A copper-clad plate copper cladding quality detection method and system

Abstract

The application discloses a copper-clad plate copper-clad quality detection method and system, and the method comprises the following steps: applying an ultrasonic excitation signal with a preset frequency and power to a to-be-detected copper-clad plate, and synchronously collecting a dynamic micro-vibration response image sequence of the copper-clad plate surface under ultrasonic excitation by using a high-speed camera; performing phase-locked analysis on the dynamic micro-vibration response image sequence, and constructing a copper-clad plate dynamic response feature map; inputting the copper-clad plate dynamic response feature map into a trained defect recognition model, and outputting a preliminary defect region distribution map; performing multi-scale morphological analysis and region connectivity verification on the preliminary defect region distribution map, and generating an accurate positioning defect region map; and generating a comprehensive evaluation report of the copper-clad quality based on the accurate positioning defect region map and according to a preset quality scoring rule. By using the embodiment of the application, high-precision positioning and classification of the copper-clad plate defects can be realized, and an efficient and reliable automatic solution is provided for copper-clad plate quality evaluation.

Description

Technical Field

[0001] This invention belongs to the field of quality inspection technology, and in particular to a method and system for inspecting the copper clad laminate quality. Background Technology

[0002] As the core substrate of printed circuit boards (PCBs), copper-clad laminates (CCLs) directly affect circuit performance and reliability due to the bonding quality between the copper layer and the substrate. Traditional inspection methods, such as visual inspection, electrical performance testing, or X-ray scanning, are limited by low efficiency, high cost, or difficulty in detecting micron-level defects. Especially for latent defects such as poor copper bonding and microcracks, current technologies struggle to achieve high-precision, non-destructive, and rapid localization. While vibration-based non-destructive testing technologies have seen some development in recent years, they are limited by environmental noise interference and signal resolution capabilities, making it difficult to effectively separate minute vibration characteristics under complex operating conditions. Furthermore, conventional image processing algorithms lack the ability to extract features from the dynamic response of CCLs, resulting in a high false detection rate. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for inspecting the copper clad laminate quality, so as to overcome the shortcomings of the prior art, achieve high-precision positioning and classification of copper clad laminate defects, and provide an efficient and reliable automated solution for copper clad laminate quality assessment.

[0004] One embodiment of this application provides a method for inspecting the copper clad laminate quality, the method comprising:

[0005] An ultrasonic excitation signal with a preset frequency and power is applied to the copper-clad laminate under test, and a high-speed camera is used to simultaneously acquire a sequence of dynamic micro-vibration response images of the surface of the copper-clad laminate under ultrasonic excitation. The ultrasonic frequency is related to the natural frequency of the copper-clad laminate substrate and the copper layer to excite micro-vibration of a specific mode.

[0006] Phase-locked analysis is performed on the dynamic micro-vibration response image sequence to extract the amplitude-phase change features of each pixel within the ultrasonic excitation cycle, and a dynamic response feature map of the copper-clad laminate is constructed. The phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by the excitation from the environmental noise.

[0007] The dynamic response feature map of the copper-clad laminate is input into the trained defect identification model to identify and locate potential poor copper-clad laminate bonding areas, microcrack areas and local stress anomaly areas, and output a preliminary defect area distribution map. The defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples.

[0008] Multi-scale morphological analysis and regional connectivity verification are performed on the preliminary defect area distribution map. Combined with the design circuit topology information of the copper-clad laminate, the real defect areas that conform to physical laws are screened out to generate a precise defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad areas.

[0009] Based on the precisely located defect area map, the area ratio, distribution density, and distance to critical paths of various defects are calculated, and a comprehensive evaluation report on copper cladding quality is generated according to the preset quality scoring rules. The comprehensive evaluation report includes defect type statistics, location marking, and quality level determination.

[0010] Another embodiment of this application provides a copper clad laminate copper clad laminate quality inspection system, the system comprising:

[0011] The acquisition module is used to apply an ultrasonic excitation signal with a preset frequency and power to the copper-clad laminate under test, and simultaneously use a high-speed camera to acquire a sequence of dynamic micro-vibration response images of the surface of the copper-clad laminate under ultrasonic excitation. The ultrasonic frequency is related to the natural frequency of the copper-clad laminate substrate and the copper layer to excite micro-vibration in a specific mode.

[0012] The analysis module is used to perform phase-locked analysis on the dynamic micro-vibration response image sequence, extract the amplitude-phase change features of each pixel in the ultrasonic excitation cycle, and construct a dynamic response feature map of the copper-clad laminate. The phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by the excitation from the environmental noise.

[0013] The identification module is used to input the dynamic response feature map of the copper-clad laminate into the trained defect identification model, identify and locate potential poor copper-clad bonding areas, microcrack areas and local stress anomaly areas, and output a preliminary defect area distribution map. The defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples.

[0014] The verification module is used to perform multi-scale morphological analysis and regional connectivity verification on the preliminary defect area distribution map. Combined with the design circuit topology information of the copper-clad laminate, it filters out the real defect areas that conform to physical laws and generates a precise location defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad areas.

[0015] The generation module is used to calculate the area ratio, distribution density, and distance to critical paths of various defects based on the precisely located defect area map, and generate a comprehensive evaluation report of copper cladding quality according to preset quality scoring rules. The comprehensive evaluation report includes defect type statistics, location marking, and quality level determination.

[0016] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0017] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0018] Compared with existing technologies, the present invention provides a method for inspecting the copper clad laminate (CCL) quality. This method applies an ultrasonic excitation signal of preset frequency and power to the CCL under test, and simultaneously uses a high-speed camera to acquire a sequence of dynamic micro-vibration response images of the CCL surface under ultrasonic excitation. Phase-locked analysis is performed on the dynamic micro-vibration response image sequence to construct a dynamic response feature map of the CCL. This dynamic response feature map is then input into a trained defect identification model to output a preliminary defect region distribution map. Multi-scale morphological analysis and region connectivity verification are performed on the preliminary defect region distribution map to generate a precisely located defect region map. Based on the precisely located defect region map, a comprehensive evaluation report of the CCL quality is generated according to preset quality scoring rules. This method enables high-precision location and classification of CCL defects, providing an efficient and reliable automated solution for CCL quality assessment. Attached Figure Description

[0019] Figure 1 A hardware structure block diagram of a computer terminal for a copper-clad laminate copper-clad board copper-clad quality inspection method provided in an embodiment of the present invention;

[0020] Figure 2 This is a flowchart illustrating a method for detecting the copper clad laminate quality according to an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of a copper-clad laminate copper-clad board quality inspection system provided in an embodiment of the present invention. Detailed Implementation

[0022] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] This invention first provides a method for inspecting the copper clad laminate quality. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0024] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a copper-clad laminate copper-clad laminate copper-clad quality inspection method provided in an embodiment of the present invention. Figure 1As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0025] See Figure 2 The present invention provides a method for detecting the copper clad laminate quality, which may include the following steps:

[0026] S201, apply an ultrasonic excitation signal with a preset frequency and power to the copper-clad laminate under test, and simultaneously use a high-speed camera to acquire a sequence of dynamic micro-vibration response images of the surface of the copper-clad laminate under ultrasonic excitation, wherein the ultrasonic frequency is related to the natural frequency of the copper-clad laminate substrate and the copper layer to excite micro-vibration of a specific mode.

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

[0028] The system first receives the material thickness parameter (TP, in millimeters) and density parameter (DP, in grams per cubic centimeter) of the copper-clad laminate. These parameters can be obtained through user input or automatic scanning equipment (such as a laser thickness gauge). Taking a typical FR-4 epoxy resin substrate as an example, the substrate thickness value TP_substrate is 0.2 mm, the copper layer thickness value TP_copper is 0.035 mm; the substrate density DP_substrate is 1.8 g / cm³, and the copper density DP_copper is 8.9 g / cm³. Based on these parameters, a digital mechanical model of the copper-clad laminate is constructed using the Finite Element Modal Analysis (FEMA) algorithm. This algorithm discretizes the copper-clad laminate into millions of micro-mesh elements (e.g., tetrahedral elements), and each element is assigned mechanical properties such as the elastic modulus (EM) and Poisson's ratio (PR) according to the material properties. By solving large-scale eigenvalue equations, the free vibration modes of the material under unconstrained conditions are simulated. The calculation process is executed in parallel by a high-performance computing cluster, outputting a natural frequency spectrum (NFS) that includes substrate-dominated vibration modes (such as overall bending) and copper-dominated modes (such as local film vibration). The NFS is plotted with frequency values ​​(in kilohertz / kHz) on the x-axis and vibration energy ratio (VER) on the y-axis, including key peak points such as the substrate's first-order bending frequency of 85.3 kHz and the copper film vibration frequency of 217.6 kHz. The final generated NFS serves as a 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. Based on the physical boundary between the substrate and the copper layer, the algorithm establishes independent material layers (ML) in the mesh model and assigns corresponding TP and DP values ​​to each layer. The boundary condition is set to "Free-Free Boundary Condition" (FFBC), simulating the real state of the copper-clad laminate in air without clamp constraints. The solver uses the Lanczos Iteration Method (LIM) to extract the first N (e.g., the first 50) vibration modes. Each mode contains three key data points: natural frequency F_n (in kilohertz), mode shape vector V_m (describing spatial displacement distribution), and modal quality factor Q_m (reflecting energy dissipation characteristics). For composite structures like copper-clad laminates, the algorithm focuses particularly on two types of modes:

[0030] The dominant mode of the substrate is characterized by overall bending or torsion of the substrate, with a frequency range of 50-150 kHz (example value: first bending mode 85.3 kHz, mode shape is the maximum displacement of the plate center).

[0031] Copper layer dominant mode: manifested as local thin film vibration of copper layer, with a frequency range of 200-500 kHz (example value: copper layer thin film mode 217.6 kHz, the mode shape is high-frequency micro-amplitude vibration of copper foil region).

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

[0033] To improve computational efficiency, the algorithm employs Adaptive Mesh Refinement (AMR). The initial mesh size is set to 1 mm, and it is automatically refined to 0.1 mm in stress concentration areas such as copper layer edges and holes. The material constitutive model is a linear elastic model (LEM), with the substrate elastic modulus (EM_substrate) set to 22 GPa and the copper layer elastic modulus (EM_copper) set to 110 GPa. After calculation, the system verifies the reliability of the results through Modal Assurance Criterion (MAC): comparing the frequency deviations at adjacent mesh densities, if the deviation is less than 1% (e.g., 85.3 kHz vs. 85.2 kHz), the results are considered convergent. The final output PFCS strictly corresponds to the physical properties of the copper-clad laminate, providing precise frequency targets for subsequent ultrasonic excitation.

[0034] Using a preset frequency candidate set, the output frequency of the ultrasonic generator is dynamically tuned by an adaptive resonance tracking algorithm. The frequency point that can excite the maximum micro-vibration amplitude is selected, and the optimized ultrasonic excitation signal is output.

[0035] The Adaptive Resonance Tracking Algorithm (ARTA) uses the PFCS (Potential Frequency Control System) as input to control the ultrasonic generator for 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 watts). Simultaneously, a laser Doppler vibrometer measures the vibration velocity amplitude V_amp (in millimeters per second) at a designated point (usually the central region) on the copper-clad laminate surface in real time. ARTA scans within a frequency window of ±5% (e.g., 85.3 ± 4.3 kHz) with a 10 Hz frequency step size (FSS). When V_amp exceeds three times the background noise (e.g., increasing from 0.01 mm / s to 0.03 mm / s), the algorithm determines that it has entered the resonance region and immediately switches to the Golden Section Search (GSS) mode, reducing the step size to 1 Hz for fine scanning.

[0036] During the fine-scan 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 using cubic spline interpolation (CSI), and the resonant frequency F_res corresponding to the global maximum point is calculated (example value: 86.7 kHz). The system then automatically verifies whether F_res is within the theoretical tolerance range predicted by the PFCS (e.g., ±2%). If the verification is successful (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), a modal rematch procedure is triggered: F_res is compared with the mode shape of the nearest frequency point in the PFCS (e.g., 127.8 kHz), and the mode shape correlation coefficient (MCC) is used to determine if a match is found (MCC > 0.8 is considered a match). If a match is successful, the PFCS is updated to a new candidate set centered on F_res.

[0037] After determining F_res, the algorithm further optimizes the ultrasonic power parameters. At a fixed frequency F_res, the USG power increases from 10 watts to 100 watts in 5-watt increments, while simultaneously monitoring V_amp and the temperature rise ΔT (in degrees Celsius) fed back from the thermal imager. The algorithm stops when ΔT exceeds a safety threshold (e.g., 5°C) or V_amp growth saturates (increase less than 1%). Finally, a power (e.g., 60 watts) is selected that satisfies ΔT ≤ 5°C and V_amp ≥ 80% of its maximum value. This generates the optimized ultrasonic excitation signal (OES) with parameters {frequency: 86.7 kHz, power: 60 watts}.

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

[0039] The periodicity of OES is represented by a 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) requires the high-speed camera's frame rate F_frame (in frames per second) to satisfy an integer multiple matching principle: F_frame = N × F_res, where N is an integer (usually N ≥ 4). Taking a common camera as an example, if its maximum frame rate is 200,000 frames per second, then calculate the value of N that satisfies F_frame ≤ 200,000:

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

[0041] If a high-speed camera with a frame rate of 500,000 frames per second is actually selected, then the maximum N = 5 (F_frame = 433.5 kiloframes per second). At this point, it is necessary to verify the sampling theorem: since the highest frequency component of the vibration is about 3 times F_res (260.1 kHz), 433.5 kiloframes per second > 2 × 260.1 kHz, which satisfies the Nyquist sampling requirement.

[0042] The core of synchronization parameter configuration is phase locking. The system aligns the camera exposure signal with the rising edge of the OES (Output Sequence of Exposure) using a phase-locked loop (PLL) circuit. Setting parameters:

[0043] Exposure time T_exp: set to T_period / 4≈2.88 microseconds to ensure capture of key phase points (0°, 90°, 180°, 270°) within the vibration cycle;

[0044] Trigger delay T_delay: Adjusted in 0.1 microsecond steps to compensate for circuit propagation delay (typical value 0.5 microseconds);

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

[0046] The output 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 (PCD) system was employed. The master clock source used a 100 MHz temperature-compensated crystal oscillator (TCXO), with jitter less than 1 picosecond. The OES signal and camera trigger signal were transmitted via coaxial cables of equal length, with path difference controlled within 1 mm (delay difference of approximately 3.3 picoseconds). The final measured synchronization error was less than 0.1°, ensuring phase consistency of the image sequence.

[0048] An optimized ultrasonic excitation signal is applied to the copper-clad laminate, and a high-speed camera is triggered to acquire images based on a synchronous parameter set, generating a dynamic micro-vibration response image sequence.

[0049] The ultrasonic excitation subsystem executes OES parameters: a piezoelectric transducer (PZT) is attached to the edge of the copper-clad laminate, receiving an 86.7 kHz sinusoidal electrical signal, which is then amplified to output 60 watts of acoustic energy. An acoustic coupling agent (silicone grease, impedance matching value Z_match = 2.5 MNayl) is coated between the PZT and the board to reduce energy loss. As the ultrasonic waves propagate within the board, they excite specific modes (such as copper film vibration), generating micro-vibration displacement D_vib on the surface (typical amplitude 0.1-5 micrometers). Simultaneously, the high-speed camera subsystem operates based on SPS parameters: 0.5 microseconds after the rising edge of the OES signal (T_delay), the camera starts capturing images at a frame rate of 433.5 kHz, with each frame exposed for 2.88 microseconds, for a total of 40 frames acquired.

[0050] The imaging system employs 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 × 1024 pixels, with a pixel size of 5.5 micrometers. Combined with a 100 mm macro lens, the imaging field of view covers the central area of ​​the copper-clad laminate (30 × 24 mm), achieving a spatial resolution of 23 micrometers per pixel, capable of resolving microcracks in the copper layer (>50 micrometers). Each frame is stored in 12-bit grayscale format, with a dynamic range that accommodates grayscale changes corresponding to a 0.1 micrometer displacement (approximately 20 grayscale levels).

[0051] Real-time quality monitoring was implemented during the acquisition process: the vibration amplitude was verified to be stable at 0.25±0.02 mm / s through LDV; the temperature rise ΔT was monitored by the thermal imager to be ≤3℃; the camera's built-in image signal-to-noise ratio (SNR) calculation module ensured that SNR>30 dB.

[0052] If any indicator exceeds the limit (e.g., SNR < 25 dB), the system will automatically interrupt and restart the acquisition. The final output is a Dynamic Vibration Image Sequence (DVIS), containing 40 frames of 1280×1024 pixel images, with a total data size of approximately 2.5 gigabytes (GB).

[0053] S202, perform phase-locked analysis on the dynamic micro-vibration response image sequence, extract the amplitude-phase change features of each pixel in the ultrasonic excitation cycle, and construct a dynamic response feature map of the copper-clad laminate. The phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by the excitation from the environmental noise.

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

[0055] The cornerstone of the entire phase-locked analysis is obtaining 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 (such as a preset optimized frequency point, assumed to be 500 kHz) and power (such as 50 watts W). This electrical signal drives the ultrasonic transducer to produce mechanical vibrations, while its original, unamplified drive signal waveform (DSW) or a strictly synchronized reference clock signal (RCS) is output in real time to a high-precision data acquisition card (DAC). The data acquisition card's sampling rate (SR) is set very high, for example, 100 MHz (i.e., 100 million points per second), to ensure accurate capture of the details of the excitation signal. The system software (or embedded firmware) precisely detects the start point (i.e., zero-crossing point, ZCP) or peak point (PP) of each complete sine wave from this high-speed sampled excitation signal waveform. For example, the system may be configured to record a timestamp (TS) at each positive slope zero-crossing of the excitation signal. This timestamp is a high-precision absolute time value, typically provided by the system's high-resolution timer (HRT), with an accuracy down to the nanosecond level (e.g., 10 nanoseconds). These sequentially recorded timestamps (TS1, TS2, TS3, ..., TSN) constitute the Phase Reference Vector (PRV). Each element in the PRV precisely corresponds to the start time of a complete cycle of the ultrasonic excitation signal, providing a precise phase alignment anchor for all subsequent analysis steps. The length N of this vector depends on the total duration of the acquired dynamic image sequence and the frequency of the ultrasound. For example, if the excitation frequency is 500 kHz (period of 2 microseconds μs) and the acquisition lasts for 0.1 seconds, then the PRV will contain approximately 50,000 timestamps.

[0056] The accuracy of constructing the Phase Reference Vector (PRV) is crucial. Besides relying on high sampling rates (SR) and high-resolution timers (HRT), the system employs digital signal processing techniques to further improve the accuracy of timestamp detection. For example, when detecting zero-crossing points (ZCPs), the system does not simply use the sign change between two adjacent sampling points for a rough judgment, but instead combines an interpolation algorithm (IA). A common method is three-point linear interpolation or sine fitting interpolation. Suppose that in the sampling sequence of the data acquisition card (DAC), the Kth sampling point has a negative value and the (K+1)th sampling point has a positive value; then the true zero-crossing point must lie 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 can be calculated more accurately using a linear or sine function model. This process significantly reduces the error introduced by sampling discreteness (quantization error), making the timestamp (TS) accuracy in the PRV much higher than the sampling interval (for a 100MHz SR, with an interval of 10ns, the accuracy after interpolation can reach 1ns or even higher). In addition, the system monitors the frequency stability of the ultrasonic generator (UG) output. Although the optimized frequency is preset, there may be slight drift during actual operation. The PRV construction process itself also implicitly involves measuring the instantaneous frequency of the actual excitation signal, and this information is recorded for handling possible non-steady-state conditions in subsequent analysis.

[0057] The Phase Reference Vector (PRV) is not merely a list of time points; it is also closely related to the triggering and synchronization logic of the entire experiment. Claim 2 states that the frame rate (FR) of the high-speed camera is matched to an integer multiple of the frequency of the ultrasonic excitation signal. This means that before the system begins acquiring the Dynamic Micro-Vibration Response Image Sequence (DMVRIS), the frame rate of the high-speed camera has been precisely calculated and set (FR = 500kHz * 8 = 4 million frames per second (MFPS) based on the optimized ultrasonic frequency (e.g., 500kHz) and the desired number of sampling points per cycle (e.g., 8 frames per ultrasonic cycle). The Exposure Trigger Signal (ETS) that triggers the high-speed camera to begin exposure is also precisely emitted by the same main control system based on the start timestamp (TS1) in the PRV, ensuring that the exposure start time of the first frame is strictly aligned with a certain phase of the excitation signal (e.g., the initial zero-crossing). After the PRV is constructed, it serves not only as the basis for the temporal alignment of subsequent image sequences but also as a verification benchmark for the time synchronization of the entire system. The system checks whether the exposure time (ET) and readout time (RT) of each camera frame maintain the expected timing relationship with the excitation period and PRV, ensuring that there are no dropped frames or timing drift. This strictly synchronized PRV is a key guarantee for the subsequent extraction of weak micro-vibration signals from a noisy background.

[0058] Based on the phase reference vector, a time-domain alignment operation is performed on the dynamic micro-vibration response image sequence, and the aligned image sequence is output.

[0059] Although the frame rate (FR) of the Dynamic Micro-Vibration Response Image Sequence (DMVRIS) acquired by the high-speed camera matches an integer multiple of the excitation frequency, the ultrasonic excitation phases corresponding to the image frames arranged directly in the acquisition order are not strictly aligned due to the slight delays and uncertainties in camera exposure and image transmission, as well as the extremely slight frequency jitter that may exist in the ultrasonic excitation signal itself. In other words, the surface state captured by the M-th frame does not precisely correspond to the K-th fixed phase point of the N-th cycle of the excitation signal. The purpose of Temporal Alignment Operation (TAO) is to use the previously constructed Phase Reference Vector (PRV) to rearrange (or interpolate) each frame in the image sequence so that it precisely corresponds to the surface state at a specific phase moment of the ultrasonic excitation signal (e.g., the start point, 1 / 4 cycle point, 1 / 2 cycle point, etc.). This is the core preprocessing step for achieving high-sensitivity vibration measurement.

[0060] The specific process of performing Temporal Alignment (TAO) is as follows: The system first needs to determine the Target Phase Point (TPP). Typically, multiple equally spaced phase points within the excitation signal cycle are selected as the targets for analysis. For example, if it is desired to analyze 8 phase states (PS) within each ultrasonic cycle, the target phase point TPP can be defined as: PS0 (0° / start point), PS1 (45°), PS2 (90°), PS3 (135°), PS4 (180°), PS5 (225°), PS6 (270°), and PS7 (315°). For each raw frame (RF) in DMVRIS, the system finds the two nearest timestamps (e.g., TS_i and TS_i+1) in the Phase Reference Vector (PRV) based on its Exposure Start Time (EST) or Exposure Midpoint Time (EMT). These two timestamps define a complete ultrasonic cycle. Then, the time offset (TO) of the frame's EST or EMT relative to the start point TS_i of this period is calculated. Dividing this time offset TO by the actual duration of the period (i.e., TS_i+1-TS_i) yields the actual phase angle (APA) of the frame, typically ranging from 0 to 360 degrees. Next, based on the preset target phase points (TPPs, such as PS0, PS1, ..., PS7), the system needs to find (or construct) a corresponding, precisely phase-aligned image for each TPP. If, for a certain TPP (e.g., PS0), the APA of an original frame (RF) happens to be very close to 0° (within a very small tolerance range, such as ±0.5°), then this RF frame is directly marked as belonging to PS0. However, in most cases, the APA of the original frame may not fall precisely on the preset TPP. In this case, image interpolation (II) techniques are needed. Common methods include linear interpolation based on adjacent frames or more advanced motion-compensated interpolation. For example, to obtain an image at PS2 (90°), we find that the APA of RF_j in the previous frame is 85°, and the APA of RF_k in the next frame is 95°. Therefore, the target image can be viewed as a weighted average of RF_j and RF_k. The weights are calculated based on the differences between the target phase (90°) and the phases of RF_j (85°) and RF_k (95°). The smaller the difference, the greater the weight of that frame. Through this interpolation, a surface image that precisely corresponds to the 90° phase point in time can be "synthesized".This process is performed on each frame in the sequence, constructing a sub-sequence for each target phase point (TPP), with each sub-sequence containing images of the same phase point at different excitation periods.

[0061] The output is an aligned image sequence (AIS). After the above processing, the original DMVRIS is reorganized into a new sequence, AIS. This new sequence typically has a three-dimensional structure: 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 location (X, Y), under each phase state PS, there is now a series of images captured (or interpolated) at the same phase time but with different excitation cycles. This alignment process effectively decouples the periodic micro-vibration motion caused by ultrasonic excitation, synchronized with the excitation signal, from random environmental noise (such as air disturbance, equipment vibration, and electronic noise). Because environmental noise is usually aperiodic or its frequency components differ from the excitation frequency, after being aligned in the time domain and arranged according to the excitation phase, its intensity is randomly distributed between different phase points and does not exhibit a regular change related to the phase of the excitation signal. The actual micro-vibration response, however, exhibits a regular, periodic pattern of intensity or positional change with the phase of the excitation signal (0° to 360°). This structured AIS provides a perfect foundation for the next step of analyzing the vibrational characteristics (amplitude and phase) of each pixel as a function of the excitation phase. The system stores the AIS and marks the precise phase state PS corresponding to each image.

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

[0063] Now, we have the aligned image sequence (AIS), meaning that for each pixel (X, Y) on the copper-clad laminate surface, we have a series of intensity values ​​(grayscale values) at eight (or other numbers) equally spaced phase points (PS0-PS7) of the ultrasonic excitation signal. This time series of intensity values ​​(along the same phase point across different excitation cycles) should theoretically be constant unless defects or stress anomalies cause micro-vibrations at that point. These micro-vibrations result in a slight spatial shift (amplitude) of the point, and this movement may have a delay (phase angle) relative to the excitation signal. To detect these minute vibrations at the same frequency as the excitation from the intensity sequence, a powerful tool capable of analyzing the local frequency characteristics of the signal is needed. The Wavelet Packet Transform Algorithm (WPTA) is used for this purpose. Wavelet packet transform (WPT) is an extension of wavelet transform (WT). It provides a finer frequency resolution than standard wavelet decomposition and can decompose signals into a series of preset, narrower frequency bands (FB).

[0064] For each pixel (X, Y) and each phase state (PS), the system processes the intensity value sequence (IVS) of gray values ​​across all image frames corresponding to that PS in AIS. The length of this IVS is equal to the number of excitation cycles acquired (e.g., 1000 cycles). The system applies WPTA to this IVS. First, a suitable wavelet basis function (WBF) needs to be selected. Commonly used and suitable for vibration analysis are the Daubechies (Db) series (e.g., Db4, Db8) or Symlets (Sym) series (e.g., Sym8) wavelets, which have good time-frequency localization characteristics. Then, the number of decomposition levels (DL) needs to be determined. The selection of DL needs to balance frequency resolution and time resolution. For ultrasonic frequencies in the hundreds of kilohertz (kHz) range, with a sampling frequency of frame rate (e.g., 4MFPS, i.e., 4e6 Hz), a suitable DL might be 6 to 8 levels. For example, by selecting DL=7, WPT will decompose the original signal (bandwidth 0-2MHz) into increasingly narrower sub-bands layer by layer. At each layer, each node represents a specific frequency band range. Ultimately, at layer 7, 128 (2^7) sub-band nodes will be obtained. The system will identify which sub-band node's center frequency (CF) is closest to the ultrasonic excitation frequency (e.g., 500kHz). This band is considered the Micro-Vibration Dominant Frequency Band (MVDFB). WPT not only decomposes the signal into different frequency bands, but more importantly, it provides a complex value coefficient for each band at each time point (actually, each excitation cycle point). This complex coefficient contains information about the instantaneous amplitude (IA) and instantaneous phase angle (IPA) of the signal in that band at that time point. For an IVS at pixel (X,Y) in phase state PS, the most important result in the WPTA output is the complex coefficient sequence corresponding to MDFFB. The change of the magnitude of this sequence with time (excitation period) reflects the trend of the micro-vibration amplitude IA at that point in phase PS (e.g., whether there are defects 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 stress concentration causing phase lag).

[0065] To obtain a stable estimate characterizing the vibration properties of a pixel (X,Y), the system does not rely solely on the IA and IPA of a single excitation cycle, but instead performs statistical analysis on the MDFFB complex coefficient sequence. Typically, the mean instantaneous amplitude (MIA) and mean instantaneous phase angle (MIPA) of the sequence are calculated over the entire acquisition period. MIA reflects the average intensity of the micro-vibration amplitude at that point under a specific phase state PS; MIPA reflects the average offset of the vibration relative to the phase of the excitation signal. This process needs to be performed independently for each pixel (X,Y) and each phase state PS (PS0-PS7). Ultimately, for each spatial location (X,Y), the system obtains 8 MIA values ​​(corresponding to PS0-PS7) and 8 MIPA values ​​(corresponding to PS0-PS7). Organizing these values ​​for all pixels forms the pixel-level amplitude-phase matrix (PLAPM). This is a four-dimensional data structure: spatial dimensions X and Y; feature dimensions including MIA and MIPA; and phase dimension PS (0-7). PLAPM is the core data for subsequent analysis, capturing the complete dynamic characteristics of the micro-vibration response of each point on the copper-clad laminate surface under ultrasonic excitation, as the excitation phase changes. Amplitude information (MIA) can reveal material stiffness, bonding strength (small amplitude may indicate good bonding and high stiffness; large amplitude may indicate weak bonding or cracks), damping characteristics, etc.; phase information (MIPA) can reveal local stress state (stress concentration may cause phase lag), material inhomogeneity, etc. The advantage of WPTA lies in its excellent time-frequency localization capability, which can effectively separate weak vibration signals near the excitation frequency, even if they are submerged in broadband noise.

[0066] Using a phase coherence filtering algorithm with the phase reference vector as a reference, the environmental noise component in the amplitude-phase matrix is ​​suppressed, the pure micro-vibration signal is separated, and the noise-suppressed amplitude-phase feature set is output.

[0067] Although temporal alignment and wavelet packet transform have greatly enhanced micro-vibration signals, some noise inevitably remains in the pixel-level amplitude-phase matrix (PLAPM). This noise mainly originates from: 1) camera readout noise (RN) and shot noise (SN); 2) ambient light fluctuations; 3) weak mechanical vibrations not fully correlated with ultrasonic excitation; and 4) electronic system noise. The core idea of ​​the Phase Coherence Filtering Algorithm (PCFA) is to utilize a key characteristic: the phase (represented as MIPA in PLAPM) of a genuine, 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). In contrast, signal fluctuations caused by noise have random and incoherent phases.

[0068] The specific process of performing Phase Coherence Filtering (PCFA) is as follows: For each pixel (X,Y) and each phase state PS, we have already obtained an average instantaneous phase angle MIPA from PLAPM. This MIPA is obtained by averaging the complex coefficient sequence of WPT previously. However, to evaluate phase stability (i.e., coherence), we need to examine the instantaneous phase angle IPA for each excitation cycle in more detail. Recall that when applying WPTA, for each pixel (X,Y), each PS, and each excitation cycle T, WPT outputs the instantaneous phase angle IPA(X,Y,PS,T) of MVDFB. The system calculates the standard deviation (SD_IPA) of the IPA(X,Y,PS,T) values ​​over all excitation cycles T. This standard deviation SD_IPA quantitatively measures the degree to which the instantaneous phase at that point fluctuates around its average MIPA at 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 more unstable the phase and the more likely it is noise-dominated. Simultaneously, the system calculates the ratio or difference between the average instantaneous amplitude MIA(X,Y,PS) of that point under that PS and the average amplitude of all pixels under that PS (or the amplitude of a certain reference area), serving as an auxiliary amplitude confidence indicator (ACI). True micro-vibration regions should exhibit spatially smooth amplitude and phase changes, while 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; a higher value indicates a higher probability that the signal at that point under that PS is a true coherent micro-vibration and less susceptible to noise contamination. The calculation formula for PCC may be a multidimensional function, such as PCC = exp(-k1*SD_IPA)*sigmoid(k2*ACI), where k1 and k2 are empirical constants, and sigmoid is a sigmoid function used for normalization. After the PCC is calculated, it is applied to 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] A more common approach is to primarily use PCC to weight and suppress noise in MIA, while keeping MIPA unchanged or smoothing only the MIPA of low PCC points. This process significantly suppresses the amplitude contribution of pixels with unstable phase (high SD_IPA) or abnormally low / high amplitude (low ACI) (these pixels are likely noise), while preserving information from pixels with high phase coherence (low SD_IPA) and reasonable amplitude (these pixels are genuine vibration signals). Essentially, PCC acts as an adaptive weight in both spatial and feature dimensions.

[0070] The output is a 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 reduced to zero in noisy regions, while it is largely preserved or enhanced in real micro-vibration regions (because noise is suppressed, the signal-to-noise ratio is relatively improved). The filtered MIPA value (Filtered_MIPA) usually does not change much, but phase values ​​in low PCC regions may be marked as unreliable or smoothed out by neighborhood values. NSAPFS retains the same structure as PLAPM (X, Y, PS, Filtered_MIA, Filtered_MIPA), but its data quality is significantly improved, and environmental noise components are effectively suppressed. This feature set is the direct input for constructing the final dynamic response feature map. The success of PCFA relies on the accurate phase reference (PRV) and temporal alignment (AIS) established in the previous steps, ensuring that only signals strictly synchronized with the excitation exhibit high coherence. It cleverly utilizes the periodicity of signals as a "fingerprint" to distinguish between signals and noise, which is a key step in improving the sensitivity and reliability of defect detection.

[0071] The amplitude-phase feature set after noise suppression is mapped to two-dimensional spatial coordinates to construct the dynamic response feature map of copper-clad laminate.

[0072] The noise-suppressed amplitude-phase feature set (NSAPFS) contains key vibration features for each pixel (X,Y) in eight (or more) phase states (PS): the filtered average instantaneous amplitude (Filtered_MIA) and the filtered average instantaneous phase angle (Filtered_MIPA). To visually represent the vibration response characteristics of the entire copper-clad laminate surface and use them as input for subsequent deep learning models, these features need to be mapped back to their corresponding two-dimensional spatial coordinates, forming an easily processed image format: the Copper Clad Laminate Dynamic Response Feature Map (CCLDRFM).

[0073] The process of constructing a feature map involves feature selection, combination, and visual encoding. The system typically doesn't directly cram all 16 features (8 PS * [Filtered_MIA + Filtered_MIPA]) into a single massive feature map; instead, it performs meaningful combinations or selects the most representative subset of features. Common strategies include:

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

[0075] Maximum Amplitude Map (MAM): For each pixel (X,Y), calculate the maximum value of the Filtered_MIA across all 8 PS. MAM(X,Y) = max(Filtered_MIA(X,Y,PS0),...,Filtered_MIA(X,Y,PS7)). This map visually displays the peak distribution of micro-vibration amplitudes across the copper-clad laminate surface. Poorly bonded areas and microcracked areas typically exhibit abnormally high amplitudes (intense vibrations). Areas with abnormal local stress may result in abnormally low amplitudes (material hardening) or high amplitudes (local softening).

[0076] Mean Phase Lag Map (MPLM): For each pixel (X,Y), calculate the average of the Filtered_MIPA values ​​over 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 distribution of the average phase delay of the vibration response relative to the excitation signal. Stress concentration areas and the vicinity of internal material defects can cause changes in wave propagation velocity, thus resulting in phase lag.

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

[0078] Feature fusion map: This merges amplitude and phase information onto a pseudo-color map. For example, hue is used to represent MPLM (phase lag), and saturation or value / intensity is used to represent MAM (amplitude intensity). Thus, different colors represent different phase lag states, and the intensity or shade of the color represents the magnitude of the amplitude.

[0079] Multi-channel feature map: To retain richer 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 (which may need to be normalized to 0-255). In this way, each spatial location (X,Y) has a feature vector containing 16 values, and the entire copper-clad laminate surface constitutes a 16-channel "image".

[0080] Regardless of the strategy employed, the final output of the Copper Clad Laminate Dynamic Response Feature Map (CCLDRFM) is an array of rich vibrational characteristic data for each point on a two-dimensional spatial structure (with width W and height H, corresponding to the camera's field of view). This feature map is the culmination of all the preceding complex signal processing steps (phase reference, temporal alignment, wavelet packet analysis, and phase coherence filtering). It effectively presents the complex dynamic micro-vibration response (including amplitude and phase information) of the Copper Clad Laminate surface under ultrasonic excitation in the form of a spatial distribution map. In high-quality, defect-free areas, the amplitude map should exhibit uniform and low values, and the phase map should also be relatively uniform. However, in areas with poor copper bonding, due to debonding of the copper layer from the substrate or the presence of voids, the local stiffness of this area decreases, and the vibration amplitude under ultrasonic action will significantly increase, appearing as bright spots (high values) on the MAM map; simultaneously, the reflection and scattering of sound waves at the debonding interface may lead to abnormal phase distribution. In microcrack areas, the crack tip is a place of high stress concentration and a strong source of wave scattering, which will also lead to abnormally high amplitude and phase distortion near the crack. Localized stress anomalies (such as excessive residual stress) can alter the elastic modulus and sound velocity of the local material, affecting wave propagation and potentially causing a decrease in amplitude (hardening) or a phase shift (hysteresis or lead). CCLDRFM captures these subtle dynamic response differences closely related to the material's internal state and defects, providing high-quality, highly informative input data for the next step of automatically identifying and locating defects using deep learning models. Once constructed, CCLDRFM is passed to the defect identification model for further analysis.

[0081] S203, input the dynamic response feature map of the copper-clad laminate into the trained defect identification model, identify and locate potential poor copper-clad laminate bonding areas, microcrack areas and local stress anomaly 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, multi-channel feature enhancement can be performed on the dynamic response feature map of copper-clad laminates, and abnormal response areas can be highlighted through an attention mechanism to output an enhanced feature map;

[0083] When the dynamic response feature map of the copper clad laminate (containing amplitude and phase information for each pixel) is input, the system first performs multi-channel feature enhancement. This feature map is essentially dual-channel data: the first channel stores the amplitude value (AV), reflecting the local vibration intensity of the material; the second channel stores the phase angle (PA), reflecting the temporal relationship of the vibration. The enhancement process first uses Contrast Limited Adaptive Histogram Equalization (CLAHE) technology to process the two channels separately. Taking the amplitude channel as an example: the system divides the image into several local blocks (e.g., 32×32 pixels), calculates the grayscale histogram independently for each block, and limits over-enhancement by using a preset contrast limiting threshold (Clip Limit, CL, typically 2.0). Then, pixel values ​​are redistributed to stretch the local contrast. This operation significantly improves the distinction between weak vibration areas (e.g., the edge of a potential microcrack) and strong vibration areas (e.g., an intact copper layer). Meanwhile, the phase channel is enhanced using Histogram of Oriented Gradients (HOG) to calculate the directional gradient of the phase angle within the neighborhood of each pixel (e.g., a 5×5 window), thereby enhancing the texture features of the vibration wave propagation direction. This is crucial for identifying the anisotropic response in areas of stress anomalies.

[0084] The enhanced dual-channel data input channel attention module (CAM) first performs global average pooling (GAP) on each channel, compressing the two-dimensional feature map of each channel into a single scalar (i.e., a channel importance score). The pooling value of the amplitude channel (denoted as S_AV) reflects the overall vibrational energy level, and the pooling value of the phase channel (S_PA) reflects the overall temporal consistency. These two scalars are input into a small fully connected network (containing one hidden layer and 16 nodes), which outputs the weight coefficients of the two channels (Weight Coefficient, WC_AV and WC_PA, ranging from 0 to 1). If the current sample's S_AV is significantly higher than the training set mean, then WC_AV approaches 1 (emphasizing amplitude anomalies); if S_PA deviates from the typical value, then WC_PA increases (highlighting phase disorder regions). Finally, the original amplitude map is multiplied by WC_AV, and the phase map is multiplied by WC_PA to achieve channel-level weighting.

[0085] The channel-weighted feature map is further input into the Spatial Attention Module (SAM). This mechanism first stitches the dual-channel data along the depth direction to form a two-dimensional map. By calculating the standard deviation (SD) of each pixel location within its neighborhood (e.g., a 7×7 window), a Spatial Saliency Map (SSM) is generated. Regions with high standard deviations (such as poorly bound areas with abrupt amplitude changes) are highlighted in the SSM. After the SSM is normalized to the 0-1 range using the sigmoid function, it is multiplied pixel-by-pixel with the original stitched map. This allows the model to focus on the spatial locations of abnormal fluctuations in vibration response (such as sudden amplitude drops at microcracks or phase lag in stress concentration areas). The final output is an Enhanced Feature Map (EFM) that fuses the channel and spatial attention data, which improves the signal-to-noise ratio of abnormal regions by approximately 60%, providing a highly discriminative input for subsequent defect identification.

[0086] The enhanced feature map is input into a pre-trained defect recognition model, which adopts a graph convolutional network architecture, integrates dynamic response features and prior knowledge of material stress distribution, and outputs the original defect probability map.

[0087] A pre-trained defect recognition model is constructed using Enhanced Feature Map (EFM) input as its core. This model employs a Graph Convolutional Network (GCN) architecture. Unlike traditional Convolutional Neural Networks (CNNs), GCN treats the copper-clad laminate (CCL) as a graph structure in non-Euclidean space. First, graph construction is performed: each superpixel block (SPB, generated by the SLIC algorithm, approximately 10×10 pixels) of the EFM serves as a graph node (ND). The node feature vector contains five statistical parameters, including the mean amplitude and phase variance within the block. Edges (ED) between nodes are established according to two rules: 1) Spatially adjacent superpixels are forcibly connected; 2) Based on prior knowledge of material stress propagation, if two nodes are located on the same copper foil trace (according to the design topology) and the included angle is less than 30 degrees, a weighted edge is added (the edge weight is positively correlated with the cosine of the included angle).

[0088] Graph-structured data is input into a multi-layered GCN for message passing (MP). Each GCN layer performs three key operations: 1) Neighbor aggregation: For each node, the feature vectors of its first-order neighbors are collected and aggregated into an aggregated vector (AV) using mean pooling; 2) Feature transformation: The node's own features are concatenated with the AV and input to a weight matrix (WM, whose size is the product of the input and output dimensions) for linear transformation; 3) Nonlinear activation: Nonlinearity is introduced using the modified linear unit (ReLU) function. Through two layers of GCN iteration, node features can be fused with three-hop neighborhood information. For example, a node located at a copper foil corner will have features including vibration propagation characteristics from the straight line segment (reflecting stress concentration effects) and damping characteristics of the adjacent resin region.

[0089] The final GCN layer outputs node-level feature vectors, which are input to the Classification Head (CH). CH contains a fully connected layer (output dimension three) and a Softmax function, generating a probability distribution (PD) for each node belonging to one of the three defect types (poor bonding / microcrack / stress anomaly). Subsequently, an image-to-image mapping is performed: based on the correspondence between superpixel blocks and the original image, node probability values ​​are assigned to corresponding pixels, generating a full-resolution Raw Defect Probability Map (RDPM). Each pixel in this map contains a three-dimensional probability value (P_bond, P_crack, P_stress), summing to 1. For example, the P_crack probability in a microcrack region can reach above 0.95, while the probabilities for all three types in a healthy region are below 0.1.

[0090] An adaptive threshold segmentation algorithm is applied to process the original defect probability map, dynamically dividing the defect category boundaries according to the probability values, and outputting a binary defect mask map.

[0091] The three probability channels of the original defect probability map (RDPM) are processed separately. Taking the poor bonding probability channel (P_bond channel) as an example: First, Gaussian Adaptive Thresholding (GAT) is used. The algorithm slides a window (e.g., 50×50 pixels) on the image and calculates the local mean (LM) and local standard deviation (LSD) of P_bond within the window. The dynamic threshold (DT) is determined by the formula DT = LM + k × LSD, where k is the sensitivity coefficient (preset value 1.5). If a pixel's P_bond > DT, it is temporarily marked as a candidate poor bonding point. This method overcomes the sensitivity problem of the global threshold to uneven illumination (probability value)—it automatically increases the threshold in high-probability background areas (such as areas with existing large bonding defects) to prevent false expansion, and decreases the threshold in low-probability areas to increase sensitivity.

[0092] After initial segmentation, Probability-Guided Region Growing (PGRG) is performed to optimize the boundary. Starting from each candidate point, its eight neighboring pixels are examined. If the difference between the P_bond value of a neighboring pixel and that point is less than the tolerance (Tol, set to 0.15) and they belong to the same copper foil region (according to the design topology), then that pixel is included in the current defect region. During the growth process, the growth threshold is dynamically updated: the mean P_bond value of newly added pixels and the standard deviation of the region are recalculated, and subsequent neighbor comparisons are based on the updated statistics. This process iterates until no new pixels are added, ensuring that the segmentation boundary is aligned with the probabilistic gradient transition position, avoiding jagged edges caused by traditional fixed thresholds.

[0093] After segmenting the three types of defect channels as described above, conflict resolution and mask synthesis are performed. If a pixel is labeled by multiple channels (e.g., P_bond>0.8 and P_crack>0.7), its final category is determined using the Maximum Probability Ownership (MPO) principle. Subsequently, three binary defect mask maps (BDMMs) are generated: a Bonding Mask (BM), a Crack Mask (CM), and a Stress Mask (SM). In each mask, the pixel value of the defect region is one (1), and the background is zero (0). For example, a microcrack with a width of two micrometers (2μm) appears as a continuous bright line in the CM.

[0094] Mark connected regions on the binary defect mask and classify them as poor copper bonding, microcracks or stress anomalies, and output a preliminary defect region distribution map.

[0095] Connected Component Labeling (CCL) is performed on each binary mask (BM / CM / SM). A two-pass scanning algorithm is used: the first pass traverses the image from left to right and top to bottom, assigning a temporary label to each foreground pixel and recording equivalence relationships (e.g., adjacent pixels belonging to the same defect); the second pass parses the equivalence relationship table, merges the temporary labels, and assigns a unique label (Label ID, LID) to each independent connected component. Simultaneously, the geometric properties of each connected component are recorded: bounding box (BB), area (A, number of pixels), perimeter (P), centroid coordinates (CC), etc. For example, a circular layered region is labeled LID_001 in BM, and its area is recorded as 150 pixels.

[0096] Secondary verification of defect types is performed based on geometric and topological features. The system predefines a morphological rule base for three types of defects:

[0097] Poorly bonded areas: area threshold (A>200 pixels) and form factor (FF=4πA / P) 2 Approximately 1 (circular feature), mostly located at the edge of the copper layer.

[0098] Microcrack region: Aspect ratio (AR = BB_width / BB_height) is greater than 5 (5) and FF is less than 0.3 (0.3), and the direction is usually parallel to the stress direction.

[0099] Anomaly zone: The spatial distribution is less than 100 micrometers (100μm) away from the high stress points (such as pad corners) in the design topology, and it is star-shaped and radial (FF≈0.5).

[0100] If the features of a connected region conflict with the mask type (such as a circular region in CM), the manual rule engine is activated to adjust the type or mark it as pending review.

[0101] The labeling results of the three types of masks are integrated to generate a preliminary defect distribution map (PDDM). This map is an RGB color image: poorly bonded areas are marked in red (R=255, G=0, B=0), microcracks are marked in blue (R=0, G=0, B=255), and stress anomalies are marked in yellow (R=255, G=255, B=0). Each connected region is labeled with its LID and type code (e.g., "C-102" represents microcrack number 102), and a crosshair is superimposed on the centroid. Simultaneously, a supplementary data table is generated, recording the LID, type, area, location coordinates, minimum bounding rectangle, and other attributes of each defect for subsequent morphological analysis and physical verification. At this point, the preliminary intelligent interpretation of non-contact ultrasonic vibration detection is completed.

[0102] The trained graph convolutional network model identifies three typical defects by analyzing features such as vibration amplitude distribution and phase delay modes. Integrating prior knowledge of materials mechanics, the model can distinguish between real defects and subtle response differences caused by normal process fluctuations, achieving automated defect classification and localization, and overcoming the subjectivity issues of manual interpretation. The deep learning model's ability to model complex features nonlinearly significantly improves the detection rate of latent defects (such as incompletely peeled bonding defects).

[0103] S204, Perform multi-scale morphological analysis and regional connectivity verification on the preliminary defect area distribution map, and combine the design circuit topology information of the copper-clad laminate to screen out the real defect areas that conform to physical laws and generate a precise location defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad laminate areas.

[0104] Specifically, multi-scale morphological opening and closing operations can be performed on the preliminary defect area distribution map, and isolated noise points can be eliminated using variable-scale structuring elements to output a morphological correction map.

[0105] The system receives a preliminary defect region map (PDRM) generated by the defect identification model. This image is a binary image, where white pixels represent areas marked as potential defects (including poor bonding, microcracks, or stress anomalies), and black pixels represent normal background. Due to potential sensitivity fluctuations or image noise interference in the preceding model, the PDRM often contains two types of interference: isolated noise points with tiny sizes (e.g., 1-5 pixels), caused by dust reflections or camera noise; and irregular boundaries, generated by discontinuous threshold segmentation. To eliminate these interferences, multi-scale morphological opening-closing operations are employed. The core operation unit is the variable-scale structuring element (VSSE). The structuring element is essentially a pixel template with a preset geometric shape (e.g., circle, square), and its size is controlled by the scale parameter (SP). The system pre-defines three typical scales: small scale SP_S (3 pixels in diameter), medium scale SP_M (7 pixels in diameter), and large scale SP_L (15 pixels in diameter). First, an opening operation is performed: the PDRM is processed sequentially using circular structuring elements in the order SP_S→SP_M→SP_L, performing erosion followed by dilation. The small-scale opening operation filters out isolated noise points (such as isolated white pixels within a 3×3 area), the medium-scale opening operation smooths small bumps, and the large-scale opening operation preserves the main structure of the true defects. Next, a closing operation is performed: dilation followed by erosion is performed in the same order to fill small holes (such as broken pixels inside microcracks) and connect adjacent fragmented defect areas (such as discrete spots in stress anomaly zones). The final output is a morphologically corrected map (MCM) with significantly suppressed noise and smooth, coherent boundaries.

[0106] The application of variable-scale structuring elements requires dynamic adaptation to defect features. The system has a built-in Scale Selection Strategy (SSS): it calculates the area distribution histogram (ADH) of all connected regions in the PDRM. If the ADH shows that 80% of the regions have an area less than 10 pixels², SP_S is prioritized; if there is a macro-defect with an area greater than 100 pixels², SP_L operation is performed to protect structural integrity. During the operation, a progressive erosion-dilation (PED) mechanism is used: for example, when performing opening operations, the image is first eroded with SP_S to remove tiny noise points; then SP_M is used to dilate and restore part of the structure; finally, SP_L is used to slightly erode and refine the edges. This mechanism avoids the loss of detail caused by a single scale (such as the tip of a micro-crack being smoothed out by a large-scale operation). All morphological operations are performed in high-speed parallel computation using image processing libraries (such as the low-level optimization of the morphologyEx function in OpenCV), processing a 300mm×300mm copper-clad laminate area in just 50 milliseconds (ms). MCM is temporarily stored as an intermediate result in the cache, and its pixel accuracy remains consistent with the original image (typically 5 micrometers / pixel), ensuring the spatial resolution of subsequent analysis.

[0107] To verify the effectiveness of morphological processing, the system performs a Noise Suppression Ratio Evaluation (NSRE). It compares the changes in the number of connected regions in the PDRM and MCM: if a region exists in the PDRM but disappears in the MCM, and its area is lower than the preset Valid Defect Area Threshold (VDAT, default 8 pixels²), it is considered a successfully eliminated noise point. Simultaneously, it monitors Defect Contour Fidelity (DCF): calculating the Area Loss Rate (ALR) of the actual defect in the MCM. If the ALR exceeds 5% (e.g., due to overfilling of microcrack widths caused by closing operations), the scale of the closing operation is automatically reduced or the processing of that region is skipped. The final output MCM must meet two criteria: an isolated noise point reduction rate ≥95%, and an actual defect deformation rate ≤3%. This step significantly improves image quality and lays the foundation for subsequent connectivity analysis.

[0108] Based on the morphological correction map, a region growing algorithm is used to label connected defect regions and output a connected region label map.

[0109] The Morphological Correction Map (MCM) is input to the Region Growing Module (RGM). This module uses the Region Growing Algorithm (RGA) to cluster and label defect regions. The algorithm first scans the entire MCM to locate all foreground pixels (i.e., defect pixels). Each unlabeled foreground pixel is used as a seed point, and its 8-neighborhood pixels are examined: if neighboring pixels are also foreground and have not been visited, they are included in the current region and recursively expanded as new seed points. The growth criteria include two aspects: spatial continuity, requiring adjacent pixel positions; and gray-level consistency. Since the MCM is a binary image, all foreground pixels have a gray value of 255 (white), so only foreground pixels need to be considered. The growth process continues until no new pixels can be included, at which point an independent Connected Defect Region (CDR) is generated. Each Corresponding Data Map (CDR) is assigned a unique Region Label (RL), with label values ​​starting from 1 and incrementing sequentially (e.g., Label_1, Label_2, ...). The final output is a Connected Region Label Map (CRLM), where each pixel value represents the label number of its region, and background pixels are marked as 0.

[0110] To improve algorithm efficiency, a queue optimization strategy (QOS) is adopted. The system maintains a first-in, first-out (FIFO) pixel queue (PixelQueue, PQ). When a new seed point is found, its coordinates (x, y) are added to the tail of the queue; during processing, a pixel is taken from the head of the queue and its neighborhood is checked, and newly found neighboring pixels are added to the tail of the queue. To avoid duplicate visits, a visited matrix (VM) of the same size as the MCM is established to record whether each pixel has been processed. For ultra-large defect regions (area > 10,000 pixels²), block-based parallel processing (BPP) is enabled: the image is divided into several 256×256 pixel sub-blocks, each sub-block performs region growing independently, and then the connected regions across sub-blocks are merged using the boundary pixel matching (BPM) algorithm. Label assignment uses a Union-Find Data Structure (UFDS) to accelerate region merging: when the boundary regions of two adjacent sub-blocks belong to the same defect, their root labels are updated quickly. The entire labeling process can be completed within 100 milliseconds (ms) for an image containing 500 defect regions.

[0111] Before output, perform Region Validity Verification (RVV). Calculate the geometric attributes of each CDR, including:

[0112] Area (AR): Total number of pixels in the region; Bounding Box (BB): Coordinates of the smallest bounding 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 smaller than the Minimum Valid Region Threshold (MVRTh = 10 pixels²), or if the Perimeter-Area Ratio (PAR) is abnormally high (>3.0, indicating fragmented noise), its label is reset to 0 (i.e., it is classified as background). Finally, the CRLM retains only connected regions that conform to the physical defect size rules, along with their Geometric Attribute Table (GAT).

[0114] Load the copper-clad laminate design circuit topology information and construct the circuit topology diagram;

[0115] The system loads the Design Circuit Topology Information (DCTI) of copper-clad laminates from the Engineering Database (EDB). This information typically originates from computer-aided design (CAD) files (such as Gerber or ODB++ format) and contains three core layers: the copper layer, the substrate layer, and the solder mask layer. Key data includes:

[0116] Trace Geometry Vectors: Describe the start / end coordinates, width, and direction of copper traces; Pad Positions: Mark the center coordinates and dimensions of component solder joints; Copper Area Boundaries: Define the polygonal outline of large copper areas; Aperture Features: Identify the locations of drill holes or slots.

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

[0118] Nodes: Represent electrical connection points (such as pad centers, line intersections); Edges: Represent copper trace segments, with weights equal to the segment length (in micrometers); Faces: Represent enclosed copper areas, storing the coordinates of their polygon vertices.

[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 CRLM, with copper areas marked as 1 (conductivity) and non-copper areas (substrate, holes) marked as 0 (insulation).

[0121] Key Point Extraction: Corner detection is used to locate line turning points and pad centers as candidate nodes.

[0122] Topology Connection: Initial connections between nodes are established based on Delaunay triangulation, and redundant edges are trimmed according to the actual routing path.

[0123] Area Segmentation: Identifies closed copper areas using a scan-line filling algorithm and calculates the area and centroid of each area.

[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, coordinate transformation calibration (CTC) is performed. Pre-defined reference markers (FM) are placed at the four corners of the copper-clad laminate, whose design coordinates (DC) and actual coordinates (Image Coordinates, IC) in the high-speed camera image are known. The affine transformation matrix (ATM) is solved using the least squares method to uniformly transform the coordinates of all nodes / regions in the CTG to the image coordinate system. The transformation error is controlled within ±2 pixels to ensure accurate overlay between the topology map and the CRLM.

[0126] By combining the circuit topology diagram, physical constraint verification is performed on the connected region label diagram to verify whether the connected regions conform to the stress propagation model, eliminate false detections in non-copper areas, and output a precise location defect map based on physical verification.

[0127] The Connected Region Label Map (CRLM) is overlaid with the Line Topology Map (CTG), and Physical Constraint Verification (PCV) is performed. The core objective is to verify whether each defective region (CDR) conforms to the Stress Propagation Model (SPM). This model is based on the propagation characteristics of ultrasonic waves in copper-clad laminates.

[0128] Poor Bonding Regions: Located at the copper-substrate interface, stress waves are reflected more strongly here, resulting in abnormally high amplitudes. The shape of these regions should be parallel to the copper cladding boundary (e.g., a strip-shaped area along the edge of a large copper cladding area).

[0129] Micro-crack regions: Stress wave propagation is blocked by cracks, which manifests as a sudden drop in amplitude, and the crack direction should be perpendicular to the principal stress direction (such as perpendicular to high-frequency signal traces).

[0130] Stress anomaly regions: concentrated at points of geometric change (such as pad edges and trace corners), appearing as localized high-amplitude spots.

[0131] The verification process involves three steps:

[0132] Location Compliance Check (LCC): If the CDR is located entirely in a non-copper-covered area (such as substrate holes or solder mask), it is considered a false positive (such as dust adhesion) and is directly rejected.

[0133] Morphology Conformity Assessment (MCA): Calculate the average boundary distance (ABD) between the CDR profile and the nearest copper boundary. The ABD in poorly bonded areas should be less than 5 pixels, while the ABD in microcrack areas can be larger but must maintain directional consistency.

[0134] Stress Propagation Simulation (SPS): Using the CDR centroid as the excitation point, the finite difference method is used to simulate stress wave propagation. Real-world defects will cause the correlation coefficient (CC) between the simulated amplitude map (SAM) and the dynamic response feature map in this region to be greater than 0.7.

[0135] Region Connectivity Validation (RCV) is specifically designed to eliminate two types of false positives:

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

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

[0138] Implemented through the Spatial Relation Reasoning Engine (SRRE):

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

[0140] For the CDRs that pass the CCR check, calculate the angle (Angle to Traces, ATT) between its bounding rectangle and adjacent copper traces. The ATT of microcracks should be close to 90° (vertical), and the ATT of poor bonding should be close to 0° (parallel).

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

[0142] Poor bonding area: Filled with red, accompanied by an interface direction arrow; Microcrack area: Blue contour line, marking the crack length; Stress anomaly area: Yellow spots, marking the stress concentration coefficient.

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

[0144] S205, based on the precisely located defect area map, calculate the area ratio, distribution density, and distance to the critical path of various defects, and generate a comprehensive evaluation report of copper cladding quality according to the preset quality scoring rules. The comprehensive evaluation report includes defect type statistics, location marking, and quality level determination.

[0145] Specifically, it can parse and accurately locate defect area maps, extract the geometric attributes of each defect area, calculate the area ratio and distribution density, and output a defect measurement dataset;

[0146] The system receives the Precisely Located Defect Region Map (PLDRM) generated in the previous step. This map is in binary image format, where white pixels represent physically verified real defect areas, and black pixels represent normal areas. Each defect area is labeled with a type label (e.g., Bonding_Failure_Label, BFL; Microcrack_Label, MCL; Stress_Anomaly_Label, SAL). The parsing process first traverses the image pixels using the Connected Component Labeling Algorithm (CCLA) to assign a unique identifier (Defect ID, DID) to each individual defect area. Then, for each pixel cluster corresponding to a DID, the system calls the Geometric Attribute Extraction Engine (GAEE). The engine calculates core geometric attributes: Defect Area (DA) is calculated by multiplying the total number of pixels under a given DID by a preset Spatial Resolution Parameter (SRP, unit: micrometer / pixel) to convert the pixel value into an actual physical area (unit: square millimeter); Defect Perimeter (DP) is calculated using a boundary tracing algorithm to determine the contour length; Minimum Bounding Rectangle (MBR) records the orientation and extent of the defect. For example, a microcrack region with a DID of 103 has a calculated DA of 0.15 square millimeters, a DP of 1.8 millimeters, and an MBR size of 0.3 millimeters × 0.5 millimeters.

[0147] The calculation of Area Percentage (AP) requires consideration of the total inspected area of ​​the copper-clad laminate. The system reads the total inspected area (TIA, unit: square centimeters) from the inspection configuration file. For each type of defect (e.g., all MCL type defects), its AP value is equal to the sum of the densities (DA) of all defects of that type divided by the TIA and then multiplied by 100%. For example, if 5 microcracks are detected on the entire board, the sum of the DA is 0.75 square millimeters, and the TIA is 200 square centimeters (i.e., 20,000 square millimeters), then the AP of the microcracks = (0.75 / 20,000) × 100% = 0.00375%. Distribution Density (DD) reflects the number of defects per unit area. The system divides the copper-clad laminate into virtual grid cells (GC), with the grid size set according to process requirements (e.g., 10 mm × 10 mm). The number of defects of a specific type within each GC is calculated and then divided by the area of ​​that GC to obtain the local density value. The final global defect DD output for this type of defect is the arithmetic mean of the densities of all non-empty GCs. For example, in a grid of 200 meshes, 30 meshes contain microcracks, for a total of 50 microcracks. The average density per cracked mesh is 50 / 30 / 1 cm² = 1.67 microcracks / cm². All geometric attributes (DA, DP, MBR) and derived metrics (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 measurement data, the system performs Data Validity Check (DVC). Validation rules include: area reasonableness (e.g., DA must not exceed 1 square centimeter to prevent mislabeling of large areas), location boundary constraints (defect coordinates must be within the TIA range), and type consistency (all pixel labels within the same DID must be consistent). If abnormal data is detected (e.g., a BFL-type defect with a DA of 5 square centimeters), the Anomaly Tracing Mechanism (ATM) is triggered, automatically retrieving the original vibration characteristic map of the area and intermediate processing results for manual review. The validated DMD is output to downstream modules in a standardized format (e.g., JSON or XML). Its data structure includes fields such as: defect ID, type, center coordinates, area, perimeter, circumscribed rectangle, area percentage, and distribution density, providing a quantitative basis for subsequent risk assessment.

[0149] Based on the defect measurement dataset and the design route topology information, the nearest neighbor algorithm is used to calculate the minimum Euclidean distance from the defect area to the critical route, and the distance risk index set is output.

[0150] The system synchronously loads the Design Circuitry Topology Information (DCTI). This information originates from the computer-aided design (CAD) file and is stored in vector graphics format. It includes the geometric paths of all conductive copper traces (composed of line segments, arcs, and other primitives) and Critical Trace Markers (CTMs). Critical traces typically refer to defect-sensitive areas such as high-frequency signal lines, power trunks, or impedance control lines, and are marked with special attributes in the DCTI (e.g., Critical_Level = High). The system first performs rasterization preprocessing (RP) on the vector trace data, converting it into a binary image with the same resolution and spatial alignment as the Precise Defect Region Map (PLDRM) (traffic areas are white, non-traffic areas are black). Simultaneously, it extracts the set of pixel coordinates for all traces marked as critical, generating the Critical Trace Coordinate Set (CTCS).

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

[0152] Constructing a spatial index: A spatial index is built for the pixel coordinates of critical lines in the CTCS using a k-dimensional tree (kD-Tree) data structure. The kD-Tree significantly improves the efficiency of nearest neighbor queries by recursively partitioning the space (e.g., alternating partitions along the X / Y axes).

[0153] Traverse the defect region: For each defect DID in the Defect Measurement Dataset (DMD), obtain its Defect Boundary Coordinate Set (DBCS).

[0154] Parallel distance calculation: For each boundary point in the DBCS, the nearest critical trace point (CTP) in the CTCS is quickly found using the kD-Tree index, and the Euclidean distance (ED = √[(x1-x2)]) between the two points is calculated. 2 +(y1-y2) 2 ]).

[0155] Determine the minimum distance: Compare the ED values of all boundary points of the defect, and take the minimum value as the MED of the defect to the critical line. For example, the ED from 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 = 0.42 mm. If the distances of other boundary points are all greater than this value, then the MED of this defect = 0.42 mm.

[0156] The calculated MED needs to be converted into a risk level (Risk Level, RL) in combination with process specifications. The system presets distance threshold parameters (Distance Threshold Parameters, DTP), for example: High Risk (HR): MED ≤ 0.1 mm; Medium Risk (MR): 0.1 mm < MED ≤ 0.3 mm; Low Risk (LR): MED > 0.3 mm.

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

[0158] Input the defect measurement data set and the distance risk indicator set into the quality scoring engine, and perform weighted fusion according to the preset rules to generate the original quality scoring vector;

[0159] The core of the Quality Scoring Engine (QSE) is the 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 the percentage of defect area (AP), the larger the AP, the more points are deducted; Density Factor (DF): Based on the defect distribution density (DD), high density indicates concentrated problems; Risk Factor (RF): Based on the distance risk level (RL), high-risk defects have the highest deduction weight; Type Factor (TF): Based on the inherent hazard of the defect type (e.g., microcracks TF > poor bonding TF > stress anomalies TF). The weighting coefficients WC are predetermined using the Analytic Hierarchy Process (AHP), for example: RF weight 0.4, AF weight 0.3, DF weight 0.2, TF weight 0.1, summing to 1.

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

[0162] Factor normalization: Converts raw data of different dimensions into scores within the [0,1] interval. For example, the AP value is transformed using a linear normalization function (LNF): AF_score = 1 - (AP / AP_max), where AP_max is the preset maximum allowable area percentage (e.g., 0.01%). Factor weighting: Calculates weighted scores according to weights. For example, if the AF_score of a defect is 0.7, then its AF contribution score = 0.7 × 0.3 = 0.21. Individual defect scoring: For each defect, calculates the weighted sum of all its factors:

[0163] Defect_Score = (AF_score × WC_AF) + (DF_score × WC_DF) + (RF_score × WC_RF) + (TF_score × WC_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: microcrack → 0, poor bonding → 0.6, stress anomaly → 1.0.

[0164] Overall Raw Score: Calculate the raw quality score for the entire board. First, calculate the average of all defects (Defect_Score), then subtract this value from 1 (the more defects and the more severe they are, the lower the total score): Raw_Quality_Score = 1 - (∑Defect_Score / N). N is the total number of defects, and this value ranges from [0,1]. The closer it is to 1, the better the quality.

[0165] The Raw Quality Score Vector (RQSV) contains multi-level scoring results: the overall raw quality score; average scores categorized by defect type (e.g., Bonding_Avg_Score, Crack_Avg_Score); and zone scores calculated by spatial partitions (e.g., four quadrants) (Zone1_Score, Zone2_Score, etc.). For example, a copper-clad laminate might have an overall raw quality score of 0.82, an average score of 0.65 for microcracks (poor), and an average score of 0.88 for poor bonding (good). This vector serves as the direct basis for determining the quality level.

[0166] Based on the original quality scoring vector, a comprehensive evaluation report is generated, including a defect type statistics table, a location annotation map, and a quality level judgment label.

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

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

[0169] Base Map Composite: The original copper-clad laminate appearance drawing, the designed circuit topology diagram, and the Precisely Targeted Defect Region Map (PLDRM) are semi-transparently overlaid to form a composite base map. Defect Annotation: Each defect area is marked with a colored outline (e.g., microcracks - red, poor bonding - yellow, abnormal stress - blue); the defect ID (e.g., Crack_103) is marked within the outline; a lead line is drawn from the defect center, ending in an Info Callout Box (ICB) displaying the defect's DID, type, area, risk level, and distance to the nearest critical path. Hotspot Highlighting: High-risk defects (RL=HR) are highlighted with a flashing outline and a warning symbol added.

[0170] Zoning Identification: If zoning scoring is used, dashed lines are used to delineate region boundaries and region numbers are labeled. This map is output in high-resolution bitmap or vector (SVG / PDF) format, supporting zooming to view details.

[0171] The Quality Grade Tag (QGT) is generated based on the Raw Quality Score Vector (RQSV):

[0172] The Grade Rule Base (GRB) predefines the following grade thresholds: Excellent (Grade A): Overall_Raw_Score ≥ 0.90; Good (Grade B): 0.80 ≤ Overall_Raw_Score < 0.90; Average (Grade C): 0.70 ≤ Overall_Raw_Score < 0.80; Poor (Grade D): Overall_Raw_Score < 0.70.

[0173] Constraint Check (CC): Even if the total score is high, if there are high-risk defects (such as HR number ≥ 3) or defects in critical areas (such as cracks under the CPU socket), the grade will be forcibly downgraded (e.g., A→B).

[0174] Label Generation: Based on the Overall_Raw_Score and constraint check results, the system outputs the final grade label (e.g., "Grade B"), along with key supporting explanations (e.g., "Total score 0.87, but due to the presence of 2 high-risk microcracks, it did not reach Grade A"). All content (statistical tables, location maps, grade labels) is automatically integrated into a Comprehensive Evaluation Report (CER), with output formats selectable as PDF / HTML, and supports one-click export to the Production Management System (MES).

[0175] Another embodiment of the present invention provides a copper clad laminate copper clad laminate quality inspection system, see [link to relevant documentation]. Figure 3 The system may include:

[0176] The acquisition module 301 is used to apply an ultrasonic excitation signal with a preset frequency and power to the copper-clad laminate under test, and simultaneously use a high-speed camera to acquire a sequence of dynamic micro-vibration response images of the surface of the copper-clad laminate under ultrasonic excitation. The ultrasonic frequency is related to the natural frequency of the copper-clad laminate substrate and the copper layer to excite micro-vibration in a specific mode.

[0177] Analysis module 302 is used to perform phase-locked analysis on the dynamic micro-vibration response image sequence, extract the amplitude-phase change features of each pixel in the ultrasonic excitation cycle, and construct a dynamic response feature map of copper-clad laminate. 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 the environmental noise.

[0178] The identification module 303 is used to input the dynamic response feature map of the copper-clad laminate into the trained defect identification model, identify and locate potential poor copper-clad bonding areas, microcrack areas and local stress anomaly areas, and output a preliminary defect area distribution map. The defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples.

[0179] The verification module 304 is used to perform multi-scale morphological analysis and regional connectivity verification on the preliminary defect area distribution map. Combined with the design circuit topology information of the copper-clad laminate, it filters out the real defect areas that conform to physical laws and generates a precise location defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad areas.

[0180] The generation module 305 is used to calculate the area ratio, distribution density, and distance to critical paths of various defects based on the precise location defect area map, and generate a comprehensive evaluation report of copper cladding quality according to preset quality scoring rules. The comprehensive evaluation report includes defect type statistics, location marking, and quality level determination.

[0181] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for inspecting the copper clad laminate quality, characterized in that, The method includes: An ultrasonic excitation signal with a preset frequency and power is applied to the copper-clad laminate under test, and a high-speed camera is used to simultaneously acquire a sequence of dynamic micro-vibration response images of the surface of the copper-clad laminate under ultrasonic excitation. The ultrasonic frequency is related to the natural frequency of the copper-clad laminate substrate and the copper layer to excite micro-vibration of a specific mode. Phase-locked analysis is performed on the dynamic micro-vibration response image sequence to extract the amplitude-phase change features of each pixel within the ultrasonic excitation cycle, constructing a dynamic response feature map of the copper-clad laminate. The phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by the excitation from environmental noise. Specifically, a phase reference timestamp is extracted from the ultrasonic excitation signal to construct a phase reference vector. Based on the phase reference vector, a time-domain alignment operation is performed on the dynamic micro-vibration response image sequence, outputting the aligned image sequence. A wavelet packet transform algorithm is used to analyze each pixel of the aligned image sequence, extracting the instantaneous amplitude and phase angle of the dominant frequency band of the micro-vibration, outputting a pixel-level amplitude-phase matrix. A phase coherence filtering algorithm, using the phase reference vector as a reference, suppresses the environmental noise component in the amplitude-phase matrix, separating the pure micro-vibration signal, and outputting a noise-suppressed amplitude-phase feature set. The noise-suppressed amplitude-phase feature set is mapped to two-dimensional spatial coordinates to construct the dynamic response feature map of the copper-clad laminate. The dynamic response feature map of the copper-clad laminate is input into the trained defect identification model to identify and locate potential poor copper-clad laminate bonding areas, microcrack areas and local stress anomaly areas, and output a preliminary defect area distribution map. The defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples. Multi-scale morphological analysis and regional connectivity verification are performed on the preliminary defect area distribution map. Combined with the design circuit topology information of the copper-clad laminate, the real defect areas that conform to physical laws are screened out to generate a precise defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad areas. Based on the precisely located defect area map, the area ratio, distribution density, and distance to critical paths of various defects are calculated, and a comprehensive evaluation report on copper cladding quality is generated according to the preset quality scoring rules. The comprehensive evaluation report includes defect type statistics, location marking, and quality level determination.

2. The method according to claim 1, characterized in that, The process involves applying an ultrasonic excitation signal of a preset frequency and power to the copper-clad laminate under test, and simultaneously using a high-speed camera to acquire a sequence of dynamic micro-vibration response images of the laminate surface under ultrasonic excitation. The ultrasonic frequency is associated with the natural frequencies of the copper-clad laminate substrate and copper layer to excite micro-vibrations of a specific mode, including: Based on the material thickness and density parameters of copper-clad laminate, the natural frequency spectrum of the substrate and copper layer is calculated by the finite element modal analysis algorithm, and the natural frequency spectrum is output as a preset frequency candidate set. Using a preset frequency candidate set, the output frequency of the ultrasonic generator is dynamically tuned by an adaptive resonance tracking algorithm. The frequency point that can excite the maximum micro-vibration amplitude is selected, and the optimized ultrasonic excitation signal is output. Based on the optimized periodic characteristics of the ultrasonic excitation signal, a frame rate synchronization module for the high-speed camera is configured to ensure that the frame rate matches an integer multiple of the excitation frequency and output a set of synchronization parameters. An optimized ultrasonic excitation signal is applied to the copper-clad laminate, and a high-speed camera is triggered to acquire images based on a synchronization parameter set, generating a dynamic micro-vibration response image sequence.

3. The method according to claim 2, characterized in that, The process involves inputting the dynamic response feature map of the copper-clad laminate into a trained defect identification model to identify and locate potential areas of poor copper-clad bonding, microcrack areas, and localized stress anomalies, and outputting a preliminary defect area distribution map. The defect identification model is a deep learning model trained based on the dynamic response feature maps of known defect samples, and includes: Multi-channel feature enhancement is performed on the dynamic response feature map of copper clad laminate. Abnormal response areas are highlighted by weighting through an attention mechanism, and an enhanced feature map is output. The enhanced feature map is input into a pre-trained defect recognition model, which adopts a graph convolutional network architecture, integrates dynamic response features and prior knowledge of material stress distribution, and outputs the original defect probability map. An adaptive threshold segmentation algorithm is applied to process the original defect probability map, dynamically dividing the defect category boundaries according to the probability values, and outputting a binary defect mask map. Mark connected regions on the binary defect mask and classify them as poor copper bonding, microcracks or stress anomalies, and output a preliminary defect region distribution map.

4. The method according to claim 3, characterized in that, The preliminary defect area distribution map is subjected to multi-scale morphological analysis and regional connectivity verification. Combined with the design circuit topology information of the copper-clad laminate, real defect areas that conform to physical laws are screened out to generate a precisely located defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad areas, including: Perform multi-scale morphological opening and closing operations on the preliminary defect area distribution map, use variable-scale structuring elements to eliminate isolated noise points, and output a morphologically corrected map. Based on the morphological correction map, a region growing algorithm is used to label connected defect regions and output a connected region label map. Load the copper-clad laminate design circuit topology information and construct the circuit topology diagram; By combining the circuit topology diagram, physical constraint verification is performed on the connected region label diagram to verify whether the connected regions conform to the stress propagation model, eliminate false detections in non-copper areas, and output a precise location defect map based on physical verification.

5. The method according to claim 4, characterized in that, Based on the precisely located defect area map, the area proportion, distribution density, and distance to critical paths of various defects are calculated. A comprehensive evaluation report on copper plating quality is generated according to preset quality scoring rules. This comprehensive evaluation report includes defect type statistics, location markings, and quality level determination, including: Analyze the precise location map of the defect area, extract the geometric attributes of each defect area, calculate the area ratio and distribution density, and output the defect measurement dataset. Based on the defect measurement dataset and the design route topology information, the nearest neighbor algorithm is used to calculate the minimum Euclidean distance from the defect area to the critical route, and the distance risk index set is output. The defect measurement dataset and the distance risk index dataset are input into the quality scoring engine, and then weighted and fused according to preset rules to generate the original quality score vector. Based on the original quality scoring vector, a comprehensive evaluation report is generated, including a defect type statistics table, a location annotation map, and a quality level judgment label.

6. A copper-clad laminate copper-clad laminate quality inspection system, characterized in that, The system includes: The acquisition module is used to apply an ultrasonic excitation signal with a preset frequency and power to the copper-clad laminate under test, and simultaneously use a high-speed camera to acquire a sequence of dynamic micro-vibration response images of the surface of the copper-clad laminate under ultrasonic excitation. The ultrasonic frequency is related to the natural frequency of the copper-clad laminate substrate and the copper layer to excite micro-vibration in a specific mode. The analysis module performs phase-locked analysis on the dynamic micro-vibration response image sequence, extracts the amplitude-phase change features of each pixel within the ultrasonic excitation cycle, and constructs a dynamic response feature map of the copper-clad laminate. The phase-locked analysis is based on the phase reference of the ultrasonic excitation signal to separate the micro-vibration signal caused by the excitation from environmental noise. Specifically, a phase reference timestamp is extracted from the ultrasonic excitation signal to construct a phase reference vector. Based on the phase reference vector, a time-domain alignment operation is performed on the dynamic micro-vibration response image sequence, outputting the aligned image sequence. A wavelet packet transform algorithm is used to analyze each pixel of the aligned image sequence, extracting the instantaneous amplitude and phase angle of the dominant frequency band of the micro-vibration, and outputting a pixel-level amplitude-phase matrix. A phase coherence filtering algorithm, using the phase reference vector as a reference, suppresses the environmental noise component in the amplitude-phase matrix, separating the pure micro-vibration signal and outputting a noise-suppressed amplitude-phase feature set. The noise-suppressed amplitude-phase feature set is mapped to two-dimensional spatial coordinates to construct the dynamic response feature map of the copper-clad laminate. The identification module is used to input the dynamic response feature map of the copper-clad laminate into the trained defect identification model, identify and locate potential poor copper-clad bonding areas, microcrack areas and local stress anomaly areas, and output a preliminary defect area distribution map. The defect identification model is a deep learning model trained based on the dynamic response feature map of known defect samples. The verification module is used to perform multi-scale morphological analysis and regional connectivity verification on the preliminary defect area distribution map. Combined with the design circuit topology information of the copper-clad laminate, it filters out the real defect areas that conform to physical laws and generates a precise location defect area map. The regional connectivity verification is used to eliminate false detections caused by image noise or non-copper-clad areas. The generation module is used to calculate the area ratio, distribution density, and distance to critical paths of various defects based on the precisely located defect area map, and generate a comprehensive evaluation report of copper cladding quality according to preset quality scoring rules. The comprehensive evaluation report includes defect type statistics, location marking, and quality level determination.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Based on the material thickness and density parameters of copper-clad laminate, the natural frequency spectrum of the substrate and copper layer is calculated by the finite element modal analysis algorithm, and the natural frequency spectrum is output as a preset frequency candidate set. Using a preset frequency candidate set, the output frequency of the ultrasonic generator is dynamically tuned by an adaptive resonance tracking algorithm. The frequency point that can excite the maximum micro-vibration amplitude is selected, and the optimized ultrasonic excitation signal is output. Based on the optimized periodic characteristics of the ultrasonic excitation signal, a frame rate synchronization module for the high-speed camera is configured to ensure that the frame rate matches an integer multiple of the excitation frequency and output a set of synchronization parameters. An optimized ultrasonic excitation signal is applied to the copper-clad laminate, and a high-speed camera is triggered to acquire images based on a synchronization parameter set, generating a dynamic micro-vibration response image sequence.

8. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images