthermore, through the cell state update formula The historical cell state and the new candidate state are fused according to the forgetting and input weights to obtain the latest cell state. This enables the spatial-temporal fusion of stress mode memory; The exposure ratio of the current cell state is calculated using the output gate mechanism, and a new hidden layer output is generated. This is achieved using the formula... in, For the output gate vector, and This is the output gate weight matrix. For output gate bias, use the formula Generate noise-resistant intermediate prediction sequences for stress dynamic evolution. ; By integrating long-range dependent memory updates of gated cyclic units with prior knowledge of historical processes, the high-dimensional hidden layer features of the previous step are transformed into robust intermediate prediction sequences, thereby improving the anti-interference capability of future stress wave diffusion path prediction. For example, in a high-multilayer PCB lamination scenario, the spatial topological dimension of the four-dimensional stress gradient spatiotemporal tensor is configured as 64×64, the time step dimension is 50, and the input gate weights are... Initialize as a normal distribution matrix with mean 0 and standard deviation 0.01, with forget gate bias. Set to 1.2 to enhance short-term memory retention, candidate state weights Assume LeCun normal initialization. When executing the input gate formula, the input vector... The maximum value corresponds to the peak region of the stress change rate. The Sigmoid function output ranges from 0.7 to 0.9, indicating that highly correlated features are fully preserved. When executing the forgetting gate formula, the forgetting coefficient... In the low stress change rate region, the value drops to 0.3, eliminating non-critical stress waveforms; when executing the candidate state generation formula, the region where the tanh output is close to ±0.8 represents a strong driving force for the evolution of stress peaks. After updating the element state, in the output gate calculation, The values ​​are concentrated around 0.85, combined with The output achieves accurate matching between the intermediate prediction sequence and the historical stress diffusion pattern, which significantly reduces the error of the peak position in subsequent predictions within 200 ms, and the prediction results maintain high stability in the determination of the stress peak criticality. S5.4: Perform a fully connected mapping transformation on the intermediate prediction sequence of stress dynamic evolution to decode the high-dimensional hidden layer features into a cloud map of the probability distribution of stress spatial diffusion within the next 200 milliseconds and the confidence interval of potential peak positions in physical space coordinates; For the intermediate prediction sequence of stress dynamic evolution generated by gated cyclic units, a fully connected mapping transformation method (parameters: number of layers = 3, number of hidden units = 256, activation function = ReLU) is adopted to realize the mapping function from the high-dimensional hidden layer state representation to the physical space coordinate domain. Furthermore, by performing weight matrix multiplication and bias vector addition operations (weight matrix size: 256×N, where N is the total number of spatial coordinates × the number of time steps), a linear transformation from the hidden layer feature vector to the predicted output vector is achieved, and a preliminary stress spatial diffusion probability matrix is ​​obtained. Furthermore, the Softmax normalization method (temperature coefficient = 1.0) is used to normalize the probability value matrix at each point in the spatial coordinates, and to generate a numerical representation of the stress diffusion probability distribution cloud map at each location within the next two hundred milliseconds. Furthermore, by transforming with the Sigmoid function, the probability value of the corresponding potential peak position index in the output vector is compressed to the [0,1] interval, thereby realizing the peak position confidence calculation and obtaining the potential peak position confidence interval data within the next two hundred milliseconds; Through fully connected mapping transformation, the high-dimensional hidden layer feature representation of the previous step is transformed into a stress diffusion probability distribution cloud map and a confidence interval of potential peak position in physical space coordinates, thereby realizing the visualization and structuring of the prediction results and providing standardized input for subsequent extreme value search and trajectory fitting. For example, in the high-multilayer PCB lamination process, the length of the intermediate prediction sequence for stress dynamic evolution is set to 40 time steps, and the dimension of the hidden layer feature vector at each time step is 256. When using fully connected mapping transformation, the weight matrix size is set to 256×512 for the first layer, 512×1024 for the second layer, and 1024×(X×Y×T) for the third layer, where X×Y is the total number of spatial grid points on the lamination board surface, and T is 4 future time steps. The first and second layers both use ReLU activation, and the third layer uses Softmax normalization. The probability mapping formula is as follows: in It is a probability matrix. This is the weight matrix. These are the hidden layer feature vectors. This is the bias vector. It is obtained after Softmax normalization. The matrix elements represent the stress diffusion probability at the corresponding spatial location and time step. The index of the location with the maximum value corresponds to the potential peak location, which is then processed by the Sigmoid function. Implement confidence interval calculation, where The mean of the probability matrix is ​​used. The results show that the probability distribution cloud map can significantly improve the spatial resolution of future stress diffusion region predictions. The stability of the peak position confidence interval output is verified under multi-material combinations, meeting the prediction accuracy requirements of closed-loop control. S5.5: Based on the stress spatial diffusion probability distribution cloud map and the confidence interval of potential peak positions, extreme value search and trajectory fitting are performed to output the determined future stress spatial diffusion path vector set and the accurate stress peak position prediction results, thus completing the sub-period stress evolution trend deduction.

[0017] Step S6: Determine whether the predicted stress peak location exceeds the critical instability threshold. If it does, generate a pressure compensation command for the corresponding pressure plate micro-area to determine the triggering condition and effective area for active stress mitigation. Specifically, this includes: S6.1: Obtain the predicted results of the spatial diffusion path and peak position of stress within the next two hundred milliseconds from the output of the spatiotemporal convolutional neural network model, and perform extreme value search processing on the multidimensional stress gradient quantization index in the prediction results to extract the peak data of the predicted local shear stress change rate corresponding to each detection point. For the predicted stress spatial diffusion path and peak position within the next 200 milliseconds output by the spatiotemporal convolutional neural network model, an extreme value search algorithm (parameter settings: search window covers the entire prediction space, step size is 1 detection point index unit) is used to scan and process the multidimensional stress gradient quantization index matrix in the prediction results and capture the candidate peak points of the local shear stress change rate. Furthermore, by using a local quadratic surface fitting method (parameter settings: the fitting neighborhood radius is 2 detection points, and the fitting basis function is a quadratic polynomial), peak interpolation correction processing of candidate peak points is achieved, and the corrected peak value of stress change rate is obtained, so as to improve the spatial accuracy of peak data extraction. Furthermore, by using a threshold pre-screening algorithm (parameter setting: the screening threshold is the mean of the predicted rate of change sequence plus three standard deviations), preliminary anomaly detection of the corrected peak estimate is achieved, and a peak candidate vector containing anomaly markers is generated to reduce the risk of subsequent misjudgment; Furthermore, a multi-scale gradient direction consistency verification method is adopted (parameter setting: the direction consistency criterion is that the gradient direction difference is less than 5°) to realize the directionality verification of the candidate vector of the peak of the anomaly label and generate a peak data subset with consistent direction; By using feature index mapping, the peak data subset with consistent direction is transformed into the peak data of the predicted local shear stress change rate corresponding to each detection point, thus achieving the expected technical effect of this sub-step. For example, in the high-multilayer PCB lamination process, the input stress space diffusion path prediction matrix has a size of 128×128 points, a time dimension of 20 sampling steps, and a prediction window of 200 milliseconds. An extreme value search algorithm is used to scan the entire domain, obtaining 512 candidate peak points. A local quadratic surface fitting method is used to interpolate and correct each candidate peak, with a neighborhood radius of 2 detection points. The corrected peak change rate value is... Between MPa / ms. A threshold pre-screening algorithm is applied, setting the screening threshold to the mean of the predicted rate of change. Add three standard deviations Filter out those with a change rate greater than A total of 74 peak points were identified at MPa / ms. Further, a multi-scale gradient direction consistency verification method was used, setting the direction difference criterion to be less than... The system retains 59 peak data points with consistent orientation. The final output contains the peak data of the predicted local shear stress change rate from these 59 detection points, which serves as input for subsequent comparison of critical instability thresholds, effectively improving the accuracy and stability of abnormal stress zone identification. S6.2: Based on the pre-stored database of critical instability thresholds for multilayer board material interfaces, call the critical instability threshold parameters that match the current glass cloth stiffness and resin glass transition temperature, and perform boundary comparison calculations on the extracted predicted local shear stress change rate peak data to generate a stress stability discrimination vector containing overthreshold state markers. S6.3: Based on the overthreshold state marker in the stress stability discrimination vector, the set of coordinates of abnormal detection points that exceed the critical instability threshold is selected. The set of coordinates of abnormal detection points is then processed by neighborhood fusion using a spatial clustering algorithm to delineate the range of action of the micro-region to be compensated with continuous stress concentration characteristics. Anomaly detection point screening is performed based on the over-threshold state marker in the stress stability discrimination vector. A matrix index retrieval method (parameters: discrimination vector, spatial coordinate matrix) is used to extract the target point dataset with stress change rate exceeding the critical instability threshold from the coordinate set of detection points on the entire plate surface. Furthermore, the spatial relationship between any two points in the target point dataset is calculated using a spatial distance metric algorithm (parameters: Euclidean distance, neighborhood radius threshold), and a distance matrix containing the pairwise distances of all outlier points is generated. Furthermore, the density-based spatial clustering method DBSCAN (parameters: minimum number of samples, neighborhood radius threshold) is used to perform clustering analysis on the distance matrix, so as to automatically assign the outlier points that are close to each other and whose density meets the set conditions to the same cluster on the spatial coordinate plane, and obtain a set of cluster division identifiers with continuous stress concentration characteristics. Furthermore, a centroid calculation algorithm (parameter: coordinates of all points within the cluster) is used to perform centroid calculation on the spatial range of each cluster, generating the centroid coordinates of the corresponding cluster and the cluster range boundary rectangle data; By fusing cluster centroid and boundary rectangle data, the cluster division results from the previous step are transformed into a dataset of the range of micro-areas to be compensated, thereby achieving accurate mapping from anomaly detection points to continuous stress concentration feature areas. For example, in the lamination process of high-multilayer PCBs, the input stress stability discrimination vector contains 1024 detection points, of which 86 points are marked as exceeding the threshold state. The spatial coordinates of these 86 points are extracted using matrix index retrieval and the Euclidean distance formula is applied. in , The coordinates of the current point. , Using the coordinates of reference points, the distance matrix between any two points was calculated. When performing DBSCAN clustering, a neighborhood radius threshold of 3.5 mm and a minimum sample size of 4 were set, resulting in a partition of 9 clusters, each satisfying the continuous stress concentration characteristic condition. Centroid coordinates were calculated for each cluster, and boundary rectangles were generated using the minimum and maximum x and y values ​​of points within the cluster, ultimately forming 9 data entries for the effective range of the micro-regions to be compensated. In the verification, the defined effective range highly correlated with the stress concentration region in the phonon response spectrum, ensuring the accurate application location of subsequent pressure compensation commands and significantly improving the targeting and effectiveness of the compensation action. S6.4: For the defined range of action of the micro-area to be compensated, combined with the magnitude of the amplitude of the predicted peak value of the local shear stress change rate, perform pressure pulse amplitude and timing mapping calculation to generate a pressure plate micro-area pressure compensation command data package containing the specific compensation amplitude, duration and trigger timing. S6.5: Perform protocol encapsulation and verification on the generated pressure plate micro-area pressure compensation command data packet, and send the encapsulated pressure plate micro-area pressure compensation command to the piezoelectric stack drive control interface to determine the final triggering condition for active stress suppression and lock the target micro-area position of the actuator action.

[0018] Step S7: Activate the micro-area pressure actuator driven by the piezoelectric stack according to the micro-area pressure compensation command of the pressure plate, and implement local pressure pulse compensation with an amplitude of 0.15–0.42 MPa and a duration of 150–320 μs to complete the instantaneous control of the interface stress gradient. Specifically, this includes: S7.1: Parse and process the received pressure plate micro-area pressure compensation command, extract the target action area coordinates, target pressure amplitude parameters and target duration parameters, so as to generate a micro-area pressure drive data packet containing spatial positioning information and waveform control parameters; S7.2: Based on the target pressure amplitude parameter and target duration parameter in the micro-area pressure driving data packet, the required driving voltage step value and pulse width modulation duty cycle are calculated using the high-pressure piezoelectric ceramic inverse piezoelectric effect mapping model, so as to generate a discretized voltage excitation sequence that adapts to the physical characteristics of the piezoelectric stack. S7.3: Perform high-speed digital-to-analog conversion and power amplification on the discretized voltage excitation sequence, and convert the digital voltage signal into a high-response-rate analog high-current drive signal through a broadband linear power amplifier to form a strong electric field drive source that can instantaneously excite mechanical deformation of the piezoelectric stack. S7.4: Using a strong electric field driving source to act on the piezoelectric stacked actuator unit in a designated area, the piezoelectric lattice structure is triggered to undergo rapid axial expansion and contraction deformation, converting electrical energy into mechanical displacement output with micron-level precision, so as to generate a local pressure pulse mechanical wave with specific amplitude and duration. For the analog high-current drive signal output by the broadband linear power amplifier, the electric field localization method (parameters: target action area coordinates, drive voltage step value of 75V, pulse width modulation duty cycle of 60%) is used to realize the local coupling injection of high voltage electric field in the piezoelectric stack actuation unit. Furthermore, by using piezoelectric lattice orientation consistency modulation technology (parameters: piezoelectric ceramic lattice orientation coefficient, driving electric field strength), transient axial stretching deformation of the piezoelectric lattice structure is achieved, and initial deformation data are obtained. Furthermore, based on the inverse piezoelectric effect physical mapping model (parameters: dielectric constant ε is 1600, piezoelectric coefficient d...), 33 With an applied electric field strength of 1.5 kV / mm and a voltage of 450 pm / V, the quantification of electrical energy into mechanical displacement is achieved, generating mechanical displacement output values ​​with micron-level precision. The axial deformation is calculated using the following formula: in, For the axial deformation of the piezoelectric stack, The longitudinal piezoelectric coefficient of the piezoelectric material. To apply an electric field strength; Furthermore, by utilizing a mechanical wave modulation algorithm (parameters: target pressure amplitude, duration, displacement waveform control curve), the time-domain shaping conversion from axial deformation to local pressure pulse mechanical wave is realized, and a pressure pulse conforming to the set amplitude and duration is generated; By using a mechanical wave coupling transmission method (parameters: micro-area action range, interface material elastic modulus), the mechanical displacement output of the previous step is converted into a local pressure pulse mechanical wave, thereby achieving efficient concentration and localization of local pressure energy. For example, with a target area coordinate of (24.5 mm, 12.0 mm), a target pressure amplitude of 0.30 MPa, a duration of 200 μs, a driving voltage step of 75 V, a pulse width modulation duty cycle of 60%, and a longitudinal piezoelectric coefficient of the piezoelectric material... Set as pm / V, dielectric constant is Apply electric field strength kV / mm. Based on the formula... Calculate axial deformation for The deformation, measured in nm, is transmitted to the interfacial material (elastic modulus of 3.5 GPa), generating a mechanical wave of precise amplitude. After modulation, this wave forms a pressure pulse with a peak value of 0.30 MPa and a constant duration of 200 μs. This pressure pulse acts before the interfacial stress gradient forms a critical instability, eliminating the stress peak and significantly improving interfacial stability, thus achieving the instantaneous control target. S7.5: The local pressure pulse mechanical wave is coupled and transmitted to the glass cloth resin interface to apply a reverse mechanical cancellation effect on the local shear stress gradient of the interface, so as to complete the instantaneous control of the interface stress gradient and eliminate the predicted stress peak instability risk. The input condition is a local pressure pulse mechanical wave generated by S7.4. This mechanical wave has a micrometer-level axial deformation amplitude and a hundred-microsecond-level duration parameter output by the piezoelectric stack actuation unit, and is located in the preset glass cloth resin interface action area. An elastic mechanical wave propagation model (parameters: interface material density ρ, elastic modulus E, acoustic impedance Z boundary) is adopted to realize the effective coupling propagation path mapping of local pressure pulse mechanical waves through the pressure plate matrix and the interlayer bonding interface. Furthermore, by using an interface acoustic impedance matching algorithm (parameter: threshold of difference between Z boundary and Z wave), the maximum transmission efficiency of mechanical wave energy at the glass cloth resin interface is achieved, and the transient stress distribution matrix at the interface is obtained. Furthermore, by using the shear stress reverse superposition calculation method (parameters: the existing stress gradient matrix σstep and the compensation wave stress matrix σcomplement), the reverse mechanical cancellation effect of the local shear stress gradient at the interface is achieved, and a new stress gradient correction matrix is ​​generated. Furthermore, the stress gradient correction value is calculated using the following formula. : in Given the existing stress gradient matrix, The sign of the reverse stress matrix generated under the action of the compensation wave depends on the phase relationship between the compensation wave and the existing stress. Furthermore, through the spatially finite difference update algorithm (parameters: The matrix and interface mesh resolution h) are used to realize the numerical update of the stress gradient of the entire interface area and obtain the stress field distribution map after instantaneous adjustment. By using interface coupling transmission and reverse mechanical cancellation processing, the result of the local pressure pulse mechanical wave action in the previous step is transformed into a low-amplitude matrix of interface stress gradient, thereby eliminating the risk of predicted stress peak instability. For example, in a certain high-multilayer PCB lamination process embodiment, the interface material density is configured to be 1.85 g / cm³, the elastic modulus is 4.2 GPa, and the acoustic impedance matching threshold is set to 5×10⁻⁶. 5 The pressure pulse mechanical wave output by the piezoelectric stack has an amplitude of 0.32 MPa and a duration of 220 μs. Using the above-mentioned elastic wave transmission model, the peak value of the stress distribution matrix at the interface is 0.29 MPa. Using the shear stress inverse superposition calculation method, the existing stress gradient matrix peak value of 0.31 MPa is corrected to 0.02 MPa after the compensation wave. The corrected value matrix is ​​input into the spatial finite difference update algorithm (interface mesh resolution set to 0.5 mm) to generate a stress field distribution map after regulation. In this distribution map, the unstable region of the original predicted peak value disappears, and the interface stress gradient is uniformly distributed, verifying that the instantaneous regulation effect significantly improves the process stability.

[0019] Step S8: Monitor the dynamics of the compensated stress distribution. If abnormal fluctuations are detected, adjust the training parameters of the spatiotemporal convolutional neural network model to optimize subsequent prediction accuracy and achieve closed-loop feedback correction. Specifically, this includes: S8.1: Perform high-frequency laser ultrasonic scanning on the interface area of ​​the pressure plate micro-area pressure actuator after local pressure pulse compensation to obtain the transient reflected echo signal sequence after compensation, and use the reconstructed multi-point phonon state dynamic characteristic sequence as the initial input data. S8.2: Based on the reconstructed multi-point phonon state dynamic feature sequence, the short-time Fourier transform algorithm is used to extract the main peak group velocity of the transverse wave phonon, and the real-time local shear stress change rate is calculated by combining the mapping relationship function between the transverse wave phonon group velocity and the local shear stress, so as to generate a set of compensated multidimensional stress gradient quantification indexes. S8.3: Perform spatiotemporal difference operation on the set of compensated multidimensional stress gradient quantification indicators and compare it with the preset critical instability threshold to extract the spatial distribution map of stress residuals that exceed the allowable range, so as to generate an abnormal fluctuation feature vector that characterizes the control deviation. S8.4: Based on the generated abnormal fluctuation feature vector, the gradient error term of the weights of each layer inside the spatiotemporal convolutional neural network model is calculated using the backpropagation algorithm to quantify the nonlinear mapping deviation between the current model prediction result and the actual stress evolution trend. S8.5: Based on the nonlinear mapping deviation obtained by quantization, the adaptive moment estimation algorithm is used to update the convolution kernel weight matrix and bias parameters of the spatiotemporal convolutional neural network model, so as to output an optimized spatiotemporal convolutional neural network model parameter set with higher prediction accuracy to complete the closed-loop feedback correction.

[0020] 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.

[0021] 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 high-multilayer board processing method based on adaptive control of pressing stress, characterized in that, Includes the following steps: S1: A micro-pulse laser transmitter and an optical fiber coupled ultrasonic receiver array are embedded in the upper and lower hot press plates of the pressing mold to form an interface phonon detection network. S2: During the pressing and heating stage, laser ultrasonic excitation pulses are synchronously emitted at each detection point on the pressing plate surface, and the reflected echo signal of the glass cloth / resin interface area is collected. S3: Perform short-time Fourier transform processing on the reflected echo signal to extract the group velocity of the main peak of the transverse wave phonon and generate a dynamic feature sequence of the phonon state. S4: Based on the mapping relationship function between transverse phonon group velocity and local shear stress, which was previously calibrated by molecular dynamics simulation and in-situ nanoindentation experiment, calculate the rate of change of local shear stress at each detection point in the dynamic characteristic sequence of the phonon state. S5: Input the local shear stress change rate into the spatiotemporal convolutional neural network model to obtain the prediction results of the stress spatial diffusion path and peak position in the future period; S6: Determine whether the predicted stress peak location exceeds the critical instability threshold. If it does, generate a pressure compensation command for the corresponding pressure plate micro-area. S7: Activate the micro-area pressure actuator according to the pressure plate micro-area pressure compensation command, implement local pressure pulse compensation, and complete the instantaneous control of the interface stress gradient; S8: Monitor the dynamics of the stress distribution after compensation. If abnormal fluctuations are detected, adjust the training parameters of the spatiotemporal convolutional neural network model to optimize the subsequent prediction accuracy and achieve closed-loop feedback correction.

2. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 1, characterized in that, The micro-pulse laser emitter is configured with a laser micro-engraving depth of 0.35 mm, a slot flatness tolerance of less than or equal to 2 micrometers, an array working wavelength of 266 nm, and an output pulse width of 120 ps. The fiber-coupled ultrasonic receiver probe has a bandwidth of 0.8–3.5 GHz and an optical fiber diameter of 125–250 μm.

3. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 1, characterized in that, Step S3 specifically includes: The acquired reflected echo signals are windowed and truncated to generate short-time signal segments; Short-time Fourier transform processing is performed on the short-time signal segment to generate a time-frequency energy spectrum. Peak search and filtering were performed on the time-frequency energy spectrum in the 0.9 to 1.3 GHz frequency band to extract the center frequency trajectory of the main peak of the transverse wave phonon. The group velocity is calculated by combining the center frequency trajectory with the known dispersion relation curve of the glass cloth resin interface, and the instantaneous values ​​of the transverse wave phonon group velocity at each detection point at the corresponding time are generated. Based on the instantaneous values ​​of the transverse wave phonon group velocity, a sequence of phonon state dynamic characteristics is generated by sequential recombination along the time axis.

4. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 3, characterized in that, The time-frequency energy spectrum contains two-dimensional distribution information of time and frequency.

5. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 1, characterized in that, Step S4 specifically includes: The non-equilibrium phonon state reconstruction dataset of the glass cloth resin interface generated by molecular dynamics simulation and the micromechanical response dataset measured by in-situ nanoindentation experiment are obtained. The timestamps and spatial coordinates of the non-equilibrium phonon state reconstruction dataset of the glass cloth resin interface and the micromechanical response dataset are registered to generate a standardized joint calibration sample library. Based on the shear wave phonon group velocity values ​​in the standardized joint calibration sample library as independent variables and the local shear stress values ​​as dependent variables, a Gaussian process regression fitting operation is performed to construct a mapping relationship function between shear wave phonon group velocity and local shear stress, thereby outputting a stress inversion mathematical model. Receive the transverse wave phonon main peak group velocity time series data in the phonon state dynamic characteristic sequence of each detection point generated in step S3, call the stress inversion mathematical model to perform point-by-point mapping transformation on the transverse wave phonon main peak group velocity time series data, and calculate the instantaneous local shear stress time series data corresponding to each time point. The central difference differential algorithm is applied to the instantaneous local shear stress time series data to perform time series derivative calculations, and the numerical sequence of the local shear stress change rate is calculated. By integrating the numerical sequence of local shear stress change rate calculated from all detection points on the entire plate, the one-dimensional change rate numerical sequence is reconstructed into a two-dimensional spatial distribution matrix using spatial tensor reconstruction technology, thereby generating a multi-dimensional stress gradient quantification index.

6. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 5, characterized in that, Each sample in the standardized joint calibration sample library contains transverse wave phonon group velocity values, local shear stress values, and spatial three-dimensional coordinate labels.

7. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 1, characterized in that, Step S5 specifically includes: Tensor reconstruction is performed on the input local shear stress rate of change sequence to generate the stress gradient spacetime tensor; Multi-scale spatiotemporal convolution is performed based on the stress gradient spatiotemporal tensor to extract the local directional feature vector of stress wave propagation at the glass cloth resin interface and the global coupled feature map to form a high-dimensional hidden layer state representation. The high-dimensional hidden layer state representation is used to drive the gated loop unit to perform long-range dependent memory updates, so as to integrate the prior knowledge of 120,000 sets of multi-material combination pressing processes in history and generate intermediate prediction sequences of stress dynamic evolution. Perform a fully connected mapping transformation on the intermediate prediction sequence of stress dynamic evolution to decode the high-dimensional hidden layer features into a cloud map of the probability distribution of stress spatial diffusion within the next 200 milliseconds and the confidence interval of potential peak positions in physical space coordinates. Based on the stress spatial diffusion probability distribution cloud map and the confidence interval of the potential peak position, extreme value search and trajectory fitting are performed to output the determined future stress spatial diffusion path vector set and the accurate stress peak position prediction result.

8. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 7, characterized in that, The spatiotemporal convolutional neural network model receives a four-dimensional stress gradient spatiotemporal tensor input and extracts features through multi-scale three-dimensional convolution and residual network. In the multi-scale three-dimensional convolution, the kernel size is set to 3, 5, and 5 according to the time dimension and spatial topology dimension, respectively. The residual network has a residual chain length of 4 layers and the activation function is ReLU.

9. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 1, characterized in that, The pressure compensation command for the pressure plate micro-area includes the specific compensation amplitude, duration, and triggering sequence.

10. The high-multilayer board processing method based on adaptive control of pressing stress according to claim 1, characterized in that, The micro-area pressure actuator is composed of a piezoelectric stack, with a driving voltage step value of 75V, a pulse width modulation duty cycle of 60%, a longitudinal piezoelectric coefficient of 450pm / V, a dielectric constant of 1600, and an applied electric field strength of 1.5kV / mm.

Images