Method and system for detecting copper-clad quality of copper-clad plate

Abstract

The invention discloses a copper-clad plate copper-clad quality detection method and system, and the method comprises the steps: applying an ultrasonic excitation signal with preset frequency and power to a to-be-detected copper-clad plate, and synchronously employing a high-speed camera to collect a dynamic micro-vibration response image sequence of the surface of the copper-clad plate under the excitation of ultrasonic waves; performing phase locking analysis on the dynamic micro-vibration response image sequence, and constructing a copper-clad plate dynamic response feature map; inputting the dynamic response feature map of the copper-clad plate into the trained defect identification model, and outputting a preliminary defect area distribution map; performing multi-scale morphological analysis and region connectivity verification on the preliminary defect region distribution map to generate an accurate positioning defect region map; and generating a comprehensive evaluation report of the copper coating quality according to a preset quality scoring rule on the basis of the accurately positioned defect area graph. According to the embodiment of the invention, the method can achieve the high-precision positioning and classification of the defects of the copper-clad plate, and provides an efficient and reliable automatic solution for the quality evaluation of the copper-clad plate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of quality detection, and particularly relates to a copper-clad plate copper-clad quality detection method and system. BACKGROUND

[0002] As a core substrate of a printed circuit board (PCB), the copper-clad layer of a copper-clad plate directly affects the circuit performance and reliability. Traditional detection methods such as visual inspection, electrical performance testing or X-ray scanning have the limitations of low efficiency, high cost or difficulty in detecting micron-level defects. Especially for hidden defects such as poor copper-clad bonding and micro-cracks, the prior art cannot achieve high-precision, non-destructive and rapid positioning. In recent years, although non-destructive detection technology based on vibration response has developed, it is limited by environmental noise interference and signal analysis capability, and cannot effectively separate the small vibration characteristics under complex working conditions. In addition, the conventional image processing algorithm has insufficient feature extraction capability for the dynamic response characteristics of the copper-clad plate, resulting in a high false detection rate. SUMMARY

[0003] The purpose of the present application is to provide a copper-clad plate copper-clad quality detection method and system to solve the problems in the prior art and achieve high-precision positioning and classification of copper-clad plate defects, thereby providing an efficient and reliable automatic solution for copper-clad plate quality evaluation.

[0004] One embodiment of the present application provides a copper-clad plate copper-clad quality detection method, which comprises the following steps:

[0005] An ultrasonic excitation signal with a preset frequency and power is applied to the copper-clad plate to be detected, and a high-speed camera is used to synchronously capture a dynamic micro-vibration response image sequence of the copper-clad plate surface under ultrasonic excitation, wherein the ultrasonic frequency is associated with the natural frequency of the copper-clad plate substrate and the copper layer to excite specific mode micro-vibration;

[0006] Phase-locked analysis is performed on the dynamic micro-vibration response image sequence to extract the amplitude-phase variation characteristics of each pixel point within an ultrasonic excitation cycle, and a copper-clad plate dynamic response feature map is constructed, wherein the phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by excitation from environmental noise;

[0007] The copper-clad plate dynamic response feature map is input into a trained defect recognition model to identify and locate potential poor copper-clad bonding areas, micro-crack areas and local stress abnormal areas, and output a preliminary defect area distribution map, wherein the defect recognition model is a deep learning model trained based on the dynamic response feature map of known defect samples;

[0008] The preliminary defect area distribution map is subjected to multi-scale morphological analysis and region connectivity verification, combined with the design line topological information of the copper-clad plate, to screen out real defect areas conforming to physical laws, and a precisely positioned defect area map is generated, wherein the region connectivity verification is used to exclude false detection caused by image noise or non-copper-clad areas;

[0009] Based on the precisely positioned defect area map, the area proportion, distribution density and distance from key lines of each type of defect are calculated, and a comprehensive evaluation report of copper-clad quality is generated according to a preset quality scoring rule, wherein the comprehensive evaluation report includes defect type statistics, location labeling and quality level determination.

[0010] Yet another embodiment of the present application provides a copper-clad plate copper-clad quality detection system, which comprises:

[0011] A collection module is configured to apply an ultrasonic excitation signal of a preset frequency and power to the copper-clad plate to be detected, and simultaneously collect a dynamic micro-vibration response image sequence of the copper-clad plate surface under ultrasonic excitation using a high-speed camera, wherein the ultrasonic frequency is associated with the inherent frequency of the copper-clad plate substrate and copper layer to excite specific mode micro-vibration;

[0012] An analysis module is configured to perform phase-locked analysis on the dynamic micro-vibration response image sequence to extract the amplitude-phase variation characteristics of each pixel point within an ultrasonic excitation cycle, and construct a copper-clad plate dynamic response feature map, wherein the phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by excitation from environmental noise;

[0013] An identification module is configured to input the copper-clad plate dynamic response feature map into a trained defect identification model to identify and locate potential poor copper-clad bonding areas, micro-crack areas and local stress abnormal areas, and output a preliminary defect area distribution map, wherein the defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples;

[0014] A verification module is configured to perform multi-scale morphological analysis and region connectivity verification on the preliminary defect area distribution map, combined with the design line topological information of the copper-clad plate, to screen out real defect areas conforming to physical laws, and generate a precisely positioned defect area map, wherein the region connectivity verification is used to exclude false detection caused by image noise or non-copper-clad areas;

[0015] A generation module is configured to calculate the area proportion, distribution density and distance from key lines of each type of defect based on the precisely positioned defect area map, and generate a comprehensive evaluation report of copper-clad quality according to a preset quality scoring rule, wherein the comprehensive evaluation report includes defect type statistics, location labeling and quality level determination.

[0016] Yet another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any of the above embodiments when run.

[0017] Yet another embodiment of the present application provides an electronic device comprising a memory having a computer program stored therein and a processor configured to execute the computer program to perform the method described in any of the above embodiments.

[0018] Compared with the prior art, the copper-clad plate copper-clad quality detection method provided by the present application applies an ultrasonic excitation signal of a preset frequency and power to the to-be-detected copper-clad plate, and synchronously uses a high-speed camera to collect a dynamic micro-vibration response image sequence of the copper-clad plate surface under the excitation of the ultrasonic wave; performs phase-locked analysis on the dynamic micro-vibration response image sequence to construct a copper-clad plate dynamic response feature map; inputs the copper-clad plate dynamic response feature map into a trained defect recognition model to output a preliminary defect region distribution map; performs multi-scale morphological analysis and region connectivity verification on the preliminary defect region distribution map to generate an accurately positioned defect region map; and generates a comprehensive evaluation report of the copper-clad quality based on the accurately positioned defect region map and according to a preset quality scoring rule, so that high-precision positioning and classification of the copper-clad plate defects can be realized, and an efficient and reliable automatic solution for copper-clad plate quality evaluation is provided. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A hardware structure block diagram of a computer terminal of the copper-clad plate copper-clad quality detection method provided by the embodiment of the present application is shown in the figure.

[0020] Figure 2 A flowchart of the copper-clad plate copper-clad quality detection method provided by the embodiment of the present application is shown in the figure.

[0021] Figure 3 A structure diagram of the copper-clad plate copper-clad quality detection system provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0022] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be explained as a limitation on the present application.

[0023] The embodiment of the present application first provides a copper-clad plate copper-clad quality detection method, which can be applied to electronic devices such as computer terminals, specifically, ordinary computers, etc.

[0024] The following will be described in detail taking the running on the computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal of the copper-clad plate copper-clad quality detection method provided by the embodiment of the present application is shown in the figure. Figure 1As shown, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0025] Referring to Figure 2 Embodiments of the present application provide a copper-clad plate copper cladding quality detection method, which can include the following steps:

[0026] S201, apply an ultrasonic excitation signal of a preset frequency and power to the copper-clad plate to be detected, and synchronously use a high-speed camera to collect a dynamic micro-vibration response image sequence of the surface of the copper-clad plate under the excitation of the ultrasonic wave, wherein the ultrasonic frequency is associated with the inherent frequency of the substrate and the copper layer of the copper-clad plate to excite a specific mode of micro-vibration;

[0027] Specifically, the inherent frequency spectrum of the substrate and the copper layer can be calculated based on the material thickness and density parameters of the copper-clad plate through a finite element modal analysis algorithm, and the inherent frequency spectrum is output as a preset frequency candidate set;

[0028] The system first receives the material thickness parameter (Thickness Parameter, TP, unit: millimeter) and density parameter (Density Parameter, DP, unit: gram / cubic centimeter) of the copper-clad plate. These parameters can be obtained by user input or automatic scanning equipment (such as a laser thickness gauge). Taking a typical FR-4 epoxy substrate as an example, the substrate thickness value TP_substrate is 0.2 millimeters, and the copper layer thickness value TP_copper is 0.035 millimeters; the substrate density DP_substrate is 1.8 grams / cubic centimeter, and the copper density DP_copper is 8.9 grams / cubic centimeter. Based on these parameters, a digital mechanical model of the copper-clad plate is constructed using the Finite Element Modal Analysis (FEMA) algorithm. This algorithm discretizes the copper-clad plate into millions of micro-grid elements (such as tetrahedral elements), and each element is assigned mechanical properties such as Elastic Modulus (EM), Poisson's Ratio (PR), etc. according to the material properties. By solving large eigenvalue equations, the free vibration modes of the material in the unconstrained state are simulated. The calculation process is performed in parallel by a high-performance computing cluster, and the output includes the Natural Frequency Spectrum (NFS) containing the dominant vibration modes of the substrate (such as overall bending) and the copper layer (such as local film vibration). The NFS takes frequency values (unit: kilohertz kHz) as the horizontal coordinate and vibration energy ratios (Vibration Energy Ratio, VER) as the vertical coordinate, for example, outputting key peak points such as the first-order bending frequency of the substrate 85.3 kilohertz and the film vibration frequency of the copper layer 217.6 kilohertz. The NFS generated finally serves as the Preset Frequency Candidate Set (PFCS) for subsequent ultrasonic excitation.

[0029] The core technology of finite element modal analysis lies in material layer modeling and boundary condition optimization. The algorithm establishes independent material layers (ML) in the grid model according to the physical division of the substrate and copper layer, and assigns corresponding TP and DP values to each layer. The boundary condition is set to "free-free" (Free-Free Boundary Condition, FFBC), which simulates the real state of the copper-clad plate without fixture constraint in the air. The solver uses the Lanczos iteration method (LIM) to extract the first N (for example, the first 50) order vibration modes. Each mode contains three key data: the natural frequency value F_n (unit kilohertz), the modal shape vector V_m (describing the spatial displacement distribution), and the modal quality factor Q_m (reflecting the energy dissipation characteristic). For composite structures such as copper-clad plates, the algorithm pays special attention to two types of modes:

[0030] Substrate dominant mode: It shows the overall bending or torsion of the substrate, with a frequency range of 50-150 kilohertz (example value: first-order bending mode 85.3 kilohertz, with a maximum displacement at the center of the plate);

[0031] Copper layer dominant mode: It shows local film vibration of the copper layer, with a frequency range of 200-500 kilohertz (example value: copper layer film mode 217.6 kilohertz, with high-frequency micro-amplitude vibration in the copper foil area).

[0032] When outputting PFCS, the system automatically filters out significant modal frequencies with VER greater than 5%, and arranges them in ascending order of frequency value to form a candidate list (for example: {85.3, 127.8, 217.6, 305.2} kilohertz).

[0033] To improve computational efficiency, the algorithm uses adaptive mesh refinement technology (Adaptive Mesh Refinement, AMR). The initial mesh size is set to 1 millimeter, and the edge of the copper layer, holes and other stress concentration areas are automatically refined to 0.1 millimeters. The material constitutive model is selected as the linear elastic model (Linear Elastic Model, LEM), with the substrate elastic modulus EM_substrate set to 22 gigapascals (GPa) and the copper layer EM_copper set to 110 gigapascals. After calculation, the system verifies the reliability of the results through modal assurance criterion (Modal Assurance Criterion, MAC): compare the frequency deviation under adjacent grid densities, if the deviation is less than 1% (for example, 85.3 kilohertz and 85.2 kilohertz), it is considered that the result converges. The final output PFCS strictly corresponds to the physical characteristics of the copper-clad plate, providing accurate frequency targets for subsequent ultrasonic excitation.

[0034] The adaptive resonance tracking algorithm is used to dynamically tune the output frequency of the ultrasonic generator with the preset frequency candidate set, select the frequency point that can stimulate the maximum micro-vibration amplitude, and output the optimized ultrasonic excitation signal.

[0035] The adaptive resonance tracking algorithm takes the PFCS as input and controls the ultrasonic generator to perform frequency scanning. The initial output frequency of the USG is set to the lowest value in the PFCS (e.g., 85.3 kHz), and the output power is fixed at a safety threshold (e.g., 50 W). At the same time, the laser Doppler vibrometer measures the vibration velocity amplitude V_amp (unit: mm / s) of the specified point (usually the center area) on the surface of the copper-clad plate in real time. ARTA scans within a ±5% frequency window (e.g., 85.3±4.3 kHz) with a step size of 10 Hz (Frequency Step Size, FSS). When V_amp is detected to exceed 3 times the background noise (e.g., from 0.01 mm / s to 0.03 mm / s), the algorithm determines that it has entered the resonance region, and then switches to the Golden Section Search (GSS) mode, reducing the step size to 1 Hz for fine scanning.

[0036] In the fine scanning phase, the algorithm records the V_amp value corresponding to each frequency point and constructs a Frequency-Amplitude Response Curve (FARC). The curve is fitted by Cubic Spline Interpolation (CSI), and the global maximum point corresponding to the resonance frequency F_res (example value: 86.7 kHz) is calculated. At this time, the system automatically verifies whether F_res is within the tolerance range (e.g., ±2%) of the theoretical value predicted by PFCS. If the verification is passed (e.g., 86.7 kHz is within 85.3±1.7 kHz), F_res is locked; if it exceeds the tolerance (e.g., 92.0 kHz is detected), the modal re-matching program is triggered: compare F_res with the modal shape of the nearest frequency point in PFCS (e.g., 127.8 kHz), and determine whether they match by the Modal Correlation Coefficient (MCC) (MCC>0.8 is considered to match). After successful matching, update PFCS to a new candidate set centered on F_res.

[0037] After determining F_res, the algorithm further optimizes the ultrasonic power parameter. At a fixed frequency F_res, the USG power is increased from 10 watts to 100 watts in 5-watt steps, while monitoring V_amp and the temperature rise ΔT (unit: Celsius) feedback from the thermal imager. Stop when ΔT exceeds the safety threshold (e.g. 5°C) or V_amp growth saturates (increment less than 1%). Finally, select the power (e.g. 60 watts) that satisfies ΔT ≤ 5°C and V_amp ≥ 80% of the maximum value. Thus far, the optimized ultrasonic excitation signal (OES) is generated, with parameters {frequency: 86.7 kHz, power: 60 watts}.

[0038] According to the periodic characteristics of the optimized ultrasonic excitation signal, configure the frame rate synchronization module of the high-speed camera to ensure that the frame rate is an integer multiple of the excitation frequency, and output a set of synchronization parameters.

[0039] The periodic characteristics of the OES are reflected in the fixed frequency value F_res (86.7 kHz), corresponding to a period T_period = 1 / F_res ≈ 11.53 microseconds. The Frame Rate Synchronization Module (FRSM) needs to make the frame rate F_frame (unit: frames / second) of the high-speed camera satisfy the integer multiple matching principle: F_frame = N × F_res, where N is an integer (usually N ≥ 4). Taking a commonly used camera as an example, if its maximum frame rate is 200,000 frames / second, then calculate the N value that satisfies F_frame ≤ 200,000:

[0040] When N = 4, F_frame = 4 × 86.7 kHz = 346.8 kHz → exceeds the camera's capability; when N = 8, F_frame = 693.6 kHz → still exceeds; when N = 16, F_frame = 1387.2 kHz → exceeds; when N = 32, F_frame = 2774.4 kHz → still exceeds (requires a higher performance camera).

[0041] In practice, a high-speed camera with a frame rate of 500,000 frames / second is selected, so the maximum N = 5 (F_frame = 433.5 kHz). At this time, the sampling theorem needs to be verified: since the highest frequency component of the vibration is about 3 times F_res (260.1 kHz), 433.5 kHz > 2 × 260.1 kHz, satisfying the Nyquist sampling requirement.

[0042] The core of the synchronization parameter configuration is phase locking (Phase Locking). The system aligns the camera exposure signal with the rising edge of the OES through a Phase-Locked Loop (PLL) circuit. Set the parameters:

[0043] Exposure time T_exp: set as T_period / 4≈2.88 microseconds, to ensure capturing key phase points (0°, 90°, 180°, 270°) within a vibration period;

[0044] Trigger delay T_delay: adjusted in 0.1-microsecond steps, to compensate for circuit transmission delay (typical value 0.5 microseconds);

[0045] Sequence length L_seq: covers at least 8 complete vibration periods (L_seq=8×N=8×5=40 frames).

[0046] The output synchronization parameter set (Synchronization Parameter Set, SPS) is: {F_frame: 433.5 kiloframes / second, T_exp: 2.88 microseconds, T_delay: 0.5 microseconds, L_seq: 40 frames}.

[0047] To eliminate timing jitter, a high-precision clock distribution system (Precision Clock Distribution, PCD) is used. The main clock source uses a 100-megahertz temperature-compensated crystal oscillator (TCXO), with a jitter of less than 1 picosecond. The OES signal and the camera trigger signal are transmitted through coaxial cables of equal length, with a path difference controlled within 1 millimeter (a delay difference of about 3.3 picoseconds). The final measured synchronization error is less than 0.1°, ensuring the phase consistency of the image sequence.

[0048] An optimized ultrasonic excitation signal is applied to the copper-clad plate, while a high-speed camera is triggered to capture based on the synchronization parameter set, generating a dynamic micro-vibration response image sequence.

[0049] The ultrasonic excitation subsystem performs OES parameters: a piezoelectric transducer (Piezoelectric Transducer, PZT) is attached to the edge of the copper-clad plate, an 86.7-kilohertz sinusoidal electrical signal is input, and 60-watt acoustic energy is output through a power amplifier. The PZT and the plate are coated with an acoustic coupling agent (silicone grease, impedance matching value Z_match=2.5 MRayl) to reduce energy loss. Ultrasonic waves propagate in the plate to excite specific modes (such as copper layer film vibration), generating micro-vibration displacement D_vib (typical amplitude 0.1-5 microns) on the surface. At the same time, the high-speed camera subsystem works based on the SPS parameters: after the rising edge of the OES signal for 0.5 microseconds (T_delay), the camera starts shooting at a frame rate of 433.5 kiloframes / second, with an exposure time of 2.88 microseconds per frame, and a total of 40 frames are captured.

[0050] The imaging system adopts a coaxial pulsed illumination (CPI) light source with a wavelength of 520 nm and a pulse width of 1 microsecond (synchronized with T exp) to eliminate motion blur. The camera resolution is set to 1280 x 1024 pixels, and the pixel size is 5.5 microns. In combination with a 100 mm micro-lens, the imaging field of view covers the center area (30 x 24 mm) of the copper-clad plate, and the spatial resolution reaches 23 microns / pixel, which can distinguish copper layer micro-cracks (> 50 microns). Each frame of image is stored in a 12-bit grayscale format, and the dynamic range meets the grayscale change (about 20 grayscale levels) corresponding to 0.1 micron displacement.

[0051] Real-time quality monitoring is implemented during the acquisition process: the LDV verifies that the vibration amplitude is stable at 0.25 ± 0.02 mm / s; the thermal imager monitors the temperature rise ΔT ≤ 3℃; the camera built-in image signal-to-noise ratio (SNR) calculation module ensures that SNR > 30 decibels.

[0052] If any index is out of limit (such as SNR < 25 decibels), the system automatically interrupts and restarts the acquisition. The final output is a dynamic vibration image sequence (DVIS) containing 40 frames of 1280 x 1024 pixel images, with a total data volume of about 2.5 gigabytes (GB).

[0053] S202, performing phase-locked analysis on the dynamic micro-vibration response image sequence to extract the amplitude-phase change characteristics of each pixel point in the ultrasonic wave excitation period, and constructing a copper-clad plate dynamic response feature map, wherein the phase-locked analysis is based on a phase reference of the ultrasonic wave excitation signal to separate the micro-vibration signal caused by excitation from environmental noise.

[0054] Specifically, a phase reference timestamp can be extracted from the ultrasonic wave excitation signal to construct a phase reference vector.

[0055] The cornerstone of the entire phase-locked analysis is to obtain a time reference that is strictly synchronized with the ultrasonic excitation signal. The ultrasonic excitation signal is generated by a dedicated ultrasonic generator (UG), which is a continuous or pulsed sinusoidal electrical signal with a specific frequency (e.g. a pre-set optimal frequency point, say 500 kilohertz kHz) and power (e.g. 50 watt W). While this electrical signal drives the ultrasonic transducer to produce mechanical vibrations, its original, un-amplified drive signal waveform (DSW) or a reference clock signal (RCS) that is strictly synchronized with it is output in real-time to a high-precision data acquisition card (DAC). The sampling rate (SR) of the data acquisition card is set very high, e.g. 100 megahertz MHz (i.e. 1 billion points per second), to ensure that the details of the excitation signal can be accurately captured. The system software (or embedded firmware) detects the starting point (i.e. zero-crossing point, ZCP) or peak point (PP) of each complete sinusoidal wave from this high-speed sampled excitation signal waveform. For example, the system can be set to record a timestamp (TS) at each positive slope zero-crossing of the excitation signal. This timestamp is a high-precision absolute time value, usually provided by the system's high-resolution timer (HRT), with a precision up to nanoseconds (e.g. 10 nanoseconds ns). These sequentially recorded timestamps (TS1, TS2, TS3,..., TSN) form the phase reference vector (PRV). Each element in the PRV accurately corresponds to the starting time of a complete cycle of the ultrasonic excitation signal, providing an accurate phase alignment anchor point for all subsequent analysis steps. The length N of this vector depends on the total duration of the acquired dynamic image sequence and the ultrasonic frequency. For example, if the excitation frequency is 500 kHz (period 2 microseconds ps), and the acquisition lasts 0.1 seconds, the PRV will contain approximately 50,000 timestamps.

[0056] The precision of the phase reference vector (PRV) is crucial. Besides relying on high sampling rate (SR) and high resolution timer (HRT), the system also employs digital signal processing techniques to further improve the accuracy of timestamp detection. For example, when detecting the zero crossing point (ZCP), the system does not simply use the sign change of the adjacent two sampling points to make a rough judgment, but combines interpolation algorithm (IA). A commonly used method is three-point linear interpolation or sinusoidal fitting interpolation. Assuming that in the sampling sequence of the data acquisition card (DAC), the value of the Kth sampling point is negative, and the value of the K+1th sampling point is positive, then the true zero crossing point must be between time points K and K+1. By calculating the values of these two sampling points and their time interval, the actual time of the zero crossing point can be calculated more accurately using linear or sinusoidal function model. This process greatly reduces the error (quantization error) introduced by sampling discreteness, so that the timestamp (TS) precision in the PRV is much higher than the sampling interval (for 100MHz SR, the interval is 10ns, and the precision after interpolation can reach 1ns or even higher). In addition, the system will monitor the frequency stability of the ultrasonic generator (UG) output. Although the optimized frequency is preset, there may be slight drift in actual operation. The construction process of PRV itself also implies the measurement of the instantaneous frequency of the actual excitation signal, and these information will be recorded for subsequent analysis to handle possible non-steady state conditions.

[0057] The phase reference vector (PRV) is not just a list of time points, it is also closely related to the triggering and synchronization logic of the whole experiment. As mentioned in claim 2, the frame rate (FR) of the high-speed camera is matched with the frequency of the ultrasonic excitation signal by an integer multiple. This means that before starting to acquire the dynamic micro-vibration response image sequence (DMVRIS), the system has calculated and set the frame rate (FR = 500 kHz * 8 = 4 million frames per second, MFPS) of the high-speed camera precisely according to the optimized ultrasonic frequency (e.g. 500 kHz) and the expected number of sampling points per period (e.g. 8 frames per ultrasonic period). The signal (ETS) that triggers the high-speed camera to start exposure is also sent out precisely by the same master system based on the start timestamp (TS1) in the PRV, ensuring that the exposure start time of the first frame of image is strictly aligned with a certain phase of the excitation signal (such as the start zero-crossing point). After the PRV is constructed, it not only serves as the basis for the subsequent temporal alignment of the image sequence, but also as the verification reference for the time synchronization of the whole system. The system will check whether the exposure time (ET) and readout time (RT) of each camera frame maintain the expected timing relationship with the excitation period and the PRV, ensuring that there is no frame loss or timing drift. This strictly synchronized PRV is the key guarantee for subsequent extraction of weak micro-vibration signals from a strong noise background.

[0058] performing a temporal alignment operation on the dynamic micro-vibration response image sequence based on the phase reference vector, to output an aligned image sequence;

[0059] The dynamic micro-vibration response image sequence (DMVRIS) acquired by the high-speed camera, although the frame rate (FR) is matched with the excitation frequency by an integer multiple, due to the uncertainty of the slight delay in camera exposure and image transmission, and the possible extremely slight frequency jitter (Jitter) of the ultrasonic excitation signal itself, the corresponding ultrasonic excitation phase (Phase) of the image frames arranged in the acquisition order is not strictly aligned. That is, the surface state captured by the Mth frame of image does not accurately correspond to the Kth fixed phase point of the Nth period of excitation signal. The purpose of the temporal alignment operation (TAO) is to use the previously constructed phase reference vector (PRV) to rearrange (or interpolate) each frame of image in the image sequence, so that it accurately corresponds to the surface state at a certain specific phase time (such as the start point, 1 / 4 cycle point, 1 / 2 cycle point, etc. of each period) of the ultrasonic excitation signal. This is the core preprocessing step to achieve high-sensitivity vibration measurement.

[0060] The process of performing temporal alignment operation (TAO) is as follows: The system first needs to determine the target phase points (TPP). Usually, multiple equally spaced phase points within the excitation signal period are chosen as the targets for analysis. For example, if 8 phase states (PS) are desired to be analyzed within each ultrasound period, the target phase points TPP can be defined as: PS0 (0° / start point), PS1 (45°), PS2 (90°), PS3 (135°), PS4 (180°), PS5 (225°), PS6 (270°), PS7 (315°). For each raw frame (RF) in DMVRIS, the system finds the two closest timestamps (e.g. TS_i and TS_i+1) in the phase reference vector (PRV) according to its exposure start time (EST) or exposure midpoint time (EMT), which define a complete ultrasound period. Then, the time offset (TO) of the EST or EMT of the frame relative to the start point TS_i of the period is calculated. Dividing this time offset TO by the actual length of the period (i.e. TS_i+1-TS_i) gives the actual phase angle (APA) of the frame, which usually ranges between 0 and 360 degrees. Next, according to the pre-defined target phase points (TPP, e.g. PS0, PS1,..., PS7), the system needs to find (or construct) a corresponding, phase-accurately aligned image for each TPP. If for a certain TPP (e.g. PS0), there is exactly one raw frame (RF) whose APA is very close to 0° (within a small tolerance range, e.g. ±0.5°), then this frame RF is directly marked as the image belonging to PS0. However, most of the time, the APA of a raw frame does not exactly fall on a pre-defined TPP. In this case, image interpolation (II) techniques are needed. A common approach is to perform linear interpolation in time based on neighboring frames or more advanced motion compensated interpolation. For example, to obtain the image of PS2 (90°), it is found that the APA of the previous frame RF_j is 85° and the APA of the next frame RF_k is 95°. Then, the target image can be seen as a weighted average of RF_j and RF_k. The weights are calculated according to the difference between the target phase (90°) and the phases of RF_j (85°) and RF_k (95°). The smaller the difference, the larger the weight of the frame. Through this interpolation, a surface image that is precisely in time corresponding to the 90° phase point can be "synthesized".This process is performed for each frame in the sequence, and a sub-sequence is constructed for each target phase point (TPP), each sub-sequence containing images at different excitation cycles but the same phase point.

[0061] The aligned image sequence (AIS) is outputted. After the above processing, the original DMVRIS is reorganized into a new sequence AIS. The structure of this new sequence is generally three-dimensional: the first dimension is the index of the target phase point (e.g. PS0 to PS7, a total of 8 phase states), and the second and third dimensions are the spatial coordinates (X, Y) of the image. For each spatial position (X, Y), there is now a series of images captured (or interpolated) at different excitation cycles but the same phase time under each phase state PS. This alignment process effectively decouples the periodic micro-vibration motion caused by ultrasonic excitation, which is synchronized with the excitation signal, from random environmental noise (such as air disturbance, device vibration, electronic noise). Because environmental noise is usually non-periodic or has different frequency components from the excitation frequency, their intensity is randomly distributed between different phase points after time-domain alignment and arrangement according to the excitation phase, and does not show regular changes related to the phase of the excitation signal. The true micro-vibration response, on the other hand, will exhibit a regular, periodic intensity or position change pattern as the phase of the excitation signal changes (0° to 360°). This structured AIS provides a perfect foundation for the next step of analyzing the vibration characteristics (amplitude and phase) of each pixel point with respect to the excitation phase. The system will store the AIS and mark the exact phase state PS corresponding to each image.

[0062] The wavelet packet transform algorithm is used to analyze each pixel point in the aligned image sequence, extract the instantaneous amplitude and phase angle of the micro-vibration dominant frequency band, and output the pixel-level amplitude-phase matrix;

[0063] Now, with the aligned image sequence (AIS), we have a series of intensity values (gray scale values) for each pixel (X, Y) on the copper clad board surface at 8 (or other number) equally spaced phase points (PS0-PS7) of the ultrasonic excitation signal. The time series of intensity values (along the same phase point of different excitation cycles) should be constant in theory, unless there is a defect or stress anomaly that causes the point to vibrate slightly. Micro-vibration will cause the point to move slightly in space (amplitude) and its movement may have a delay (phase angle) relative to the excitation signal. In order to detect such a small, same-frequency vibration signal from the intensity sequence, a powerful tool is needed to analyze the local frequency characteristics of the signal. The wavelet packet transform algorithm (WPTA) is used for this purpose. The wavelet packet transform (WPT) is an extension of the wavelet transform (WT), which provides finer frequency resolution than standard wavelet decomposition, and can decompose the signal into a series of pre-set, narrower frequency bands (FB).

[0064] For each pixel (X, Y) and each phase state PS, the system deals with the Intensity Value Sequence (IVS) of all the image frames in AIS corresponding to PS. The length of this IVS is equal to the number of excitation cycles collected (e.g. 1000 cycles). The system applies WPTA to this IVS. First, a suitable Wavelet Basis Function (WBF) needs to be chosen. Commonly used and suitable for vibration analysis are Daubechies (Db) series (e.g. Db4, Db8) or Symlets (Sym) series (e.g. Sym8) wavelets, which have good time-frequency localization properties. Then, the Decomposition Level (DL) is determined. The choice of DL needs to balance the frequency resolution and the time resolution. For ultrasonic frequencies in the order of a few hundred kHz, with a sampling frequency of the frame rate (e.g. 4 MFPS, i.e. 4e6 Hz), a suitable DL can be in the range of 6 to 8. For example, choosing DL = 7, WPT will decompose the original signal (frequency band 0-2 MHz) into narrower and narrower sub-bands layer by layer. At each layer, each Node represents a specific frequency band range. Finally at the 7th layer, 128 (2^7) sub-band Nodes are obtained. The system identifies which sub-band Node has a Center Frequency (CF) closest to the ultrasonic excitation frequency (e.g. 500 kHz). This frequency band is considered the Micro-Vibration Dominant Frequency Band (MVDFB). WPT not only decomposes the signal into different frequency bands, more importantly, it provides a Complex Value coefficient for each frequency band at each time point (actually each excitation cycle point). This complex coefficient contains the Instantaneous Amplitude (IA) and Instantaneous Phase Angle (IPA) information of the signal in that frequency band at that time point. For the IVS of pixel (X, Y) in phase state PS, the most important result in the WPTA output is the sequence of complex coefficients corresponding to the MVDFB. The Magnitude of this sequence reflects the trend of the micro-vibration amplitude IA of this point in this phase state PS (e.g. whether there is a defect causing abnormal increase or decrease in amplitude), while its Argument reflects the instantaneous phase shift IPA of the micro-vibration relative to the excitation signal (e.g. whether there is a stress concentration causing phase lag).

[0065] In order to obtain a stable estimate of the vibration characteristics of the pixel point (X, Y), the system does not rely on the IA and IPA of a single excitation cycle, but performs statistical analysis on the MVDFB complex coefficient sequence. Generally, the Mean Instantaneous Amplitude (MIA) and the Mean Instantaneous Phase Angle (MIPA) of the sequence in the entire acquisition time period are calculated. The MIA reflects the average strength of the micro-vibration amplitude of the point in a certain phase state PS; the MIPA reflects the average offset of the vibration phase relative to the excitation signal phase. This process needs to be performed independently for each pixel point (X, Y) and each phase state PS (PS0-PS7). Finally, for each spatial position (X, Y), the system will obtain 8 MIA values (corresponding to PS0-PS7) and 8 MIPA values (corresponding to PS0-PS7). Organizing these values for all pixel points forms a pixel-level amplitude-phase matrix (Pixel-Level Amplitude-Phase Matrix, PLAPM). It is a four-dimensional data structure: spatial dimensions X, Y; feature dimensions include MIA and MIPA; phase dimension PS (0-7). The PLAPM is the core data for subsequent analysis, which captures the complete dynamic characteristics of the micro-vibration response of each point on the copper-clad plate surface under ultrasonic excitation as the excitation phase changes. The amplitude information (MIA) can reveal material stiffness, bond strength (small amplitude may indicate good bond and large stiffness; large amplitude may indicate weak bond or cracks), damping characteristics, etc.; the phase information (MIPA) can reveal local stress state (stress concentration may cause phase lag), material inhomogeneity, etc. The advantage of WPTA is its excellent time-frequency localization capability, which can effectively separate the weak vibration signal near the excitation frequency, even if it is submerged in broadband noise.

[0066] By phase coherence filtering algorithm, taking the phase reference vector as the reference, the environmental noise components in the amplitude-phase matrix are suppressed, the pure micro-vibration signal is separated, and the noise-reduced amplitude-phase feature set is output.

[0067] Although time alignment and wavelet packet transform have greatly enhanced the micro-vibration signal, some noise inevitably remains in the pixel-level amplitude-phase matrix (PLAPM). These noises mainly come from: 1) the read noise (RN) and shot noise (SN) of the camera itself; 2) the fluctuation of ambient light; 3) the weak mechanical vibration unrelated to the ultrasonic excitation; 4) the electronic system noise. The core idea of the phase coherence filtering algorithm (PCFA) is to use a key characteristic: the phase (represented as MIPA in PLAPM) of the real, excitation-induced micro-vibration signal should be highly stable in time (i.e. across different excitation cycles) and maintain a fixed relationship with the phase of the excitation signal (defined by the phase reference vector PRV) (i.e. high phase coherence). The phase of the noise-induced signal fluctuation is random and incoherent.

[0068] The process of performing phase coherence filtering (PCFA) is as follows: for each pixel (X, Y) and each phase state PS, we have obtained an average instantaneous phase angle MIPA from the PLAPM. This MIPA is the average of the instantaneous phase angles IPA of the WPT complex coefficient sequence. However, to assess the phase stability (i.e. coherence), we need to examine the instantaneous phase angle IPA in more detail for each excitation period. Recall that, when applying the WPTA, for each pixel (X, Y), each PS, and each excitation period T, the WPT outputs the instantaneous phase angle IPA (X, Y, PS, T) of the MVDFB. The system calculates the standard deviation (SD_IPA) of the values of IPA (X, Y, PS, T) over all excitation periods T. This standard deviation SD_IPA quantitatively measures the degree of fluctuation of the instantaneous phase around its average value MIPA for that point in that phase state PS. The smaller the SD_IPA, the more stable the phase and the higher the coherence; the larger the SD_IPA, the less stable the phase and the more likely it is noise-dominated. At the same time, the system also calculates the ratio or difference between the average instantaneous amplitude MIA (X, Y, PS) of that point in that PS and the average amplitude of all pixels in that PS (or the amplitude of a certain reference region) as an auxiliary amplitude confidence indicator (ACI). The true micro-vibration region should have both amplitude and phase that are spatially smooth, while the noise points are isolated or exhibit salt-and-pepper noise characteristics. Based on SD_IPA and ACI (and possibly other indicators such as the average amplitude MIA itself), the system constructs a phase coherence coefficient (PCC). PCC is a value between 0 and 1, the larger the value, the more likely that the signal of that point in that PS is a true coherent micro-vibration and the less contaminated by noise. The formula for calculating PCC can be a multi-dimensional function, for example PCC = exp(-k1*SD_IPA)*sigmoid(k2*ACI), where k1, k2 are empirical constants and sigmoid is a sigmoid function for normalization. After PCC is calculated, it acts on the original PLAPM data. For each element (MIA (X, Y, PS) and MIPA (X, Y, PS)), it is multiplied by the corresponding PCC (X, Y, PS). That is: Filtered_MIA (X, Y, PS) = MIA (X, Y, PS)*PCC (X, Y, PS); Filtered_MIPA (X, Y, PS) = MIPA (X, Y, PS)*PCC (X, Y, PS).

[0069] More commonly, PCC is mainly used to weight the MIA to suppress noise, while MIPA remains unchanged or is only smoothed for low PCC points. This process significantly suppresses the amplitude contribution of those pixels with unstable phase (high SD_IPA) or abnormally low / high amplitude (low ACI) which are likely to be noise, while retaining the information of pixels with high phase coherence (low SD_IPA) and reasonable amplitude (which are true vibration signals). In essence, PCC acts as an adaptive weight in both spatial and feature dimensions.

[0070] Output the noise-suppressed amplitude-phase feature set (NSAPFS). After PCFA filtering, the original PLAPM is updated. The filtered MIA value (Filtered_MIA) is significantly attenuated or even zeroed in the noise region, while it is basically retained or enhanced in the real micro-vibration region (because the noise is suppressed, the signal-to-noise ratio is relatively improved). The filtered MIPA value (Filtered_MIPA) usually changes little, but the phase value in the low PCC region may be marked as unreliable or replaced by the neighborhood value smoothing. NSAPFS retains the same structure (X, Y, PS, Filtered_MIA, Filtered_MIPA) as PLAPM, but its data quality is significantly improved, and the environmental noise component is effectively suppressed. This feature set is the direct input to build the final dynamic response feature map. The success of PCFA depends on the accurate phase reference (PRV) and time domain alignment (AIS) established in the previous steps, ensuring that only signals strictly synchronized with the excitation can show high coherence. It cleverly uses the periodic characteristics of the signal as a "fingerprint" to distinguish between signals and noise, and is a key step to improve the sensitivity and reliability of defect detection.

[0071] Map the noise-suppressed amplitude-phase feature set to a two-dimensional spatial coordinate to construct the dynamic response feature map of the copper-clad plate.

[0072] The amplitude-phase feature set after noise suppression (NSAPFS) contains the key vibration features of each pixel (X, Y) at 8 (or more) phase states (PS): filtered mean instantaneous amplitude (Filtered_MIA) and filtered mean instantaneous phase angle (Filtered_MIPA). In order to visually visualize the vibration response characteristics of the entire surface of the copper clad laminate and serve as the input of the subsequent deep learning model, it is necessary to map these features back to their corresponding two-dimensional spatial coordinate positions to form an image format that is easy to process, i.e., copper clad laminate dynamic response feature map (CCLDRFM).

[0073] The process of constructing the feature map involves the selection, combination, and visual encoding of features. The system usually does not directly stuff all 16 features (8 PS*[Filtered_MIA+Filtered_MIPA]) into a huge feature map, but will perform meaningful combination or select the most representative feature subset. Common strategies include:

[0074] Amplitude-phase composite map: This is the most intuitive way. Two core images can be created:

[0075] Maximum amplitude map (MAM): For each pixel (X, Y), calculate the maximum value of the Filtered_MIA value at all 8 PS. MAM(X, Y) = max(Filtered_MIA(X, Y, PS0),..., Filtered_MIA(X, Y, PS7)). This map intuitively shows the peak distribution of micro-vibration amplitudes at various locations on the copper clad laminate surface. Areas with defects and micro-cracks often exhibit abnormally high amplitudes (vibrations are intense). Local stress anomaly areas may cause abnormally low (material hardening) or high (local softening) amplitudes.

[0076] Mean phase lag map (MPLM): For each pixel (X, Y), calculate the average value of the Filtered_MIPA value at all 8 PS (note the handling of angle cycles). MPLM(X, Y) = mean(Filtered_MIPA(X, Y, PS0),..., Filtered_MIPA(X, Y, PS7)). This map shows the average phase delay distribution of the vibration response relative to the excitation signal. Stress concentration areas, areas near internal defects of the material may cause changes in wave propagation speed, resulting in phase lag.

[0077] Specific phase point feature map: select the phase point where the vibration is most significant or most sensitive to the specific defect (e.g. PS2-90° or PS4-180°, where the displacement or velocity reaches the extreme value), and plot the Filtered_MIA and Filtered_MIPA maps under this PS separately.

[0078] Feature fusion map: fuse the amplitude and phase information into one pseudo-color map. For example, use Hue to represent MPLM (phase lag), and use Saturation or Value / Intensity to represent MAM (amplitude intensity). In this way, different colors represent different phase lag states, and the lightness or darkness of the color represents the amplitude size.

[0079] Multi-channel feature map: in order to preserve more abundant information for subsequent deep learning models, a multi-channel feature map can be directly constructed. For example, 16 channels can be defined: channels 1-8 correspond to the Filtered_MIA values of PS0-PS7; channels 9-16 correspond to the Filtered_MIPA values of PS0-PS7 (may need to be normalized to 0-255). In this way, each spatial position (X, Y) has a feature vector containing 16 numerical values, and the entire copper-clad plate surface constitutes a 16-channel “image”.

[0080] Regardless of the strategy, the final output of the copper-clad plate dynamic response feature map (CCLDRFM) is an array of each point with rich vibration feature data on a two-dimensional spatial structure (with width W and height H, corresponding to the camera field of view). This feature map is the final result of all the previous complex signal processing steps (phase reference, time domain alignment, wavelet packet analysis, phase coherent filtering). It effectively presents the complex dynamic micro-vibration response of the copper-clad plate surface under ultrasonic excitation (including amplitude and phase information) in the form of a spatial distribution map. In high-quality, defect-free areas, the amplitude map should present uniform, lower values, and the phase map should also be relatively uniform. In poorly bonded areas, due to the debonding or presence of gaps between the copper layer and the substrate, the local stiffness of the area is reduced, and the vibration amplitude under ultrasonic excitation will increase significantly, appearing as bright spots (high values) on the MAM map. At the same time, the reflection and scattering of the debonding interface to the sound wave can cause abnormal phase distribution. In the micro-crack area, the crack tip is a place of high stress concentration and a strong scattering source of waves, which will also cause abnormal increase of amplitude and phase distortion near the crack. Local stress abnormal areas (such as excessive residual stress) will change the local material elastic modulus and sound speed, affect the wave propagation, and cause the amplitude to decrease (hardening) or the phase to shift (lag or advance). CCLDRFM captures these subtle differences in dynamic response related to the internal state of the material and defects, providing high-quality, high-information input data for the next step of automatically identifying and locating defects using a deep learning model. After construction, CCLDRFM will be passed to the defect identification model for further analysis.

[0081] S203, input the copper-clad plate dynamic response feature map into the trained defect identification model, identify and locate potential poorly bonded areas, micro-crack areas and local stress abnormal areas, and output a preliminary defect area distribution map, wherein the defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples;

[0082] Specifically, the copper-clad plate dynamic response feature map can be subjected to multi-channel feature enhancement, and the attention mechanism is used to weight and highlight the abnormal response area, and an enhanced feature map is output.

[0083] When the dynamic response feature map of the copper-clad plate (containing the amplitude and phase information of each pixel) is input, the system first performs multi-channel feature enhancement. The feature map is essentially a two-channel data: the first channel stores the amplitude value (Amplitude Value, AV), reflecting the local vibration intensity of the material; the second channel stores the phase angle (Phase Angle, PA), reflecting the timing relationship of vibration. The enhancement process first uses the contrast limited adaptive histogram equalization (Contrast Limited Adaptive Histogram Equalization, CLAHE) technology to process the two channels respectively. Taking the amplitude channel as an example: the system divides the image into several local blocks (such as 32x32 pixels), calculates the gray level histogram for each block independently, and limits the over-enhancement through a pre-set contrast limit threshold (Clip Limit, CL, typical value 2.0), and then redistributes the pixel values to stretch the local contrast. This operation significantly improves the distinction between weak vibration areas (such as potential micro-crack edges) and strong vibration areas (such as intact copper layers). At the same time, the phase channel is enhanced using the Histogram of Oriented Gradients (Histogram of Oriented Gradients, HOG) enhancement, which calculates the direction gradient of the phase angle in the neighborhood (such as a 5x5 window) of each pixel, enhancing the texture features of the vibration wave propagation direction, which is crucial for identifying the anisotropic response of stress abnormal areas.

[0084] The enhanced two-channel data is input into the channel attention mechanism (Channel Attention Module, CAM). This mechanism first performs global average pooling (Global Average Pooling, GAP) on each channel, compressing the two-dimensional feature map of each channel into a single scalar (i.e. channel importance score). The pooling value of the amplitude channel (denoted as S_AV) reflects the overall vibration energy level, and the pooling value of the phase channel (S_PA) reflects the overall timing consistency. These two scalars are input into a small fully connected network (containing one hidden layer with 16 nodes) to output the weight coefficients (Weight Coefficient, WC_AV and WC_PA, range 0-1) of the two channels. If the current sample S_AV is significantly higher than the mean value of the training set, WC_AV will be close to 1 (emphasizing amplitude abnormalities); if S_PA deviates from the typical value, WC_PA will increase (highlighting phase disorder areas). Finally, the original amplitude map is multiplied by WC_AV, and the phase map is multiplied by WC_PA, achieving channel-level weighting.

[0085] The channel-weighted feature map is further input into a spatial attention mechanism (SAM). The mechanism first concatenates the two-channel data along the depth direction to form a two-dimensional map. By calculating the standard deviation (SD) of each pixel position in the neighborhood (such as a 7x7 window), a spatial saliency map (SSM) is generated. Areas with high standard deviation (such as poorly bonded areas with amplitude mutations) are highlighted in the SSM. After the SSM is normalized to the range of 0-1 by the Sigmoid function, it is multiplied pixel by pixel with the original concatenated map. This focuses the model on the spatial positions of abnormal fluctuations in the vibration response (such as amplitude drops at micro-cracks or phase lags at stress concentration areas). The final output is an enhanced feature map (EFM) that combines the channel and spatial attention, with an approximately 60% improvement in the signal-to-noise ratio of the abnormal area, providing a high-discrimination input for subsequent defect identification.

[0086] The enhanced feature map is input into a pre-trained defect identification model that uses a graph convolutional network architecture, combining dynamic response features and material stress distribution prior knowledge, to output an original defect probability map.

[0087] The enhanced feature map (EFM) is input into the core pre-trained defect identification model. This model uses a graph convolutional network (GCN) architecture, which, unlike traditional convolutional neural networks (CNNs), treats the copper-clad plate as graph structure data in a non-Euclidean space. First, graph construction is performed: each superpixel block (SPB) of the EFM, generated by the SLIC algorithm with a size of approximately 10x10 pixels, is taken as a graph node (ND), and the node feature vector includes five statistical quantities such as the mean amplitude and phase variance within the block. Edges (ED) between nodes are established according to two rules: 1) spatially adjacent superpixels are connected; and 2) based on material stress propagation prior knowledge, if two nodes are located on the same copper foil trace (according to the design topology map) and the included angle between the traces is less than 30 degrees, a weighted edge is added (the edge weight is positively related to the cosine value of the included angle).

[0088] The graph structure data is input into the multi-layer GCN for message passing (MP). Each layer of GCN performs three key operations: 1) neighbor aggregation: for each node, collect the feature vectors of its first-order neighbor nodes, and generate an aggregated vector (AV) through mean pooling; 2) feature transformation: concatenate the node's own features with the AV, and input into a weight matrix (WM, size is the product of input and output dimensions) for linear transformation; 3) nonlinear activation: use the rectified linear unit (ReLU) function to introduce nonlinearity. Through two-layer GCN iteration, the node features can fuse 3-hop neighborhood information. For example, a node located at the corner of a copper foil will have features that include vibration propagation characteristics from straight wire segments (reflecting stress concentration effects) and damping characteristics of adjacent resin regions.

[0089] The last layer of GCN outputs the node-level feature vector, which is input into the classification head (CH). The CH contains a fully connected layer (output dimension three) plus a softmax function, generating the probability distribution (PD) of each node belonging to three types of defects (poor bonding / microcracks / stress anomalies). Subsequently, the mapping from graph to image is performed to restore: according to the correspondence between the superpixel blocks and the original graph, the node probability values are assigned to the corresponding pixels to generate a full-resolution raw defect probability map (RDPM). Each pixel in this map contains three-dimensional probability values (P_bond, P_crack, P_stress), and the sum is 1. For example, the P_crack of a microcrack region can reach 0.95 or more, while the three probabilities of a perfect region are all below 0.1.

[0090] An adaptive threshold segmentation algorithm is applied to process the raw defect probability map, dynamically dividing the defect category boundary according to the probability value, and outputting a binary defect mask map;

[0091] Three types of probability channels of the raw defect probability map (RDPM) are processed respectively. Taking the poor bonding probability channel (P_bond channel) as an example: first, Gaussian adaptive thresholding (GAT) is adopted. The algorithm slides a window (such as 50*50 pixels) on the image, and calculates the local mean (LM) and local standard deviation (LSD) of P_bond in the window. The dynamic threshold (DT) is determined by the formula DT = LM + k*LSD, where k is a sensitivity coefficient (preset value 1.5). If a pixel P_bond > DT, it is temporarily marked as a candidate poor bonding point. This method overcomes the sensitivity problem of global threshold to uneven light (probability value) - automatically increases the threshold to prevent false expansion in the high probability background area (such as existing large bonding defects), and reduces the threshold to increase sensitivity in the low probability area.

[0092] After preliminary segmentation, probability-guided region growing (PGRG) is performed to optimize the boundary. From each candidate point, the eight-neighborhood pixels are checked. If the P_bond value of the neighborhood pixel is less than the tolerance (Tolerance, Tol, set to 0.15) and belongs to the same copper foil region (according to the design topology map), it is included in the current defect region. The growth threshold is dynamically updated during the growth process: the P_bond mean value of the newly added pixels and the region standard deviation are recalculated, and the subsequent neighborhood comparison is based on the updated statistics. This process iterates until no new pixel is added, ensuring that the segmentation boundary is aligned with the probability gradient transition position, avoiding the jagged edges caused by traditional fixed threshold.

[0093] After the above segmentation is completed for the three types of defect channels respectively, conflict resolution and mask synthesis are performed. If a pixel is marked by multiple channels (such as P_bond > 0.8 and P_crack > 0.7), the maximum probability ownership principle (MPO) is used to determine its final category. Then three binary defect mask maps (BDMM) are generated: bonding mask (BM), crack mask (CM), and stress mask (SM). In each mask, the defect area pixel value is one (1), and the background is zero (0). For example, a microcrack with a width of two microns (2 μm) appears as a continuous bright line in the CM.

[0094] Mark the connected components on the binary defect mask and classify them as CCL, micro-crack or stress anomaly. Output the preliminary defect distribution map.

[0095] Perform connected component labeling (CCL) on each binary mask (BM / CM / SM). Use a two-pass algorithm: the first pass traverses the image from left to right and top to bottom, assigning temporary labels to each foreground pixel and recording equivalence relations (e.g. neighboring pixels belong to the same defect); the second pass resolves the equivalence table, merging temporary labels and assigning a unique label (Label ID, LID) to each independent connected component. Also record the geometric properties of each connected component: Bounding Box (BB), Area (A, in pixels), Perimeter (P), Centroid Coordinate (CC), etc. For example, a circular delamination region in BM is labeled as LID_001 with an area of 150 pixels.

[0096] Perform secondary verification of defect types based on geometric and topological features. The system pre-defines a morphology rule base for three types of defects:

[0097] Poorly bonded regions: area threshold (A > 200 pixels) and Form Factor (FF = 4πA / P 2 ) close to 1 (circular feature), mostly located at the edge of the copper layer.

[0098] Micro-crack regions: Aspect Ratio (AR = BB_width / BB_height) greater than 5 and FF less than 0.3, with a strike direction mostly parallel to the stress direction.

[0099] Stress anomaly regions: spatial distribution close to high stress points (e.g. pad corners) in the design topology (distance < 100 μm), and star-shaped radiation pattern (FF ≈ 0.5).

[0100] If the features of a connected component conflict with the mask type (e.g. a circular region in CM), activate the manual rule engine to adjust the type or mark it for review.

[0101] The mark results of the three types of masks are integrated to generate a preliminary defect distribution map (PDDM). The map is an RGB color image: the bonding defects are marked in red (R=255, G=0, B=0), the micro-cracks are marked in blue (R=0, G=0, B=255), and the stress abnormalities are marked in yellow (R=255, G=255, B=0). Each connected domain is labeled with its LID and type code (for example, "C-102" represents the 102nd micro-crack), and the centroid position is superimposed with a cross mark. An accompanying data table is also generated, recording the LID, type, area, position coordinates, minimum bounding rectangle, and other attributes of each defect for subsequent morphological analysis and physical verification. Thus, the preliminary intelligent interpretation of the non-contact ultrasonic vibration detection is completed.

[0102] The trained graph convolution network model recognizes the three types of typical defects by analyzing the vibration amplitude distribution, phase delay pattern, and other characteristics. The model incorporates material mechanics prior knowledge and can distinguish between real defects and small response differences caused by normal process fluctuations, achieving automatic classification and positioning of defects and solving the subjectivity problem of manual interpretation. The non-linear modeling capability of the deep learning model for complex characteristics significantly improves the detection rate of hidden defects (such as incomplete peeling bonding defects).

[0103] S204, performing multi-scale morphological analysis and region connectivity verification on the preliminary defect distribution map, combining the design line topology information of the copper-clad plate, screening out real defects that conform to physical laws, and generating a precise defect region map, wherein the region connectivity verification is used to exclude false positives caused by image noise or non-copper-clad regions;

[0104] Specifically, multi-scale morphological opening and closing operations can be performed on the preliminary defect distribution map, and a variable-scale structure element is used to eliminate isolated noise points, outputting a morphologically corrected map.

[0105] The system receives a Preliminary Defect Region Map (PDRM) generated by the defect identification model. The map is a binary image, where white pixels represent regions labeled as potential defects (including poor bonding, micro-cracks, or stress anomalies), and black pixels represent normal background. Due to potential sensitivity fluctuations in the pre-stage model or image noise interference, the PDRM often contains two types of interference: first, Isolated Noise Points, which are small in size (e.g., 1-5 pixel points) and caused by dust reflection or camera noise; second, Irregular Boundaries, which are caused by threshold segmentation discontinuity. To eliminate these interferences, Multi-scale Morphological Opening-Closing Operations are used. The core operation unit is a Variable-scale Structuring Element (VSSE). The structuring element is essentially a pixel template of a pre-set geometric shape (such as a circle, square), whose size is controlled by a Scale Parameter (SP). The system pre-sets three typical scales: a small scale SP_S (diameter 3 pixels), a medium scale SP_M (diameter 7 pixels), and a large scale SP_L (diameter 15 pixels). First, perform Opening: on the PDRM, use circular structuring elements in the order of SP_S→SP_M→SP_L to perform Erosion followed by Dilation operations. Small-scale opening operations can filter out isolated noise points (such as isolated white pixels within a 3x3 region), medium-scale opening operations smooth small protrusions, and large-scale opening operations preserve the main structure of real defects. Then perform Closing: perform Dilation followed by Erosion operations in the same order, to fill small holes (such as broken pixels inside micro-cracks) and connect adjacent fragmented defect regions (such as discrete spots in stress anomaly areas). The final output is a Morphologically Corrected Map (MCM) with significantly reduced noise and smooth, coherent boundaries.

[0106] The application of the variable scale structuring element needs to dynamically adapt to the defect characteristics. The system has a built-in scale selection strategy (SSS): statistics the area distribution histogram (ADH) of all connected regions in PDRM. If the ADH shows that 80% of the area is less than 10 pixels2, SP_S is preferentially enabled; if there is a macro-defect with an area greater than 100 pixels2, SP_L operation is implemented on it to protect the structure integrity. During the operation process, a progressive erosion-dilation (PED) mechanism is used: for example, when performing the opening operation, first erode the image with SP_S to remove the tiny noise points; then dilate with SP_M to restore part of the structure; finally, slightly erode with SP_L to trim the edge. This mechanism avoids the loss of details caused by a single scale (such as the tip of the micro-crack being smoothed out by a large-scale operation). All morphological operations are implemented through image processing libraries (such as the underlying optimization of the morphologyEx function in OpenCV) to achieve high-speed parallel computing, and it only takes 50 milliseconds (ms) to process a 300mm×300mm copper-clad plate area. The MCM is temporarily stored in the cache area, and its pixel precision remains consistent with the original image (usually 5 microns / pixel), ensuring the spatial resolution of subsequent analysis.

[0107] To verify the effect of morphological processing, the system performs noise suppression ratio evaluation (NSRE). Compare the number of connected regions between PDRM and MCM: if a region exists in PDRM but disappears in MCM, and its area is lower than the preset valid defect area threshold (VDAT, default is 8 pixels2), it is determined as a successfully eliminated noise point. At the same time, monitor the defect contour fidelity (DCF): calculate the area loss rate (ALR) of the real defect in MCM, if ALR exceeds 5% (such as the width of micro-crack is excessively filled due to closing operation), automatically reduce the scale of closing operation or skip the processing of this region. The final output MCM needs to meet the double standards: isolated noise point reduction rate ≥95%, real defect deformation rate ≤3%. This step significantly improves the image quality and lays the foundation for subsequent connectivity analysis.

[0108] Based on the morphologically corrected image, use the region growing algorithm to label the connected defect regions, and output the connected region label map;

[0109] The morphological correction map (MCM) is input to the region growing module (RGM). This module employs a region growing algorithm (RGA) to cluster label the defect regions. The algorithm first scans the entire MCM to locate all foreground pixels (i.e. defect pixels). With each unlabelled foreground pixel as a seed point, it checks its 8-neighborhood pixels: if a neighborhood pixel is also foreground and unvisited, it is included in the current region and recursively expanded as a new seed point. The growth criterion contains two items: one is spatial continuity, which requires the pixel positions to be adjacent; the other is gray-level consistency. Since the MCM is a binary map, all foreground pixels have a gray-level value of 255 (white), so it is only necessary to satisfy the same foreground. The growth process continues until there are no new pixels to be included, at which time an independent connected defect region (CDR) is generated. Each CDR is assigned a unique region label (RL), with the label value starting from 1 and sequentially increasing (e.g. Label_1, Label_2,...). The final output is a connected region label map (CRLM), in which each pixel value represents the label number of the region to which it belongs, and the background pixels are labeled as 0.

[0110] To improve the efficiency of the algorithm, a queue optimization strategy (QOS) is adopted. A first-in-first-out (FIFO) pixel queue (PQ) is maintained. When a new seed point is found, its coordinates (x, y) are added to the tail of the queue. When processing, a pixel is taken from the head of the queue and its neighborhood is checked. New found neighborhood pixels are added to the tail of the queue. To avoid repeated access, a visited matrix (VM) of the same size as the MCM is established to record whether each pixel has been processed. For super large defect regions (area > 10,000 pixels2), block-based parallel processing (BPP) is enabled: the image is divided into several 256x256 pixel sub-blocks, each of which independently performs region growing, and then the boundary pixel matching (BPM) algorithm is used to merge the connected regions across sub-blocks. The label assignment uses the union-find data structure (UFDS) to accelerate region merging: when the boundary regions of two adjacent sub-blocks belong to the same defect, the root label (Root Label) is quickly updated. The entire marking process can be completed within 100 milliseconds (ms) for an image containing 500 defect regions.

[0111] Region validity verification (RVV) is performed before output. The geometric attributes of each CDR are counted, including:

[0112] Area (AR): total number of region pixels; Bounding Box (BB): coordinates of the minimum enclosing rectangle; Centroid (CT): coordinates of the geometric center of the region; Perimeter (PR): length of the boundary pixel chain.

[0113] If the area of a region is less than the minimum valid region threshold (MVRTh = 10 pixels2), or the Perimeter-Area Ratio (PAR) is abnormally high (> 3.0, indicating fragmented noise), its label is reset to 0 (i.e. classified as background). Only the connected regions that meet the physical defect size rule are retained in the final CRLM, along with their geometric attribute table (GAT).

[0114] Load the copper-clad plate design line topology information and construct the line topology graph;

[0115] The system loads the Design Circuit Topology Information (DCTI) from the Engineering Database (EDB). This information is usually derived from Computer-Aided Design (CAD) files (e.g. Gerber or ODB++) and contains three core layers: Copper Layer, Substrate Layer, Solder Mask Layer. Key data includes:

[0116] Trace Geometry Vectors: describe the start / end coordinates, width, and orientation of Copper Traces; Pad Positions: mark the center coordinates and dimensions of component pads; Copper Area Boundaries: define the polygonal outlines of large-area copper regions; Aperture Features: identify the locations of drilled holes or slots.

[0117] Based on this, the Circuit Topology Graph (CTG) is constructed. This graph is a Weighted Undirected Graph, where:

[0118] Nodes: represent electrical connection points (e.g. pad centers, trace intersections); Edges: represent segments of Copper Traces, with weights (Lengths) in microns; Faces: represent closed copper regions, storing their polygonal vertex coordinates.

[0119] The construction process employs a Layered Parsing Algorithm (LPA):

[0120] Geometry Rasterization: converts vector data into a Grid Map with the same resolution as the CRLM, marking copper regions as 1 (conductive) and non-copper regions (substrate, holes) as 0 (insulating).

[0121] Key Point Extraction: uses Corner Detection to locate trace junctions and pad centers as candidate nodes.

[0122] Topology Connection: Establish initial connections between nodes based on Delaunay Triangulation, and then prune redundant edges according to the actual routing path.

[0123] Area Segmentation: Identify closed copper areas through the Scan-line Filling algorithm, and calculate the area and centroid of each region.

[0124] The final output CTG includes: Node List: Each node contains ID, coordinates, and type (pad / via / intersection); Edge List: Each edge contains start and end node IDs, length, and average width; Area List: Each area contains vertex coordinate set, area, and electrical network number.

[0125] To align with the detection image space, perform Coordinate Transformation Calibration (CTC). Pre-set Fiducial Marks (FM) at the four corners of the copper-clad board, with known Design Coordinates (DC) and Image Coordinates (IC) in the high-speed camera image. Solve the Affine Transformation Matrix (ATM) by the Least Squares Method, and convert all node / region coordinates in the CTG to the image coordinate system. The conversion error is controlled within ±2 pixels, ensuring accurate overlap of the topology graph and CRLM.

[0126] Combine the circuit topology graph with the connected region label map to perform physical constraint verification, verify whether the connected region meets the stress propagation model, and exclude false positives in non-copper areas, output the precise positioning defect map after physical verification.

[0127] Overlay the connected region label map (CRLM) and the circuit topology graph (CTG) to perform physical constraint verification (PCV). The core is to verify whether each defect region (CDR) meets the stress propagation model (SPM). This model is based on the propagation characteristics of ultrasonic waves in copper-clad boards:

[0128] Poor Bonding Regions: Located at the copper-substrate interface, stress wave reflection is enhanced here, leading to unusually high amplitude, and the region shape should be parallel to the copper-coated boundary (e.g. a strip-shaped region along the edge of a large copper-coated area).

[0129] Micro-crack Regions: Stress wave propagation is blocked by cracks, showing a sudden drop in amplitude band, and the crack direction should be perpendicular to the principal stress direction (e.g. perpendicular to the high-frequency signal trace).

[0130] Stress Anomaly Regions: Concentrated at geometric mutations (e.g. pad edges, trace corners), showing local high-amplitude spots.

[0131] The verification is divided into three steps:

[0132] Location Compliance Check (LCC): If the CDR is completely located in the non-copper-coated area (e.g. substrate hole or solder mask layer), it is determined as a false detection (e.g. dust attachment) and directly rejected.

[0133] Morphology Conformity Assessment (MCA): Calculate the average distance (ABD) between the CDR contour and the nearest copper boundary. The ABD of the poor bonding area should be less than 5 pixels, and the ABD of the micro-crack area can be larger but needs to have directional consistency.

[0134] Stress Propagation Simulation (SPS): Take the CDR center as the excitation point, and use the finite difference method (Finite Difference Method) to simulate stress wave diffusion. Real defects will cause the correlation coefficient (Correlation Coefficient, CC) between the simulated amplitude map (Simulated Amplitude Map, SAM) and the dynamic response feature map (Dynamic Response Feature Map) in this area to be greater than 0.7.

[0135] Region Connectivity Validation (RCV) is specifically used to exclude two types of false detections:

[0136] Non-copper False Alarms: According to the grid map of CTG, if the Copper Coverage Ratio (CCR) of a CDR < 10% (i.e. > 90% on substrate), it is determined as invalid.

[0137] Cross-region Pseudo-defects: If a single CDR spans multiple electrically isolated copper regions (e.g. two unconnected pads), it violates the continuity principle of force propagation and needs to be split into independent regions for verification.

[0138] The verification is implemented by a Spatial Relation Reasoning Engine (SRRE):

[0139] Input the pixel coordinate set of a CDR in CRLM; query the CTG grid map to count the proportion of copper pixels in the coordinate set (CCR value); if CCR ≥ 90%, perform SPM verification; if CCR ≤ 10%, directly reject; if 10% < CCR < 90%, mark as "suspicious area" for manual review.

[0140] For CDRs that pass the CCR check, calculate the Angle to Traces (ATT) of their bounding rectangles and adjacent copper traces. The ATT of micro-cracks should be close to 90° (perpendicular), and the ATT of poor combinations should be close to 0° (parallel).

[0141] Finally, generate a Precisely Localized Defect Region Map (PLDRM). This map retains all CDRs that pass physical verification and labels the types:

[0142] Poor combination area: red fill with interface direction arrow; micro-crack area: blue outline with crack length label; stress anomaly area: yellow spot with stress concentration coefficient label.

[0143] At the same time, output a Defect Attribute Table (DAT) containing the Label ID, Defect Type, Centroid Coordinates, Area, Equivalent Diameter, and Distance to Critical Trace of each CDR.

[0144] S205, based on the precisely located defect region map, calculating the area proportion, distribution density, and distance to the adjacent key line of each type of defect, and generating a comprehensive evaluation report of the copper cladding quality according to a preset quality scoring rule, wherein the comprehensive evaluation report includes defect type statistics, location labeling, and quality grade determination.

[0145] Specifically, the precisely located defect region map can be parsed, the geometric properties of each defect region are extracted, the area proportion and distribution density are calculated, and the defect measurement data set is output.

[0146] The system receives the precisely located defect region map (PLDRM) generated in the previous step. The map is in a binary image format, where white pixels represent real defect regions that have passed physical verification, and black pixels represent normal regions. Each defect region is labeled with a type label (such as a bonding failure label BFL, a microcrack label MCL, and a stress anomaly label SAL). The parsing process first traverses the image pixels using a connected component labeling algorithm (CCLA) to assign a unique identifier (Defect ID, DID) to each independent defect region. Then, for each pixel cluster corresponding to a DID, the system calls a geometric attribute extraction engine (GAEE). The engine calculates the core geometric properties: the defect area (DA) is calculated by counting the total number of pixels under the DID and multiplying by a preset spatial resolution parameter (SRP, unit: microns / pixel) to convert the pixel value to the actual physical area (unit: square millimeters); the defect perimeter (DP) is calculated by a boundary tracking algorithm; and the minimum bounding rectangle (MBR) records the orientation and extension range of the defect. For example, a microcrack region with DID 103 has a DA of 0.15 square millimeters, a DP of 1.8 millimeters, and an MBR size of 0.3 millimeters x 0.5 millimeters.

[0147] The calculation of the area percentage (AP) needs to combine the total detection area of the copper-clad plate. The system reads the total inspected area (TIA) from the detection configuration file, which is the total area of the effective detection area (unit: square centimeter). For each type of defect (such as all MCL type defects), the AP value is equal to the sum of the DA of all defects of this type divided by the TIA and multiplied by one hundred. For example, if 5 microcracks are detected on the whole board, the total DA is 0.75 square millimeters, and the TIA is 200 square centimeters (i.e. 20000 square millimeters), then the AP of the microcracks is (0.75 / 20000) x 100% = 0.00375%. The distribution density (DD) reflects the number of defects per unit area. The system divides the copper-clad plate into virtual grid cells (GC), and the grid size is set according to the process requirements (such as 10 mm x 10 mm). The number of specific type defects in each GC is calculated, and then divided by the area of the GC to obtain the local density value. Finally, the global DD output of this type of defect is the arithmetic mean of the densities of all non-empty GCs. For example, in 200 grids, there are 30 grids containing microcracks, and the total number of microcracks is 50. The average density of each grid containing cracks is 50 / 30 / 1 square centimeter = 1.67 per square centimeter. All geometric properties (DA, DP, MBR) and derived indicators (AP, DD) are stored according to DID and defect type, forming a structured defect metric dataset (DMD).

[0148] To ensure the engineering significance of the metric data, the system performs data validity check (DVC). The check rules include: area rationality (such as DA cannot be greater than 1 square centimeter to prevent mislabeling large areas), position boundary constraint (defect coordinates need to be within the TIA range), and type consistency (all pixel labels in the same DID must be consistent). If abnormal data is detected (such as a BFL type defect with a DA of 5 square centimeters), the anomaly tracing mechanism (ATM) is triggered to automatically retrieve the original vibration feature map and intermediate processing results of the area for manual review. The DMD that passes the check is output to the downstream module in a standardized format (such as JSON or XML), and its data structure includes fields such as defect ID, type, center coordinates, area, perimeter, bounding rectangle, area percentage, and distribution density, providing a quantitative basis for subsequent risk assessment.

[0149] Based on the defect metric dataset and the design line topology information, the nearest neighbor algorithm is used to calculate the minimum Euclidean distance from the defect area to the key line, and output the distance risk indicator set;

[0150] System synchronously loads Design Circuitry Topology Information (DCTI). This information is derived from Computer Aided Design (CAD) files, stored in vector graphics format, containing the geometric paths (composed of line segments, circular arcs, etc.) of all conductive copper circuits and Critical Trace Markers (CTM). Critical traces usually refer to high-frequency signal lines, power supply trunks, or impedance control lines, etc. which are sensitive to defects, with special attribute labels (such as Critical_Level = High) in DCTI. The system first performs Rasterization Preprocessing (RP) on the vector circuit data, converting it into a binary image (circuit area is white, non-circuit area is black) that is spatially aligned with the Precise Location Defect Region Map (PLDRM) and has the same resolution. At the same time, it extracts the pixel coordinate set of all critical traces, generating the Critical Trace Coordinate Set (CTCS).

[0151] The Minimum Euclidean Distance (MED) calculation uses the Spatial Index Accelerated Nearest Neighbor Search (SIANNS) algorithm. The specific process is as follows:

[0152] Constructing spatial index: Use the k-Dimensional Tree (kD-Tree) data structure to establish a spatial index for the critical trace pixel coordinates in CTCS. The kD-Tree recursively partitions the space (such as alternating division along the X / Y axis), significantly improving the efficiency of nearest neighbor queries.

[0153] Traversing defect regions: For each defect DID in the Defect Metric Dataset (DMD), obtain its Defect Boundary Coordinate Set (DBCS).

[0154] Parallel distance calculation: For each boundary point in DBCS, quickly find the nearest critical trace point (Critical Trace Point, CTP) in CTCS through the kD-Tree index, and calculate the Euclidean distance (Euclidean Distance, ED = √[(x1-x2) 2 +(y1-y2) 2 ]) between the two points.

[0155] Determination of the minimum distance: compare the ED values of all boundary points of the defect, take the minimum value as the MED of the defect to the key line. For example, the ED of the boundary point P1 (10.2, 15.7) of a microcrack to the nearest CTP (10.5, 16.0) is √[(0.3) 2 +(0.3) 2 If the distances of other boundary points are all greater than this value, then the MED of the defect is 0.42 mm.

[0156] The calculated MED needs to be converted into the risk level (Risk Level, RL) in combination with the process specification. The system presets distance threshold parameters (Distance Threshold Parameters, DTP), for example: high risk (High Risk, HR): MED≤0.1 mm; medium risk (Medium Risk, MR): 0.1 mm

[0157] The MED value of each defect and its corresponding RL are recorded. The final output distance risk indicator set (Distance Risk Indicator Set, DRIS) includes: defect ID, nearest key line ID, minimum Euclidean distance, risk level. This data set directly reflects the potential threat degree of the defect to the circuit function reliability, and is the key input of quality scoring.

[0158] The defect metric data set and the distance risk indicator set are input into the quality scoring engine, and the preset rules are weighted and fused to generate the original quality score vector;

[0159] The core of the quality scoring engine (Quality Scoring Engine, QSE) is a multi-factor weighted fusion model (Multi-Factor Weighted Fusion Model, MFWFM). This model predefines scoring factors (Scoring Factors, SF), weight coefficients (Weight Coefficients, WC) and normalization rules (Normalization Rules, NR). The main scoring factors include:

[0160] Area Factor (AF): based on Area Proportion (AP), the larger AP, the more points deducted; Density Factor (DF): based on Defect Distribution (DD), high density indicates that the problem is concentrated; Risk Factor (RF): based on Risk Level (RL), high-risk defects have the highest deduction weight; Type Factor (TF): based on the inherent hazard of defect types (e.g., micro-crack TF > poor bonding TF > stress anomaly TF). The weight coefficient WC is determined in advance by Analytic Hierarchy Process (AHP), for example: RF weight 0.4, AF weight 0.3, DF weight 0.2, TF weight 0.1, and the total is 1.

[0161] The engine performs the following steps to generate the score:

[0162] Factor normalization: convert raw data of different dimensions to [0, 1] interval scores. For example, AP values are converted by a Linear Normalization Function (LNF): AF_score = 1-(AP / AP_max), where AP_max is the maximum allowed area proportion (e.g., 0.01%). Factor weighting: calculate the weighted score according to the weight. For example, if the AF_score of a certain defect is 0.7, then its AF contribution score = 0.7x0.3 = 0.21. Defect individual score: for each defect, calculate the weighted sum of all factors:

[0163] Defect_Score = (AF_scorexWC_AF) + (DF_scorexWC_DF) + (RF_scorexWC_RF) + (TF_scorexWC_TF). Where. RF_score is mapped according to risk level: HR→0, MR→0.5, LR→1.0; TF_score is mapped according to type: micro-crack→0, poor bonding→0.6, stress anomaly→1.0.

[0164] Overall raw score: calculate the overall board quality raw score. First, find the average of all defect Defect_Score, then subtract the value by 1 (the more defects and the more serious, the lower the total score): Raw_Quality_Score = 1-(∑Defect_Score / N). N is the total number of defects, the value range is [0, 1], the closer to 1 indicates the better quality.

[0165] The Raw Quality Score Vector (RQSV) contains multi-level score results: Overall_Raw_Score; average score of each defect type (e.g. Bonding_Avg_Score, Crack_Avg_Score); zone score of each spatial partition (e.g. Zone1_Score, Zone2_Score...). For example, a copper clad laminate has an overall raw score of 0.82, an average score of micro-cracks of 0.65 (poor), and an average score of bad bonding of 0.88 (good). This vector will be used as the direct basis for quality grade determination.

[0166] Based on the Raw Quality Score Vector, a comprehensive evaluation report is generated, including Defect Type Statistics Table, Defect Location Annotation Map, and Quality Grade Determination Label.

[0167] The Defect Type Statistics Table (DTST) is a structured data table, automatically generated from the Defect Metric Dataset (DMD) and the Distance Risk Indicator Set (DRIS). The table columns include: Type; Count; TotalAP; Avg DD; High-Risk Count (i.e. the number of RL=HR); Avg MED. This table provides a global view of defect distribution, facilitating quick identification of major problem types.

[0168] The Defect Location Annotation Map (DLAM) is the core visual output:

[0169] Base map synthesis: The original copper clad laminate appearance map, the designed circuit topology map, and the precisely located defect region map (PLDRM) are semi-transparently overlaid to form a comprehensive base map. Defect labeling: Each defect region is marked with a colored contour line (e.g. micro-cracks - red, bad bonding - yellow, stress anomaly - blue); the defect ID is labeled inside the contour (e.g. Crack_103); a lead line is drawn from the center of the defect, and the end is connected to an information callout box (ICB) showing the DID, type, area, risk level, distance to the nearest key circuit, etc. Hot spot highlighting: For high-risk defects (RL=HR), the contour flashes and adds a warning symbol.

[0170] Partition identification: If partition scoring is used, the area boundaries are divided by dashed lines and labeled with area numbers. This map is output in high-resolution bitmap or vector (SVG / PDF) format, supporting zoom-in to view details.

[0171] Quality Grade Tag (QGT) is generated from Raw Quality Score Vector (RQSV):

[0172] Grade Rule Base (GRB) predefines binning thresholds: Grade A: Overall_Raw_Score ≥ 0.90; Grade B: 0.80 ≤ Overall_Raw_Score < 0.90; Grade C: 0.70 ≤ Overall_Raw_Score < 0.80; Grade D: Overall_Raw_Score < 0.70.

[0173] Constraint Check (CC): Even if the overall score is high, if there are high-risk defects (such as HR number ≥ 3) or critical area defects (such as cracks under CPU socket), it is forced to downgrade (such as A→B).

[0174] Tag Generation: The system outputs the final grade tag (such as "Grade B") according to Overall_Raw_Score and constraint check results, and attaches key basis explanation (such as "total score 0.87, but due to the existence of 2 high-risk microcracks, it does not reach A level"). All contents (statistical table, position map, grade tag) are automatically integrated into a comprehensive evaluation report (Comprehensive Evaluation Report, CER), and the output format can be selected as PDF / HTML, supporting one-key export to the production management system (MES).

[0175] Another embodiment of the present application provides a copper-clad plate copper-clad quality detection system, referring to Figure 3 , the system can include:

[0176] The acquisition module 301 is configured to apply an ultrasonic excitation signal of a preset frequency and power to the copper-clad plate to be tested, and synchronously collect a dynamic micro-vibration response image sequence of the surface of the copper-clad plate under ultrasonic excitation by using a high-speed camera, wherein the ultrasonic frequency is associated with the natural frequency of the copper-clad plate substrate and the copper layer to excite micro-vibration of a specific mode.

[0177] The analysis module 302 is configured to perform phase-locked analysis on the dynamic micro-vibration response image sequence, extract amplitude-phase variation characteristics of each pixel point within an ultrasonic excitation cycle, and construct a dynamic response feature map of the copper-clad plate, wherein the phase-locked analysis is based on a phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by excitation from environmental noise.

[0178] The recognition module 303 is configured to input the dynamic response feature map of the copper-clad plate into a trained defect recognition model, recognize and locate potential poor copper-clad bonding areas, micro-crack areas and local stress abnormal areas, and output a preliminary defect area distribution map, wherein the defect recognition model is a deep learning model trained based on dynamic response feature maps of known defect samples.

[0179] The verification module 304 is configured to perform multi-scale morphological analysis and region connectivity verification on the preliminary defect area distribution map, filter out real defect areas conforming to physical laws in combination with design line topological information of the copper-clad plate, and generate a precise defect area map, wherein the region connectivity verification is configured to exclude false positives caused by image noise or non-copper-clad areas.

[0180] The generation module 305 is configured to calculate area proportions, distribution densities and distances from key lines of various defects based on the precise defect area map, and generate a comprehensive evaluation report of copper-clad quality according to a preset quality score rule, wherein the comprehensive evaluation report includes defect type statistics, location labeling and quality grade determination.

[0181] The above embodiments according to the drawings illustrate the structure, features and effects of the present application. The above description is only a preferred embodiment of the present application, but the present application is not limited by the drawings. Any changes or modifications made in accordance with the concept of the present application, or equivalent embodiments with equivalent changes, are still within the scope of the present application.

Claims

1. A method for detecting copper-clad mass of copper-clad plate, characterized in that, The method comprises: applying an ultrasonic excitation signal of a preset frequency and power to a copper-clad plate to be tested, and synchronously using a high-speed camera to collect a dynamic micro-vibration response image sequence of the surface of the copper-clad plate under ultrasonic excitation, wherein the ultrasonic frequency is associated with the natural frequency of the substrate and the copper layer of the copper-clad plate to excite specific mode micro-vibration; performing phase-locked analysis on the dynamic micro-vibration response image sequence to extract the amplitude-phase variation characteristics of each pixel point in the ultrasonic excitation cycle, and constructing a copper-clad plate dynamic response feature map, wherein the phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by excitation from environmental noise; inputting the copper-clad plate dynamic response feature map into a trained defect recognition model to identify and locate potential copper-clad bonding failure areas, micro-crack areas and local stress abnormal areas, and outputting a preliminary defect area distribution map, wherein the defect recognition model is a deep learning model trained based on the dynamic response feature map of known defect samples; performing multi-scale morphological analysis and region connectivity verification on the preliminary defect area distribution map, combining the design circuit topology information of the copper-clad plate, and screening out real defect areas that conform to physical laws to generate a precisely positioned defect area map, wherein the region connectivity verification is used to exclude false positives caused by image noise or non-copper-clad areas; based on the precisely positioned defect area map, calculating the area proportion, distribution density and distance from the key circuit of each type of defect, and generating a comprehensive evaluation report of copper-clad quality according to a preset quality scoring rule, wherein the comprehensive evaluation report includes defect type statistics, location labeling and quality level determination.

2. The method of claim 1, wherein, The method comprises: applying an ultrasonic excitation signal of a preset frequency and power to a copper-clad plate to be tested, and synchronously using a high-speed camera to collect a dynamic micro-vibration response image sequence of the surface of the copper-clad plate under ultrasonic excitation, wherein the ultrasonic frequency is associated with the natural frequency of the substrate and the copper layer of the copper-clad plate to excite specific mode micro-vibration, comprising: based on the material thickness and density parameters of the copper-clad plate, calculating the natural frequency spectrum of the substrate and the copper layer through a finite element modal analysis algorithm, and outputting the natural frequency spectrum as a preset frequency candidate set; using the preset frequency candidate set, dynamically tuning the output frequency of the ultrasonic generator using an adaptive resonance tracking algorithm, selecting a frequency point that can excite the maximum micro-vibration amplitude, and outputting an optimized ultrasonic excitation signal; according to the periodic characteristics of the optimized ultrasonic excitation signal, configuring a frame rate synchronization module of the high-speed camera to ensure that the frame rate matches an integer multiple of the excitation frequency, and outputting a synchronization parameter set; 3. The method of claim 2, wherein, applying the optimized ultrasonic excitation signal to the copper-clad plate, and simultaneously triggering the high-speed camera to collect based on the synchronization parameter set to generate a dynamic micro-vibration response image sequence. The method comprises: performing phase-locked analysis on the dynamic micro-vibration response image sequence to extract the amplitude-phase variation characteristics of each pixel point in the ultrasonic excitation cycle, and constructing a copper-clad plate dynamic response feature map, wherein the phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by excitation from environmental noise, comprising: extract a phase reference timestamp from the ultrasonic excitation signal, and construct a phase reference vector; perform a time domain alignment operation on the dynamic micro-vibration response image sequence based on the phase reference vector, and output an aligned image sequence; extract the instantaneous amplitude and phase angle of the micro-vibration dominant frequency band of each pixel point in the aligned image sequence by using a wavelet packet transform algorithm, and output a pixel-level amplitude-phase matrix; suppress the environmental noise components in the amplitude-phase matrix by using a phase coherence filtering algorithm with the phase reference vector as a reference, separate a pure micro-vibration signal, and output a noise-suppressed amplitude-phase feature set; map the noise-suppressed amplitude-phase feature set to a two-dimensional spatial coordinate to construct a copper-clad plate dynamic response feature map.

4. The method of claim 3, wherein, input the copper-clad plate dynamic response feature map into a trained defect identification model to identify and locate potential copper-clad bonding failure areas, micro-crack areas, and local stress abnormal areas, and output a preliminary defect area distribution map, wherein the defect identification model is a deep learning model trained based on dynamic response feature maps of known defect samples, and includes: perform multi-channel feature enhancement on the copper-clad plate dynamic response feature map, highlight abnormal response areas by using an attention mechanism, and output an enhanced feature map; input the enhanced feature map into a pre-trained defect identification model, which uses a graph convolution network architecture to fuse dynamic response features and material stress distribution prior knowledge, and output an original defect probability map; process the original defect probability map by using an adaptive threshold segmentation algorithm, dynamically divide defect category boundaries according to probability values, and output a binary defect mask map; label connected regions on the binary defect mask map, and classify them into copper-clad bonding failure, micro-crack, or stress abnormality, and output a preliminary defect area distribution map.

5. The method of claim 4, wherein, perform multi-scale morphological analysis and region connectivity verification on the preliminary defect area distribution map, combine the design line topology information of the copper-clad plate, filter out real defect areas that meet physical laws, and generate a precisely located defect area map, wherein the region connectivity verification is used to exclude false detections caused by image noise or non-copper-clad areas, and includes: perform multi-scale morphological opening and closing operations on the preliminary defect area distribution map, use a variable-scale structure element to eliminate isolated noise points, and output a morphological correction map; based on the morphological correction map, label connected defect areas by using a region growing algorithm, and output a connected region label map; load the design line topology information of the copper-clad plate to construct a line topology map; combine the line topology map to perform physical constraint verification on the connected region label map, verify whether the connected regions meet a stress propagation model, exclude non-copper-clad area false detections, and output a physically verified precisely located defect map.

6. The method of claim 5, wherein, based on the precisely located defect area map, calculate the area ratio, distribution density, and distance from nearby key lines of each type of defect, and generate a comprehensive evaluation report of copper-clad quality according to a preset quality score rule, wherein the comprehensive evaluation report includes defect type statistics, location labeling, and quality level determination, and includes: analyze the precisely located defect area map, extract the geometric properties of each defect area, calculate the area ratio and distribution density, and output a defect metric data set; Based on the defect metric dataset and the design line topology information, the nearest neighbor algorithm is used to calculate the minimum Euclidean distance from the defect area to the key line, and output the distance risk index set; The defect metric dataset and the distance risk index set are input into the quality score engine, and the preset rules are weighted and fused to generate an original quality score vector; According to the original quality score vector, a comprehensive evaluation report including a defect type statistical table, a location annotation map and a quality level judgment label is generated.

7. A copper-clad plate copper-clad quality detection system, characterized in that, The system comprises: The acquisition module is configured to apply an ultrasonic excitation signal of a preset frequency and power to the copper-clad plate to be tested, and synchronously use a high-speed camera to collect a dynamic micro-vibration response image sequence of the surface of the copper-clad plate under the ultrasonic excitation, wherein the ultrasonic frequency is associated with the inherent frequency of the substrate and the copper layer of the copper-clad plate to excite micro-vibration of a specific mode; The analysis module is configured to perform phase-locked analysis on the dynamic micro-vibration response image sequence to extract amplitude-phase variation characteristics of each pixel point within an ultrasonic excitation cycle, and construct a dynamic response feature map of the copper-clad plate, wherein the phase-locked analysis is based on a phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by the excitation from environmental noise; The identification module is configured to input the dynamic response feature map of the copper-clad plate into a trained defect identification model to identify and locate potential poor bonding areas, micro-crack areas and local stress abnormal areas of the copper-clad plate, and output a preliminary defect area distribution map, wherein the defect identification model is a deep learning model trained based on dynamic response feature maps of known defect samples; The verification module is configured to perform multi-scale morphological analysis and region connectivity verification on the preliminary defect area distribution map, and filter out real defect areas that conform to physical laws in combination with design line topology information of the copper-clad plate to generate an accurately positioned defect area map, wherein the region connectivity verification is used to exclude false positives caused by image noise or non-copper-clad areas; The generation module is configured to calculate the area proportion, distribution density and distance from key lines of each type of defect based on the accurately positioned defect area map, and generate a comprehensive evaluation report of the copper-clad quality according to a preset quality score rule, wherein the comprehensive evaluation report includes defect type statistics, location annotation and quality level judgment.

8. The system of claim 7, wherein, The acquisition module is specifically configured to: Based on the material thickness and density parameters of the copper-clad plate, calculate the inherent frequency spectrum of the substrate and the copper layer by a finite element modal analysis algorithm, and output the inherent frequency spectrum as a preset frequency candidate set; Use the preset frequency candidate set to dynamically tune the output frequency of the ultrasonic generator by using an adaptive resonance tracking algorithm, select a frequency point that can excite the maximum micro-vibration amplitude, and output an optimized ultrasonic excitation signal; According to the periodic characteristics of the optimized ultrasonic excitation signal, configure a frame rate synchronization module of the high-speed camera to ensure that the frame rate is an integer multiple of the excitation frequency, and output a synchronization parameter set; Apply the optimized ultrasonic excitation signal to the copper-clad plate, and trigger the high-speed camera to collect based on the synchronization parameter set to generate a dynamic micro-vibration response image sequence.

9. A storage medium, characterized by The storage medium has stored therein a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that, The memory has stored therein a computer program, and the processor is configured to execute the computer program to execute the method of any one of claims 1-6.

Images